DATA CLASSIFICATION USING MACHINE LEARNING TECHNIQUES
Systems, methods and computer program products for classifying documents are presented. Systems, methods and computer program products for analyzing documents, e.g. for verifying an association of an invoice with an entity are also presented. Systems, methods and computer program products for managing medical records are presented. One exemplary system includes a memory; and a processor in communication with the memory, the processor being configured to process at least some instructions stored in the memory. The memory stores computer executable program code comprising instructions for: training a classifier based on an invoice format associated with a first entity; accessing a plurality of invoices labeled as being associated with at least one of the first entity and other entities; and outputting an identifier of at least one of the invoices having a high probability of not being associated with the first entity.
Latest Kofax, Inc. Patents:
 Range and/or polaritybased thresholding for improved data extraction
 Smart optical input/output (I/O) extension for contextdependent workflows
 System and method for providing a universal endpoint address schema to route documents and manage document workflows
 Iterative recognitionguided thresholding and data extraction
 Systems and methods for organizing data sets
This application is a continuation of U.S. patent application Ser. No. 13/090,216, filed Apr. 19, 2011, which is a continuation of U.S. patent application Ser. No. 11/752,673 filed May 23, 2007; each of which claims priority to U.S. Provisional Patent Application Ser. No. 60/830,311, filed Jul. 12, 2006, and each of which are herein incorporated by reference.
FIELD OF THE INVENTIONThe present invention relates generally to methods and apparatus for data classification. More particularly, the present invention relates to novel applications using machine learning techniques.
BACKGROUNDHow to handle data has gained in importance in the information age and more recently with the explosion of electronic data in all walks of life including, among others, scanned documents, web material, search engine data, text data, images, audio data files, etc.
One area just starting to be explored is the nonmanual classification of data. In many classification methods the machine or computer must learn based upon manually input and created rule sets and/or manually created training examples. In machine learning where training examples are used, the number of learning examples is typically small compared to the number of parameters that have to be estimated, i.e. the number of solutions that satisfy the constraints given by the training examples is large. A challenge of machine learning is to find a solution that generalizes well despite the lack of constraints. There is thus a need for overcoming these and/or other issues associated with the prior art.
What is further needed are practical applications for machine learning techniques of all types.
SUMMARYA system for verifying an association of an invoice with an entity according to another embodiment of the present invention includes a memory; and a processor in communication with the memory, the processor being configured to process at least some instructions stored in the memory, wherein the memory stores computer executable program code comprising instructions for: training a classifier based on an invoice format associated with a first entity; accessing a plurality of invoices labeled as being associated with at least one of the first entity and other entities; performing a document classification technique on the invoices using the classifier; and outputting an identifier of at least one of the invoices having a high probability of not being associated with the first entity.
A computer program product for verifying an association of an invoice with an entity according to another embodiment of the present invention includes a program storage medium readable by a computer, where the medium tangibly embodies one or more programs of instructions executable by the computer to perform a method, comprising: training a classifier based on an invoice format associated with a first entity; accessing a plurality of invoices labeled as being associated with at least one of the first entity and other entities; performing a document classification technique on the invoices using the classifier; and outputting an identifier of at least one of the invoices having a high probability of not being associated with the first entity.
A system for managing medical records according to another embodiment of the present invention includes a memory; and a processor in communication with the memory, the processor being configured to process at least some instructions stored in the memory, wherein the memory stores computer executable program code comprising instructions for: training a classifier based on a medical diagnosis; accessing a plurality of medical records; performing a document classification technique on the medical records using the classifier, and outputting an identifier of at least one of the medical records having a low probability of being associated with the medical diagnosis.
The following description is the best mode presently contemplated for carrying out the present invention. This description is made for the purpose of illustrating the general principles of the present invention and is not meant to limit the inventive concepts claimed herein. Further, particular features described herein can be used in combination with other described features in each of the various possible combinations and permutations.
Unless otherwise specifically defined herein, all terms are to be given their broadest possible interpretation including meanings implied from the specification as well as meanings understood by those skilled in the art and as defined in dictionaries, treatises, etc.
The interest and need for classification of textual data has been particularly strong, and several methods of classification have been employed. A discussion of classification methods for textual data is described in U.S. Pat. No. 6,192,360 to Dumais, the content and substance of which is incorporated herein by reference.
The current state of the art in commercially used automatic classification systems is either rule based or utilizes inductive machine learning, i.e. using manually labeled training examples. Both methods typically entail a large manual setup effort compared to transductive methods. The solutions provided by rule based systems or inductive methods are static solutions that cannot adapt to drifting classification concepts without manual effort.
Inductive machine learning is used to ascribe properties or relations to types based on tokens (i.e., on one or a small number of observations or experiences); or to formulate laws based on limited observations of recurring patterns. Inductive machine learning involves reasoning from observed training cases to create general rules, which are then applied to the test cases.
Particularly, preferred embodiments use transductive machine learning techniques. Transductive machine learning is a powerful method that does not suffer from these disadvantages.
Transductive machine techniques may be capable of learning from a very small set of labeled training examples, automatically adapting to drifting classification concepts, and automatically correcting the labeled training examples. These advantages make transductive machine learning an interesting and valuable method for a large variety of commercial applications.
Transduction learns patterns in data. It extends the concept of inductive learning by learning not only from labeled data but also from unlabeled data. This enables transduction to learn patterns that are not or only partly captured in the labeled data. As a result transduction can, in contrast to rule based systems or systems based on inductive learning, adapt to dynamically changing environments. This capability enables transduction to be utilized for document discovery, data cleanup, and addressing drifting classification concepts, among other things.
The following is an explanation of one embodiment of transductive classification utilizing Support Vector Machine (SVM) classification as well as the Maximum Entropy Discrimination (MED) framework.
Support Vector MachinesSupport Vector Machines (SVM) is one employed method of text classification, and such method approaches the problem of the large number of solutions and the resulting generalization problem by deploying constraints on the possible solutions utilizing concepts of regularization theory. For example, a binary SVM classifier selects from all hyperplanes that separate the training data correctly as solution the hyperplane that maximizes the margin. The maximum margin regularization under the constraint that training data is classified correctly addresses the aforementioned learning problem of selecting the appropriate tradeoff between generalization and memorization: The constraint on the training data memorizes the data, whereas the regularization ensures appropriate generalization. Inductive classification learns from training examples that have known labels, i.e. every training example's class membership is known. Where inductive classification learns from known labels, transductive classification determines the classification rules from labeled as well as unlabeled data. An example of transductive SVM classification is shown in table 1.
Principle of Transductive SVM Classification
Table 1 shows the principle of a transductive classification with Support Vector Machines: The solution is given by the hyperplane that yields the maximum margin over all possible label assignments of the unlabeled data. The possible label assignments grow exponentially in the number of unlabeled data and for practically applicable solutions, the algorithm in Table 1 must be approximated. An example of such an approximation is described in T. Joachims, Transductive inference for text classification using support vector machines, Technical report, Universitaet Dortmund, LAS VIII, 1999 (Joachims).
The uniform distribution over label assignments in Table 1 implies that an unlabeled data point has a probability of ½ to be a positive example of the class and a probability of ½ of being a negative example, i.e. its two possible label assignments of y=+1 (positive example) and y=−1 (negative example) are equally likely and the resulting expected label is zero. A label expectation of zero can be obtained by a fixed class prior probability equal to ½ or a class prior probability that is a random variable with an uniform prior distribution, i.e. an unknown class prior probability. Accordingly, in applications with known class prior probabilities that are not equal to ½ the algorithm could be improved by incorporating this additional information. For example, instead of using a uniform distribution over label assignments in Table 1, one could elect to prefer some label assignments over others according to the class prior probability. However, the tradeoff between a smaller margin solution with a likely label assignment and a higher margin solution but less likely label assignment is difficult. The probability of label assignments and the margin are on different scales.
Maximum Entropy DiscriminationAnother method of classification, Maximum Entropy Discrimination (MED) (see e.g. T. Jebara, Machine Learning Discriminative and Generative, Kluwer Academic Publishers) (Jebara) does not encounter the problems associated with SVMs since the decision function regularization term as well as the label assignment regularization term are both derived from prior probability distributions over solutions and, thus are both on the same probabilistic scale. Accordingly, if the class priors and, thus, the label priors are known, transductive MED classification is superior to transductive SVM classification, since it allows for the incorporation of prior label knowledge in a principled way.
Inductive MED classification assumes a prior distribution over the parameters of the decision function, a prior distribution over the bias term, and a prior distribution over margins. It selects as a final distribution over these parameters the one that is closest to the prior distributions and yields an expected decision function that classifies the data points correctly.
Formally, for example given a linear classifier, the problem is formulated as follows: Find the distribution over hyperplane parameters p(Θ), the bias distribution p(b), the data points classification margins p(γ) whose combined probability distribution has a minimal Kullback Leibler divergence KL to the combined respective prior distributions p_{0}, i.e.
subject to the constraint
∀t:∫dΘdγdbp(Θ)p(γ)p(b)(y_{i}(ΘX_{i}−b))−y_{t})≧0, (2)
where the ΘX_{t }is the dot product between the separating hyperplane's weight vector and the tth data point's feature vector. Since the label assignments y_{t }are known and fixed, no prior distribution over the binary label assignments is needed. Accordingly, a straightforward method to generalize inductive MED classification to transductive MED classification is to treat the binary label assignments as parameters that are constrained by a prior distribution over possible label assignments. An example of transductive MED is shown in Table 2.
Transductive MED Classification
For the labeled data, the label prior distribution is a δ function, thus, effectively fixing the label to be either +1 or −1. For the unlabeled data, a label prior probability p_{0}(y) is assumed that assigns to every unlabeled data point a positive label of y=+1 with a probability of p_{0}(y) and a negative label of y=−1 with a probability of 1−p_{0}(y). Assuming a noninformative label prior (p_{0}(y)=½), yields a transductive MED classification analogous to the transductive SVM classification discussed above.
As in the case of the transductive SVM classification, a practically applicable implementation of such an MED algorithm must approximate the search through all possible label assignments. The method described in T. Jaakkola, M. Meila, and T. Jebara, Maximum entropy discrimination, Technical Report AITR1668, Massachusetts Institute of Technology, Artificial Intelligence Laboratory, 1999 (Jaakkola) elects as an approximation to decompose the procedure into two steps, similar to an Expectation Maximization (EM) formulation. In this formulation, there are two problems to solve. The first, analogous to the M step in EM algorithms, is similar to the maximization of the margin while classifying all data points correctly according to the current best guess of label assignments. The second step, analogous to the E step, uses the classification results determined in the M step and estimates new values for each example's class membership. This second step we call label induction. A general description is shown in Table 2.
The specific implementation of the method of Jaakkola, referenced herein, assumes a Gaussian with zero mean and unit variance for the hyperplane parameters, a Gaussian with zero mean and variance σ_{b}^{2 }for the bias parameter, a margin prior of the form exp [−c(1−γ)] with γ a data point's margin and c the cost factor, and a binary label prior probability of p_{0}(y) for unlabeled data as discussed above. For the following discussion of the transductive classification algorithm Jaakkola, referenced herein, a label prior probability of ½ is assumed for reasons of simplicity and without loss of generality.
The label induction step determines the label probability distribution given a fixed probability distribution for the hyperplane parameters. Using the margin and label priors introduced above yields the following objective function for the label induction step (see Table 2)
where λ_{t }is the tth training example Lagrange Multiplier, and s_{t }its classification score determined in the previous Mstep, and c the cost factor. The first two terms in the sum over the training examples is derived from the margin prior distribution, whereas the third term is given by the label prior distribution. By maximizing ℑ the Lagrange Multipliers are determined and, thus, the label probability distributions for the unlabeled data. As can be seen from Eq. 3 the data points contribute independently to the objective function and, thus, each Lagrange Multiplier can be determined irrespective of every other Lagrange Multiplier. For example, in order to maximize the contribution of an unlabeled data point with a high absolute value of its classification score s_{t} a small Lagrange Multiplier λ_{t }is required, whereas an unlabeled data point with a small value of s_{t} maximizes its contribution to ℑ with a large Lagrange Multiplier. On the other hand, the expected label y of an unlabeled data point as a function or its classification score s and its Lagrange Multiplier λ is
y=tan h(λs). (4)
The M step of the transductive classification algorithm of Jaakkola, referenced herein, determines the probability distributions for the hyperplane parameters, the bias term, and margins of the data points that are closest to the respective prior distribution under the constraints
∀t:s_{t}y_{t}−γ_{t}≧0, (5)
where s_{t }is the tth data point classification score, y_{t} its expected label and γ_{t} its expected margin. For labeled data, the expected label is fixed and either y=+1 or y=−1. The expected label for unlabeled data lies in the interval (−1, +1) and is estimated in the label induction step. According to Eq. 5 unlabeled data have to fulfill tighter classification constraints than labeled data since the classification score is scaled by the expected label. Furthermore, given the dependence of the expected label as a function of the classification score, referring to
The first term is derived from the Gaussian hyperplane parameters prior distribution, the second term is the margin prior regularization term and the last term is the bias prior regularization term derived from a Gaussian prior with zero mean and variance σ_{b}^{2}. The prior distribution over the bias term can be interpreted as a prior distribution over class prior probabilities. Accordingly, the regularization term that corresponds to the bias prior distribution constrains the weight of the positive to negative examples. According to Eq. 6, the contribution of the bias term is minimized in case the collective pull of the positive examples on the hyperplane equals the collective pull of the negative examples. The collective constraint on the Lagrange Multipliers owing to the bias prior is weighted by the expected label of the data points and is, therefore, less restrictive for unlabeled data than for labeled data. Thus, unlabeled data have the ability of influencing the final solution stronger than the labeled data.
In summary, at the M step of the transductive classification algorithm of Jaakkola, referenced herein, unlabeled data have to fulfill stricter classification constraints than the labeled data and their cumulative weight to the solution is less constrained than for labeled data. In addition, unlabeled data with an expected label close to zero that lie within the margin of the current M step influence the solution the most. The resulting net effect of formulating the E and M step this way is illustrated by applying this algorithm to the dataset shown in
This instability of this algorithm is not restricted to the example shown in
One preferred approach of the present invention employs transductive classification using the framework of Maximum Entropy Discrimination (MED). It should be understood that various embodiments of the present invention, while applicable for classification may also be applicable to other MED learning problems using transduction, including, but not limited to transductive MED regression and graphical models.
Maximum Entropy Discrimination constrains and reduces the possible solutions, by assuming a prior probability distribution over the parameters. The final solution is the expectation of all possible solutions according to the probability distribution that is closest to the assumed prior probability distribution under the constraint that the expected solution describes the training data correctly. The prior probability distribution over solutions maps to a regularization term, i.e. by choosing a specific prior distribution one has selected a specific regularization.
Discriminative estimation as applied by Support Vector Machines is effective in learning from few examples. This method and apparatus of one embodiment of the present invention has this in common with Support Vector Machines and does not attempt to estimate more parameters than necessary for solving the given problem and, consequently, yields a sparse solution. This is in contrast to generative model estimation that attempts to explain the underlying process and, in general needs higher statistics than discriminative estimation. On the other hand, generative models are more versatile and can be applied to a larger variety of problems. In addition, generative model estimation enables straightforward inclusion of prior knowledge. The method and apparatus of one embodiment of the present invention using Maximum Entropy Discrimination bridges the gap between pure discriminative, e.g. Support Vector Machine learning, and generative model estimation.
The method of one embodiment of the present invention as shown in Table 3 is an improved transductive MED classification algorithm that does not have the instability problem of the method discussed in Jaakkola, referenced herein. Differences include, but are not limited to, that in one embodiment of the present invention every data point has its own cost factor proportional to its absolute label expectation value y. In addition, each data points label prior probability is updated after each M step according to the estimated class membership probability as function of the data point's distance to the decision function. The method of one embodiment of the present invention is described in Table 3 as follows:
Improved Transductive MED Classification
Scaling the data points cost factors by y mitigates the problem that the unlabeled data can have a stronger cumulative pull on the hyperplane than the labeled data, since the cost factors of unlabeled data are now smaller than labeled data cost factors, i.e. each unlabeled data point's individual contribution to the final solution is always smaller than labeled data points individual contribution. However, in case the amount of unlabeled data is much larger then the number of labeled data, the unlabeled data still can influence the final solution more than the labeled data. In addition, the conjunction of cost factor scaling with updating the label prior probability using the estimated class probability solves the problem of the bridge effect outlined above. At the first M steps, unlabeled data have small cost factors yielding an expected label as function of the classification score that is very flat (see
In a specific implementation of the method of one embodiment of the present invention, by assuming a Gaussian prior with zero mean and unit variance for the decision function parameters Θ
The prior distribution over decision function parameters incorporates important prior knowledge of the specific classification problem at hand. Other prior distributions of decision function parameters important for classification problem are for example a multinomial distribution, a Poisson distribution, a Cauchy distribution (BreitWigner), a MaxwellBoltzman distribution or a BoseEinstein distribution.
The prior distribution over the threshold b of the decision function is given by a Gaussian distribution with mean μ_{b }and variance σ_{b}^{2}
As prior distribution of a data point's classification margin γ_{t}
Was elected, where c is the cost factor. This prior distribution differs from the one used in Jaakkola, referenced herein, which has the form exp[−c(1−γ)]. Preferably, the form given in Eq. 9 over the form used in Jaakkola, referenced herein, since it yields a positive expected margin even for cost factor smaller than one, whereas exp [−c(1−γ)] yields a negative expected margin for c<1.
Given these prior distributions, determining the corresponding partition functions Z is straightforward (see for example T. M. Cover and J. A. Thomas, Elements of Information Theory, John Wiley & Sons, Inc.) (Cover), and the objective functions ℑ=−log Z are
According to Jaakkola, referenced herein the objective function of the M step is
ℑ_{M}(λ)=ℑ_{Θ}(λ)+ℑ_{b}(λ)+ℑ_{γ}(λ) (11)
and the E step's objective function is
where s_{t }is the tth data point's classification score determined in the previous M step and p_{0,t}(y_{t}) the data point's binary label prior probability. The label prior is initialized to p_{0,t}(y_{t})=1 for labeled data and to either the noninformative prior of p_{0,t}(y_{t})=½ or the class prior probability for unlabeled data.
The section herein entitled M STEP describes the algorithm to solve the M step objective function. Also, the section herein entitled E STEP describes the E step algorithm.
The step EstimateClassProbability in line 5 of Table 3 uses the training data to determine the calibration parameters to turn classification scores into class membership probabilities, i.e. the probability of the class given the score p(cs). Relevant methods for estimating the score calibration to probabilities are described in J. Platt, Probabilistic outputs for support vector machines and comparison to regularized likelihood methods, pages 6174, 2000 (Platt) and B. Zadrozny and C. Elkan. Transforming classifier scores into accurate multiclass probability estimates, 2002 (Zadrozny).
Referring particularly to
Referring particularly to
Once data is accessed at step 104, the method of one embodiment of the present invention at step 108 then determines the label prior probabilities of the data point using the label information of data point. Then, at step 110 the expected labels of the data point are determined according to the label prior probability. With the expected labels calculated in step 110, along with the labeled data, unlabeled data and cost factors, step 112 includes iterative training of the transductive MED classifier by the scaling of the cost factor unlabeled data points. In each iteration of the calculation the unlabeled data points' cost factors are scaled. As such, the MED classifier learns through repeated iterations of calculations. The trained classifier then accessed input data 114 at step 116. The trained classifier can then complete the step of classifying input data at step 118 and terminates at step 120.
It is to be understood that the unlabeled data of 106 and the input data 114 may be derived from a single source. As such, the input data/unlabeled data can be used in the iterative process of 112 which is then used to classify at 118. Furthermore, one embodiment of the present invention contemplates that the input data 114 maybe include a feedback mechanism to supply the input data to the stored data at 106 such that the MED classifier of 112 can dynamically learn from new data that is input.
Referring particularly to
At step 208, expected labels are calculated from the data of 206. The expected labels then used in step 210 along with labeled data, unlabeled data and cost factors to conduct iterative training of a transductive MED classifier. The iterative calculations of 210 scale the cost factors of the unlabeled data at each calculation. The calculations continue until the classifier is properly trained.
The trained classifier then accessed input data at 214 from input data 212. The trained classifier can then complete the step of classifying input data at step 216. As with the process and method described in
In both methods as shown in
Referring particularly to
In step 308, expected labels are determined from the data from the prior probability information of 306. In step 310, the cost factor is scaled for each unlabeled data set proportional to the absolute value of the expected label of a data point. An MED classifier is then trained in step 312 by determining the decision function that maximizes the margin between the included training and excluded training examples utilizing the labeled as well as the unlabeled data as training examples according to their expected labels. In step 314 classification scores are determined using the trained classifier of 312. In 316 classification scores are calibrated to class membership probability. In step 318, label prior probability information is updated according to the class membership probability. An MED calculation is preformed in step 320 to determine label and margin probability distributions, wherein the previously determined classification scores are used in the MED calculation. As a result, new expected labels are computed at step 322 and the expected labels are updated in step 324 using the computations from step 322. At step 326 the method determines whether convergence has been achieved. If so, the method terminates at step 328. If convergence is not reached, another iteration of the method is completed starting with step 310. Iterations are repeated until convergence is reached thus resulting in an iterative training of the MED classifier. Convergence may be reached when change of the decision function between each iteration of the MED calculation falls below a predetermined value. In an alternative embodiment of the present invention, convergence may be reached when the change of the determined expected label value falls below a predetermined threshold value.
In use, the gateway 707 serves as an entrance point from the remote networks 702 to the proximate network 708. As such, the gateway 707 may function as a router, which is capable of directing a given packet of data that arrives at the gateway 707, and a switch, which furnishes the actual path in and out of the gateway 707 for a given packet.
Further included is at least one data server 714 coupled to the proximate network 708, and which is accessible from the remote networks 702 via the gateway 707. It should be noted that the data server(s) 714 may include any type of computing device/groupware. Coupled to each data server 714 is a plurality of user devices 716. Such user devices 716 may include a desktop computer, laptop computer, handheld computer, printer or any other type of logic. It should be noted that a user device 717 may also be directly coupled to any of the networks, in one embodiment.
A facsimile machine 720 or series of facsimile machines 720 may be coupled to one or more of the networks 704, 706, 708.
It should be noted that databases and/or additional components may be utilized with, or integrated into, any type of network element coupled to the networks 704, 706, 708. In the context of the present description, a network element may refer to any component of a network.
The workstation shown in
Referring particularly to
The processor 402 accesses the data from the memory 404 and using transductive MED calculations trains a binary classifier enable it to classify unlabeled data. The processor 402 uses iterative transductive calculation by using the cost factor and training examples from labeled and unlabeled data and scaling that cost factor as a function of expected label value thus effecting the data of the cost factor data 422 which is then reinput into processor 402. Thus the cost factor 422 changes with each iteration of the MED classification by the processor 402. Once the processor 402 adequately trains an MED classifier, the processor can then construct the classifier to classify the unlabeled data into classified data 424.
Transductive SVM and MED formulations of the prior art lead to an exponential growth of possible label assignments and approximations have to be developed for practical applications. In an alternative embodiment of the present invention, a different formulation of the transductive MED classification is introduced that does not suffer from an exponential growth of possible label assignments and allows a general closed form solution. For a linear classifier the problem is formulated as follows: Find the distribution over hyperplane parameters p(Θ), the bias distribution p(b), the data points classification margins p(γ) whose combined probability distribution has a minimal Kullback Leibler divergence KL to the combined respective prior distributions p_{0}, i.e.
subject to the following constraint for the labeled data
∀t:∫dΘdγdbp(Θ)p(γ)p(b)(y_{t}(ΘX_{t}−b))−γ_{t})≧0, (14)
and subject to the following constraint for the unlabeled data
∀t′:∫dΘdγdbp(Θ)p(γ)p(b)((ΘX_{t′}−b))^{2}−γ_{t′})≧0, (15)
where the ΘX_{t }is the dot product between the separating hyperplane's weight vector and the tth data point's feature vector. No prior distribution over labels is necessary. The labeled data are constrained to be on the right side of the separating hyperplane according to their known labels, whereas the only requirement for the unlabeled data is that their squared distance to the hyperplane is greater than the margin. In summary this embodiment of the present invention finds a separating hyperplane that is a compromise of being closest to the chosen prior distribution, separating the labeled data correctly, and having no unlabeled data between the margins. The advantage is that no prior distribution over labels has to be introduced, thus, avoiding the problem of exponentially growing label assignments.
In specific implementation of the alternate embodiment of the present invention, using the prior distributions given in the Eqs. 7, 8, and 9 for the hyperplane parameters, the bias, and the margins yields the following partition function
where subscript t is the index of the labeled data and t′ the index of the unlabeled data. Introducing the notation
Eq. 16 can be rewritten as follows
yielding, after integration, the following partition function
i.e. the final objective function is
The objective function ℑ can be solved by applying similar techniques as in the case of known labels as discussed in the section herein entitled M Step. The difference is that matrix G_{3}^{−1 }in the quadratic form of the maximum margin term has now offdiagonal terms.
There exist many applications of method of the present invention employing Maximum Entropy Discrimination framework besides classification. For example MED can be applied to solve classification of data, in general, any kind of discriminant function and prior distributions, regression and graphical models (T. Jebara, Machine Learning Discriminative and Generative, Kluwer Academic Publishers) (Jebara).
The applications of the embodiments of the present invention can be formulated as pure inductive learning problems with known labels as well as a transductive learning problem with labeled as well as unlabeled training examples. In the latter case, the improvements to the transductive MED classification algorithm described in Table 3 are applicable as well to general transductive MED classification, transductive MED regression, transductive MED learning of graphical models. As such, for purposes of this disclosure and the accompanying claims, the word “classification” may include regression or graphical models.
M StepAccording to Eq. 11, the M step's objective function is
whereby the Lagrange Multipliers λ_{t }are determined by maximizing J_{M}.
Omitting the redundant constraint that λ_{t}<c, the Lagrangian for the dual problem above is
The KKT conditions, which are necessary and sufficient for optimality, are
whereby F is
At optimum, the basis equals the expected bias v=σ_{b}^{2}Σ_{t}λ_{t}y_{t}+μ_{b }yielding
y_{t}(−F_{t}−b)+δ_{t}=0 (25)
These equations can be summarized by considering two cases using the δ_{t}λ_{t}=0 constraint. The first case for all λ_{t}=0, and second for all 0<λ_{t}<c. There is no need for the third case as described in S. Keerthi, S. Shevade, C. Bhattacharyya, and K. Murthy, Improvements to platt's smo algorithm for sym classifier design, 1999 (Keerthi), applied to the SVM algorithm; the potential function in this formulation maintains that λ_{t}≠c.
λ_{t}=0,δ_{t}≧0(F_{t}+b)y_{t}≧0 (26)
0<λ_{t}<c,δ_{t}=0(F_{t}+b)=0 (27)
Until the optimum is reached violations of these conditions for some data point t will be present. Namely, F_{t}≠−b when λ_{t }is nonzero or F_{t}y_{t}<−by_{t} when it is zero. Unfortunately, calculating b is impossible without the optimum λ_{t}'s. A good solution to this is borrowed from Keerthi, referenced herein again, by constructing the following three sets.
I_{0}={t:0<λ_{t}<c} (28)
I_{1}={t:y_{t}>0,λ_{t}=0} (29)
I_{4}={t:y_{t}<0,λ_{t}=0} (30)
Using these sets we can define the most extreme violations of the optimality conditions using the following definitions. The elements in I_{0 }are violations whenever they are not equal to −b, therefore, the largest and smallest F_{t }from I_{0 }are candidates for being violations. The elements in I_{1 }are violations when F_{t}<−b so the smallest element from I_{1 }is the most extreme violation if one exists. Lastly, the elements in I_{4 }are violations when F_{t}>−b, which makes the largest elements from I_{4 }violation candidates. Therefore, −b is bounded by the min and max over these sets as shown below.
Due to the fact that at optimum −b_{up }and −b_{low }must be equal, namely −b, then reducing the gap between −b_{up }and −b_{low }will push the training algorithm to convergence. Additionally, the gap can also be measured as a way to determine numerical convergence.
As previously stated, the value of b=b is not known until convergence. The method of this alternate embodiment differs in that only one example can be optimized at a time. Therefore the training heuristic is to alternate between the examples in I_{0 }and all of the examples every other time.
E StepThe E step's objective function of Eq. 12 is
whereby s_{t }is the tth datapoint's classification score determined in the previous M step. The Lagrange Multipliers λ_{t }are determined by maximizing ℑ_{E}.
Omitting the redundant constraint that λ_{t}<c, the Lagrangian for the dual problem above is:
The KKT conditions, which are necessary and sufficient for optimality, are
Solving for the Lagrange5 multipliers by optimizing the KKT conditions can be done in one pass over the exampled since they factorize over the examples.
For labeled examples the expected label y_{t} is one with P_{0,t}(y_{t})=1 and P_{0,t}(−y_{t})=0 reducing the KKT conditions to
and yielding as solutions for the Lagrange Multipliers of labeled examples
For unlabeled examples, Eq. 35 cannot be solved analytically, but has to be determined by applying e.g. a linear search for each unlabeled example's Lagrange Multiplier that satisfies Eq. 35.
The following are several nonlimiting examples that are enabled by the techniques illustrated above, derivations or variations thereof, and other techniques known in the art. Each example includes the preferred operations, along with optional operations or parameters that may be implemented in the basic preferred methodology.
In one embodiment, as presented in
Further, at step 1006 a transductive MED classifier is trained through iterative calculation using said at least one cost factor and the labeled data points and the unlabeled data points as training examples. For each iteration of the calculations, the unlabeled data point cost factor is adjusted as a function of an expected label value, e.g. the absolute value of the expected label of a data point, etc., and a data point label prior probability is adjusted according to an estimate of a data point class membership probability, thereby ensuring stability. Also, the transductive classifier may learn using prior probability information of the labeled and unlabeled data, which further improves stability. The iterative step of training a transductive classifier may be repeated until the convergence of data values is reached, e.g. when the change of the decision function of the transductive classifier falls below a predetermined threshold value, when the change of the determined expected label value falls below a predetermined threshold value, etc.
Additionally, in step 1008 the trained classifier is applied to classify at least one of the unlabeled data points, the labeled data points, and input data points. Input data points may be received before or after the classifier is trained, or may not be received at all. Also, the decision function that minimizes the KL divergence to the prior probability distribution of the decision function parameters given the included and excluded training examples may be determined utilizing the labeled as well as the unlabeled data points as learning examples according to their expected label. Alternatively, the decision function may be determined with minimal KL divergence using a multinomial distribution for the decision function parameters.
In step 1010 a classification of the classified data points, or a derivative thereof, is output to at least one of a user, another system, and another process. The system may be remote or local. Examples of the derivative of the classification may be, but are not limited to, the classified data points themselves, a representation or identifier of the classified data points or host file/document, etc.
In another embodiment, computer executable program code is deployed to and executed on a computer system. This program code comprises instructions for accessing stored labeled data points in a memory of a computer, where each of said labeled data points has at least one label indicating whether the data point is a training example for data points for being included in a designated category or a training example for data points being excluded from a designated category. In addition, the computer code comprises instructions for accessing unlabeled data points from a memory of a computer as well as accessing at least one predetermined cost factor of the labeled data points and unlabeled data points from a memory of a computer. Prior probability information of labeled and unlabeled data points stored in a memory of a computer may also be accessed. Also, the label of the included training example may be mapped to a first numeric value, e.g. +1, etc., and the label of the excluded training example may be mapped to a second numeric value, e.g. −1, etc.
Further, the program code comprises instructions for training a transductive classifier through iterative calculation, using the at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples. Also, for each iteration of the calculation, the unlabeled data point cost factor is adjusted as a function of the expected label value of the data point, e.g. the absolute value of the expected label of a data point. Also, for each iteration, the prior probability information may be adjusted according to an estimate of a data point class membership probability. The iterative step of training a transductive classifier may be repeated until the convergence of data values is reached, e.g. when the change of the decision function of the transductive classifier falls below a predetermined threshold value, when the change of the determined expected label value falls below a predetermined threshold value, etc.
Additionally, the program code comprises instructions for applying the trained classifier to classify at least one of the unlabeled data points, the labeled data points, and input data points, as well as instructions for outputting a classification of the classified data points, or derivative thereof, to at least one of a user, another system, and another process. Also, the decision function that minimizes the KL divergence to the prior probability distribution of the decision function parameters given the included and excluded training examples may be determined utilizing the labeled as well as the unlabeled data as learning examples according to their expected label.
In yet another embodiment, a data processing apparatus comprises at least one memory for storing: (i) labeled data points, wherein each of said labeled data points have at least one label indicating whether the data point is a training example for data points being included in a designated category or a training example for data points being excluded from a designated category; (ii) unlabeled data points, and (iii) at least one predetermined cost factor of the labeled data points and unlabeled data points. The memory may also store prior probability information of labeled and unlabeled data points. Also, the label of the included training example may be mapped to a first numeric value, e.g. +1, etc., and the label of the excluded training example may be mapped to a second numeric value, e.g. −1, etc.
In addition, the data processing apparatus comprises a transductive classifier trainer to iteratively teach the transductive classifier using transductive Maximum Entropy Discrimination (MED) using the at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples. Further, at each iteration of the MED calculation the cost factor of the unlabeled data point is adjusted as a function of the expected label value of the data point, e.g. the absolute value of the expected label of a data point, etc. Also, at each iteration of the MED calculation, the prior probability information may be adjusted according to an estimate of a data point class membership probability. The apparatus may further comprise a means for determining the convergence of data values, e.g. when the change of the decision function of the transductive classifier calculation falls below a predetermined threshold value, when the change of the determined expected label values falls below a predetermined threshold value, etc., and terminating calculations upon determination of convergence.
In addition, a trained classifier is used to classify at least one of the unlabeled data points, the labeled data points, and input data points. Further, the decision function that minimizes the KL divergence to the prior probability distribution of the decision function parameters given the included and excluded training examples may be determined by a processor utilizing the labeled as well as the unlabeled data as learning examples according to their expected label. Also, a classification of the classified data points, or derivative thereof, is output to at least one of a user, another system, and another process.
In a further embodiment, an article of manufacture comprises a program storage medium readable by a computer, where the medium tangibly embodies one or more programs of instructions executable by a computer to perform a method of data classification. In use, labeled data points are received, where each of the labeled data points has at least one label which indicates whether the data point is a training example for data points for being included in a designated category or a training example for data points being excluded from a designated category. In addition, unlabeled data points are received, as well as at least one predetermined cost factor of the labeled data points and unlabeled data points. Prior probability information of labeled and unlabeled data points may also be stored in a memory of a computer. Also, the label of the included training example may be mapped to a first numeric value, e.g. +1, etc., and the label of the excluded training example may be mapped to a second numeric value, e.g. −1, etc.
Further, a transductive classifier is trained with iterative Maximum Entropy Discrimination (MED) calculation using the at least one stored cost factor and the stored labeled data points and the unlabeled data points as training examples. At each iteration of the MED calculation, the unlabeled data point cost factor is adjusted as a function of an expected label value of the data point, e.g. the absolute value of the expected label of a data point, etc. Also, at each iteration of the MED calculation, the prior probability information may be adjusted according to an estimate of a data point class membership probability. The iterative step of training a transductive classifier may be repeated until the convergence of data values is reached, e.g. when the change of the decision function of the transductive classifier falls below a predetermined threshold value, when the change of the determined expected label value falls below a predetermined threshold value, etc.
Additionally, input data points are accessed from the memory of a computer, and the trained classifier is applied to classify at least one of the unlabeled data points, the labeled data points, and input data points. Also, the decision function that minimizes the KL divergence to the prior probability distribution of the decision function parameters given the included and excluded training examples may be determined utilizing the labeled as well as the unlabeled data as learning examples according to their expected label. Further, a classification of the classified data points, or a derivative thereof, is output to at least one of a user, another system, and another process.
In yet another embodiment, a method for classification of unlabeled data in a computerbased system is presented. In use, labeled data points are received, each of said labeled data points having at least one label indicating whether the data point is a training example for data points for being included in a designated category or a training example for data points being excluded from a designated category.
Additionally, labeled and unlabeled data points are received, as are prior label probability information of labeled data points and unlabeled data points. Further, at least one predetermined cost factor of the labeled data points and unlabeled data points is received.
Further, the expected labels for each labeled and unlabeled data point are determined according to the label prior probability of the data point. The following substeps are repeated until substantial convergence of data values:

 generating a scaled cost value for each unlabeled data point proportional to the absolute value of the data point's expected label;
 training a Maximum Entropy Discrimination (MED) classifier by determining the decision function that minimizes the KL divergence to the prior probability distribution of the decision function parameters given the included training and excluded training examples utilizing the labeled as well as the unlabeled data as training examples according to their expected label;
 determining the classification scores of the labeled and unlabeled data points using the trained classifier;
 calibrating the output of the trained classifier to class membership probability;
 updating the label prior probabilities of the unlabeled data points according to the determined class membership probabilities;
 determining the label and margin probability distributions using Maximum Entropy Discrimination (MED) using the updated label prior probabilities and the previously determined classification scores;
 computing new expected labels using the previously determined label probability distribution; and
 updating expected labels for each data point by interpolating the new expected labels with the expected label of previous iteration.
Also, a classification of the input data points, or derivative thereof, is output to at least one of a user, another system, and another process.
Convergence may be reached when the change of the decision function falls below a predetermined threshold value. Additionally, convergence may also be reached when the change of the determined expected label value falls below a predetermined threshold value. Further, the label of the included training example may have any value, for example, a value of +1, and the label of the excluded training example may have any value, for example, a value of −1.
In one embodiment of the present invention, a method for classifying documents is presented in
Additionally, after at least some of the iterations, in step 1104 confidence scores are stored for the unlabeled documents, and identifiers of the unlabeled documents having the highest confidence scores are output in step 1106 to at least one of a user, another system, and another process. The identifiers may be electronic copies of the document themselves, portions thereof, titles thereof, names thereof, file names thereof, pointers to the documents, etc. Also, confidence scores may be stored after each of the iterations, wherein an identifier of the unlabeled document having the highest confidence score after each iteration is output.
One embodiment of the present invention is capable of discovering patterns that link the initial document to the remaining documents. The task of discovery is one area where this pattern discovery proves particularly valuable. For instance, in pretrial legal discovery, a large amount of documents have to be researched with regard to possible connections to the lawsuit at hand.
The ultimate goal is to find the “smoking gun.” In another example, a common task for inventors, patent examiners, as well as patent lawyers is to evaluate the novelty of a technology through prior art search. In particular the task is to search all published patents and other publications and find documents within this set that might be related to the specific technology that is examined with regard to its novelty.
The task of discovery involves finding a document or a set of documents within a set of data. Given an initial document or concept, a user may want to discover documents that are related to the initial document or concept. However, the notion of relationship between the initial document or concept and the target documents, i.e. the documents that are to be discovered, is only well understood after the discovery has taken place. By learning from labeled and unlabeled documents, concepts, etc., the present invention can learn patterns and relationships between the initial document or documents and the target documents.
In another embodiment of the present invention, a method for analyzing documents associated with legal discovery is presented in
The document classification technique may include any type of process, e.g. a transductive process, etc. For example, any inductive or transductive technique described above may be used. In a preferred approach, a transductive classifier is trained through iterative calculation using at least one predetermined cost factor, at least one seed document, and the documents associated with the legal matter. For each iteration of the calculations the cost factor is preferably adjusted as a function of an expected label value, and the trained classifier is used to classify the received documents. This process may further comprise receiving a data point label prior probability for the labeled and unlabeled documents, wherein for each iteration of the calculations the data point label prior probability is adjusted according to an estimate of a data point class membership probability. Additionally, the document classification technique may include one or more of a support vector machine process and a maximum entropy discrimination process.
In yet another embodiment, a method for analyzing prior art documents is presented in
The search query may include at least a portion of a patent disclosure. Illustrative patent disclosures include a disclosure created by an inventor summarizing the invention, a provisional patent application, a nonprovisional patent application, a foreign patent or patent application, etc.
In one preferred approach, the search query includes at least a portion of a claim from a patent or patent application. In another approach, the search query includes at least a portion of an abstract of a patent or patent application. In a further approach, the search query includes at least a portion of a summary from a patent or patent application.
An embodiment of the present invention may be used for the classification of patent applications. In the United States, for example, patents and patent applications are currently classified by subject matter using the United States Patent Classification (USPC) system. This task is currently performed manually, and therefore is very expensive and time consuming. Such manual classification is also subject to human errors. Compounding the complexity of such a task is that the patent or patent application may be classified into multiple classes.
The document classification technique is preferably a yes/no classification technique. In other words, if the probability that the document is in the proper class is above a threshold, the decision is yes, the document belongs in this class. If the probability that the documents is in the proper class is below a threshold, the decision is no, the document does not belong in this class.
In either of the methods shown in
Officially, classification of a patent should be based on the claims. However, it may also be desirable to perform matching between (any IP related content) and (any IP related content). As an example, one approach uses the Description of a patent to train, and classify an application based on its Claims. Another approach uses the Description and Claims to train, and classify based on the Abstract. In particularly preferred approaches, whatever portion of a patent or application is used to train, that same type of content is used when classifying, i.e., if the system is trained on claims, the classification is based on claims.
The document classification technique may include any type of process, e.g. a transductive process, etc. For example, any inductive or transductive technique described above may be used. In a preferred approach, the classifier may be a transductive classifier, and the transductive classifier may be trained through iterative calculation using at least one predetermined cost factor, at least one seed document, and the prior art documents, wherein for each iteration of the calculations the cost factor is adjusted as a function of an expected label value, and the trained classifier may be used to classify the prior art documents. A data point label prior probability for the seed document and prior art documents may also be received, wherein for each iteration of the calculations the data point label prior probability may be adjusted according to an estimate of a data point class membership probability. The seed document may be any document, e.g. publications of a patent office, data retrieved from a database, a collection of prior art, a website, a patent disclosure, etc.
In one approach,
Another embodiment of the present invention involves data cleanup and accurate classification, for example in conjunction with the automation of business processes. The cleanup and classification technique may include any type of process, e.g. a transductive process, etc. For example, any inductive or transductive technique described above may be used. In a preferred approach, the keys of the entries in the database are utilized as labels associated with some confidence level according to the expected cleanliness of the database. The labels together with the associated confidence level, i.e. the expected labels, are then used to train a transductive classifier that corrects the labels (keys) in order to achieve a more consistent organization of the data in the database. For example, invoices have to be first classified according to the company or person that originated the invoice in order to enable automatic data extraction, e.g. the determination of total dollar amount, purchase order number, product amount, shipping address, etc. Commonly, training examples are needed to set up an automatic classification system. However, training examples provided by the customer often contain misclassified documents or other noise—e.g. fax cover sheets—that have to be identified and removed prior to training the automatic classification system in order to obtain accurate classification. In another example, in the area of patient records, it is useful to detect inconsistencies between the report written by the physician and the diagnosis.
In another example, it is known that the Patent Office undergoes a continuous reclassification process, in which they (1) evaluate an existing branch of their taxonomy for confusion, (2) restructure that taxonomy to evenly distributed overly congested nodes, and (3) reclassify existing patents into the new structure. The transductive learning methods presented herein may be used by the Patent Office, and the companies they outsource to do this work, to reevaluate their taxonomy, and assist them in (1) build a new taxonomy for a given main classification, and (2) reclassifying existing patents.
Transduction learns from labeled and unlabeled data, whereby the transition from labeled to unlabeled data is fluent. At one end of the spectrum are labeled data with perfect prior knowledge, i.e. the given labels are correct with no exceptions. At the other end are unlabeled data where no prior knowledge is given. Organized data with some level of noise constitute mislabeled data and are located somewhere on the spectrum between these two extremes: The labels given by the organization of the data can be trusted to be correct to some extent but not fully. Accordingly, transduction can be utilized to clean up the existing organization of data by assuming a certain level of mistakes within the given organization of the data and interpreting these as uncertainties in the prior knowledge of label assignments.
In one embodiment, a method for cleaning up data is presented in
Further, the subsets may be selected at random and may be selected and verified by a user. The label of at least some of the data items may be changed based on the classification. Also, identifiers of data items having a confidence level below a predefined threshold after classification thereof may be output to a user. The identifiers may be electronic copies of the document themselves, portions thereof, titles thereof, names thereof, file names thereof, pointers to the documents, etc.
In one embodiment of the present invention, as illustrated in
In another embodiment, a method for managing medical records is presented in
In one embodiment, the classifier may be a transductive classifier, and the transductive classifier may be trained through iterative calculation using at least one predetermined cost factor, at least one seed document, and the medical records, wherein for each iteration of the calculations the cost factor is adjusted as a function of an expected label value, and the trained classifier may be used to classify the medical records. A data point label prior probability for the seed document and medical records may also be received, wherein for each iteration of the calculations the data point label prior probability may be adjusted according to an estimate of a data point class membership probability.
Another embodiment of the present invention accounts for dynamic, shifting classification concepts. For example, in forms processing applications documents are classified using the layout information and/or the content information of the documents to classify the documents for further processing. In many applications the documents are not static but evolve over time. For example the content and/or layout of a document may change owing to new legislation. Transductive classification adapts to these changes automatically yielding the same or comparable classification accuracy despite the drifting classification concepts. This is in contrast to rule based systems or inductive classification methods that, without manually adjustments, will start to suffer in classification accuracy owing to the concept drift. One example of this is invoice processing, which traditionally involves inductive learning, or rulebased systems are used that utilize invoice layout. Under these traditional systems, if a change in the layout occurs the systems have to be manually reconfigured by either labeling new training data or by determining new rules. However, the use of transduction makes the manual reconfiguration unnecessary by automatically adapting to the small changes in layout of the invoices. In another example, transductive classification may be applied to the analysis of customer complaints in order to monitor the changing nature of such complaints. For example, a company can automatically link product changes with customer complaints.
Transduction may also be used in the classification of news articles. For example, news articles on the war on terror starting with articles about the terrorist attacks on Sep. 11, 2001 over the war in Afghanistan to news stories about the situation in today's Iraq can be automatically identified using transduction.
In yet another example, the classification of organisms (alpha taxonomy) can change over time through evolution by creating new species of organisms and other species becoming extinct. This and other principles of a classification schema or taxonomy can be dynamic, with classification concepts shifting or changing over time.
By using the incoming data that have to be classified as unlabeled data, transduction can recognize shifting classification concepts, and therefore dynamically adapt to the evolving classification schema. For example,
In yet another embodiment, a method for adapting to a shift in document content is presented in
In addition, an unlabeled document having a confidence level below the predefined threshold may be moved into one or more new categories. Also, the transductive classifier may be trained through iterative calculation using at least one predetermined cost factor, the at least one seed document, and the unlabeled documents, wherein for each iteration of the calculations the cost factor may be adjusted as a function of an expected label value, and using the trained classifier to classify the unlabeled documents. Further, a data point label prior probability for the seed document and unlabeled documents may be received, wherein for each iteration of the calculations the data point label prior probability may be adjusted according to an estimate of a data point class membership probability.
In another embodiment, a method for adapting a patent classification to a shift in document content is presented in
Also, in step 2004 the unlabeled documents having a confidence level below the predefined threshold are classified into at least one new category using the classifier, and at least some of the categorized documents are reclassified in step 2006 into the existing categories and the at least one new category using the classifier. Further, identifiers of the categorized documents are output in step 2008 to at least one of a user, another system, and another process. Also, the transductive classifier may be trained through iterative calculation using at least one predetermined cost factor, the search query, and the documents, wherein for each iteration of the calculations the cost factor may be adjusted as a function of an expected label value, and the trained classifier may be used to classify the documents. Further, a data point label prior probability for the search query and documents may be received, wherein for each iteration of the calculations the data point label prior probability is adjusted according to an estimate of a data point class membership probability.
Yet another embodiment of the present invention accounts for document drift in the field of document separation. One use case for Document separation involves the processing of mortgage documents. Loan folders consisting of a sequence of different loan documents, e.g. loan applications, approvals, requests, amounts, etc. are scanned and the different documents within the sequence of images have to be determined before further processing. The documents used are not static but can change over time. For example, tax forms used within a loan folder can change over time owing to legislation changes.
Document separation solves the problem of finding document or subdocument boundaries in a sequence of images. Common examples that produce a sequence of images are digital scanners or Multi Functional Peripherals (MFPs). As in the case of classification, transduction can be utilized in Document separation in order to handle the drift of documents and their boundaries over time. Static separation systems like rule based systems or systems based on inductive learning solutions cannot adapt automatically to drifting separation concepts. The performance of these static separation systems degrade over time whenever a drift occurs. In order to keep the performance on its initial level, one either has to manually adapt the rules (in the case of a rule based system), or has to manually label new documents and relearn the system (in case of an inductive learning solution). Either way is time and cost expensive. Applying transduction to Document separation allows the development of a system that automatically adapts to the drift in the separation concepts.
In one embodiment, a method for separating documents is presented in
Additionally, the system as shown in
Yet another embodiment of the present invention is able to perform face recognition using transduction. As mentioned above, the use of transduction has many advantages, for example the need of a relatively small number of training examples, the ability to use unlabeled examples in training, etc. By making use of the aforementioned advantages, transductive face recognition may be implemented for criminal detection.
For example, the Department of Homeland Security must ensure that terrorists are not allowed onto commercial airliners. Part of an airport's screening process may be to take a picture of each passenger at the airport security checkpoint and attempt to recognize that person. The system could initially be trained using a small number of examples from the limited photographs available of possible terrorists. There may also be more unlabeled photographs of the same terrorist available in other lawenforcement databases that may also be used in training. Thus, a transductive trainer would take advantage of not only the initially sparse data to create a functional facerecognition system but would also use unlabeled examples from other sources to increase performance. After processing the photograph taken at the airport security checkpoint, the transductive system would be able to recognize the person in question more accurately than a comparable inductive system.
In yet another embodiment, a method for face recognition is presented in
Further, in step 2406 identifiers of the unlabeled documents having the highest confidence scores are output to at least one of a user, another system, and another process. The identifiers may be electronic copies of the document themselves, portions thereof, titles thereof, names thereof, file names thereof, pointers to the documents, etc. Also, confidence scores may be stored after each of the iterations, wherein an identifier of the unlabeled images having the highest confidence score after each iteration is output. Additionally, a data point label prior probability for the labeled and unlabeled image may be received, wherein for each iteration of the calculations the data point label prior probability may be adjusted according to an estimate of a data point class membership probability. Further, a third unlabeled image of a face, e.g., from the above airport security example, may be received, the third unlabeled image may be compared to at least some of the images having the highest confidence scores, and an identifier of the third unlabeled image may be output if a confidence that the face in the third unlabeled image is the same as the face in the seed image.
Yet another embodiment of the present invention enables a user to improve their search results by providing feedback to the document discovery system. For example, when performing a search on an internet search engine, patent or patent application search product, etc., users may get a multitude of results in response to their search query. An embodiment of the present invention enables the user to review the suggested results from the search engine and inform the engine of the relevance of one or more of the retrieved results, e.g. “close, but not exactly what I wanted,” “definitely not,” etc. As the user provides feedback to the engine, better results are prioritized for the user to review.
In one embodiment, a method for document searching is presented in
The document classification technique may include any type of process, e.g. a transductive process, a support vector machine process, a maximum entropy discrimination process, etc. Any inductive or transductive technique described above may be used. In a preferred approach, the classifier may be a transductive classifier, and the transductive classifier may be trained through iterative calculation using at least one predetermined cost factor, the search query, and the documents, wherein for each iteration of the calculations the cost factor may be adjusted as a function of an expected label value, and the trained classifier may be used to classify the documents. In addition, a data point label prior probability for the search query and documents may be received, wherein for each iteration of the calculations the data point label prior probability may be adjusted according to an estimate of a data point class membership probability.
A further embodiment of the present invention may be used for improving ICR/OCR, and speech recognition. For example, many embodiments of speech recognition programs and systems require the operator to repeat a number of words to train the system. The present invention can initially monitor the voice of a user for a preset period of time to gather “unclassified” content, e.g., by listening in to phone conversations. As a result, when the user starts training the recognition system, the system utilizes transductive learning to utilize the monitored speech to assist in building a memory model.
In yet another embodiment, a method for verifying an association of an invoice with an entity is presented in
Further, the classifier may be any type of classifier, for example, a transductive classifier, and the transductive classifier may be trained through iterative calculation using at least one predetermined cost factor, at least one seed document, and the invoices, wherein for each iteration of the calculations the cost factor is adjusted as a function of an expected label value, and using the trained classifier to classify the invoices. Also, a data point label prior probability for the seed document and invoices may be received, wherein for each iteration of the calculations the data point label prior probability is adjusted according to an estimate of a data point class membership probability.
One of the benefits afforded by the embodiments depicted herein is the stability of the transductive algorithm. This stability is achieved by scaling the cost factors and adjusting the label prior probability. For example, in one embodiment a transductive classifier is trained through iterative classification using at least one cost factor, the labeled data points, and the unlabeled data points as training examples. For each iteration of the calculations, the unlabeled date point cost factor is adjusted as a function of an expected label value. Additionally, for each iteration of the calculations the data point label prior probability is adjusted according to an estimate of a data point class membership probability.
The workstation may have resident thereon an operating system such as the Microsoft Windows® Operating System (OS), a MAC OS, or UNIX operating system. It will be appreciated that a preferred embodiment may also be implemented on platforms and operating systems other than those mentioned. A preferred embodiment may be written using JAVA, XML, C, and/or C++ language, or other programming languages, along with an object oriented programming methodology. Object oriented programming (OOP), which has become increasingly used to develop complex applications, may be used.
The above application uses transductive learning to overcome the problem of very sparse data sets which plague inductive facerecognition systems. This aspect of transductive learning is not limited to this application and may be used to solve other machinelearning problems that arise from sparse data.
Those skilled in the art could devise variations that are within the scope and spirit of the various embodiments of the invention disclosed herein. Further, the various features of the embodiments disclosed herein can be used alone, or in varying combinations with each other and are not intended to be limited to the specific combination described herein. Thus, the scope of the claims is not to be limited by the illustrated embodiments.
Claims
1. A system for verifying an association of an invoice with an entity, comprising:
 a memory; and
 a processor in communication with the memory, the processor being configured to process at least some instructions stored in the memory,
 wherein the memory stores computer executable program code comprising instructions for: training a classifier based on an invoice format associated with a first entity; accessing a plurality of invoices labeled as being associated with at least one of the first entity and other entities; and outputting an identifier of at least one of the invoices having a high probability of not being associated with the first entity.
2. The product of claim 1, wherein the invoice format includes a physical layout of markings on the invoice.
3. The system of claim 1, further comprising: performing a document classification technique on the invoices using the classifier.
4. The system of claim 3, wherein the document classification technique includes a transductive process.
5. The system of claim 4, wherein the invoice format includes a physical layout of markings on the invoice.
6. The system of claim 3, wherein the document classification technique includes a support vector machine process.
7. The system of claim 6, wherein the invoice format includes a physical layout of markings on the invoice.
8. The system of claim 3, wherein the document classification technique includes a maximum entropy discrimination process.
9. The system of claim 8, wherein the invoice format includes a physical layout of markings on the invoice.
10. A computer program product comprising a compute readable storage medium having stored thereon computer executable program code comprising instructions for:
 training a classifier based on an invoice format associated with a first entity;
 accessing a plurality of invoices labeled as being associated with at least one of the first entity and other entities; and
 outputting an identifier of at least one of the invoices having a high probability of not being associated with the first entity.
11. The computer program product of claim 10, wherein the invoice format includes a physical layout of markings on the invoice.
12. The computer program product of claim 10, further comprising computer executable program code comprising instructions for: performing a document classification technique on the invoices using the classifier.
13. The computer program product of claim 12, wherein the document classification technique includes a transductive process.
14. The computer program product of claim 13, wherein the invoice format includes a physical layout of markings on the invoice.
15. The computer program product of claim 12, wherein the document classification technique includes a support vector machine process.
16. The computer program product of claim 15, wherein the invoice format includes a physical layout of markings on the invoice.
17. The computer program product of claim 12, wherein the document classification technique includes a maximum entropy discrimination process.
18. The computer program product of claim 17, wherein the invoice format includes a physical layout of markings on the invoice.
19. A system for managing medical records, comprising:
 a memory; and
 a processor in communication with the memory, the processor being configured to process at least some instructions stored in the memory,
 wherein the memory stores computer executable program code comprising instructions for: training a classifier based on a medical diagnosis; accessing a plurality of medical records; and outputting an identifier of at least one of the medical records having a low probability of being associated with the medical diagnosis.
20. The system of claim 19, the memory further comprising computer executable program code comprising instructions for performing a document classification technique on the medical records using the classifier.
21. The system of claim 20, wherein the document classification technique includes a transductive process.
22. The system of claim 21, wherein the invoice format includes a physical layout of markings on at least one of the medical records.
23. The system of claim 20, wherein the document classification technique includes a support vector machine process.
24. The system of claim 23, wherein the invoice format includes a physical layout of markings on at least one of the medical records.
25. The system of claim 20, wherein the document classification technique includes a maximum entropy discrimination process.
Type: Application
Filed: Mar 25, 2014
Publication Date: Jul 24, 2014
Applicant: Kofax, Inc. (Irvine, CA)
Inventors: Mauritius A.R. Schmidtler (Escondido, CA), Roland G. Borrey (Villa Park, CA), Anthony Sarah (San Diego, CA)
Application Number: 14/225,298