Determining Utility Functions from Ordinal Rankings
A user's preference structure in respect of alternative “objects” with which the user is presented is captured in a multi-attribute utility function. The user ranks these competing objects in order of the user's relative preference for such objects. A utility function that defines the user's preference structure is provided as output on the basis of this relative ranking. This technique can be used to assist a buyer in selecting between multi-attribute quotes or bids submitted by prospective suppliers to the buyer.
Latest IBM Patents:
- INTERACTIVE DATASET EXPLORATION AND PREPROCESSING
- NETWORK SECURITY ASSESSMENT BASED UPON IDENTIFICATION OF AN ADVERSARY
- NON-LINEAR APPROXIMATION ROBUST TO INPUT RANGE OF HOMOMORPHIC ENCRYPTION ANALYTICS
- Back-side memory element with local memory select transistor
- Injection molded solder head with improved sealing performance
This Application is a division of U.S. patent application Ser. No. 10/308,401, filed Dec. 3, 2002.
FIELD OF THE INVENTIONThe present invention relates to determining utility functions for use in determining or representing user preferences.
BACKGROUNDAuction software often allows bidders (for example, suppliers) to specify multiple attributes associated with their bids, such as quality, terms and conditions, etc. Consequently, bidders can differentiate themselves by factors other than price.
Weighting factors enable buyers to rate the relative importance of attributes that may be associated with bids. A buyer can specify weighting factors for each attribute using a sliding scale, in which the user can choose from various options such as “do not care”, “important”, “more important”, “very important”, and “most important”.
There also exists software for performing “Request for Quotes” (RFQ). Some software permits complex configuration and bill-of-material relationships. Suppliers can specify quotes having ranges of attributes, and price variations for terms and conditions. In some cases, a buyer can post a RFQ, and the software automatically (i) searches supplier's rules (predefined and stored in a database), (ii) calculates a best offer for each supplier, and (iii) ranks all offers according to criteria that are most important to a buyer.
In some existing RFQ software, buyers (that is, users) can specify the weighting factors of each attribute explicitly. However, these softwares demand a thorough knowledge/perception about the attributes from a user. On the other hand, rest of the software accept input from a user about the order (partial or full) on the items or bids and extract the weighting factors (utilities) of the attributes from the specified order.
In software that accepts an ordered list (full or partial) of items/bids from a user, the utility function of the item/bid is expressed as a parametric function of the attributes. There exist a number of investigations concerning techniques for representing the utility function of the bids/items having multiple attributes and subsequent techniques for the parametric model fitting on the ordinal data sets (that is, ordered set of items). In general, such an approach may be classified as belonging to multi-attribute utility theory (MAUT).
In multi-attribute utility theory (MAUT), parametric utility functions can be classified according to different levels of complexity. Types of parametric utility function include multilinear, multiplicative, and additive models. Solutions to the problem of evaluating attributes are proposed for additive models only, the utilities for individual attributes are considered to be independent. For additive utility functions, evaluation of the utility function, and subsequent ranking of the objects, is performed by assessing the weights of the attributes by formulating the problem as a linear programming task.
There are two primary limitations associated with the existing tools described above.
First, users' preference structures are represented by what is essentially a linear additive utility function, namely a weighted sum of utility functions for individual attributes. For example, if attributes associated with a bid are price and quality, then a relevant utility function for the buyer, U(price, quality), is defined as U(price, quality)=w1′U(price)+w2′U(quality). In this expression, U(price) and U(quality) are respectively the buyer's individual utility functions for price and quality.
Second, individual attribute utility functions are assumed a priori. In the example above, U(price) and U(quality) are assumed to be known, and the buyer effectively specifies weights w1 and w2 by indicating the relative importance of the respective attributes on a sliding scale, such as that described above.
In view of the above observations, a need clearly exists for representing a user's preferences when selecting between competing alternatives.
SUMMARYA technique for determining a utility function having multiple attributes is described herein, assuming that there exists a partly or fully ranked set of objects. Nonlinear interactions between individual attributes are assumed, and there is also assumed to be no a priori knowledge about individual attribute utility functions.
Objects (partially or fully ordered) having multiple attributes are accepted as input, and a multi-attribute utility function is provided as output. The ordered set of objects can either be directly obtained from a user, or determined from past interactions with the user or observed behavior of the user. This multi-attribute utility function is essentially “learned” from the ordered set of objects. A technique for learning a utility function using a feed-forward neural network is described. The neural network iteratively learns the utility function from pairs of items in the ranked list.
The techniques described herein are presented in the context of a buyer selecting between alternative bids or quotes submitted by prospective suppliers. The described techniques, however, apply more generally to any user's preference structure in respect of alternative “objects” with which the user is presented. The user ranks these competing objects in order of the user's relative preference for such objects. A utility function that defines the user's preference structure is provided as output on the basis of this relative ranking.
DESCRIPTION OF DRAWINGS
In “real-life” problems, a user may be considered to have an inherent preference structure involving objects having multiple attributes. An example of such a real-life problem is evaluating responses received, from suppliers, by a buyer for the buyer's RFQ. The buyer's preference structure is represented, mathematically, by a utility function that specifies these relevant attributes.
The responses specify multiple attributes (of interest to the buyer), and the buyer chooses from amongst favorable responses received from suppliers. Examples of attributes that may comprise a response to a RFQ are “price”, “quality”, “quantity”. A buyer's preference structure over the entire attribute space is likely to involve dependencies between different attributes.
For buyers, attributes “price” and “quality” are likely to be related in a buyer's preference structure. A buyer, for example, may prefer a quote having a higher price and higher quality equally as much as a response having a lower price and medium quality. Interdependencies of this sort between different attributes are quite likely to be nonlinear.
The described techniques can be used to learn a buyer's preference structure (represented as a utility function) over a space of multi-attributed responses. The buyer provides a set of ranked responses, as a training example, so that the utility function can be determined.
Table 1 below presents a prescriptive indication of an approximate number of nodes to be used in the hidden layer of the feed-forward neural network of
Techniques described herein are presented in the form of an algorithm, and each step of the algorithm is described in detail. Each object i is represented as an m-dimensional vector
The described algorithm, which can be conveniently referred to as “bid evaluation”, receives as input r, which is a ranked list of objects. The bid evaluation algorithm provides as output ƒ(.), which is a utility function.
The input list r, an either partly or fully ranked subset of items/bids, can be represented as a set of ordered pairs in the following manner:
r={(xi,xj): xi is preferred over xj by the user and provided as input to the system} a.
The Input Ranked List r consists of object pairs provided by the user, out of n(n−1)/2 possible pairs for n objects (|r|=[n(n−1)/2) specified by the system to the user. Each object pair indicates a transitive relationship of preference such that (a,b) indicates that a is preferred over b, and given (a,b) and (b,c), a is preferred over c, by implication.
Table 3 below presents pseudocode corresponding to the steps of Table 2 and
Each of the above-described steps is described in further detail below.
Obtaining a Ranked List—Step 210
A user can explicitly provide a list of ranked objects (list r), or the list can be implicitly obtained from observed behavior of the user, or from other transactions with the user. The objects are ranked in descending order of the user's preference for the respective objects, from more preferred to less preferred. The ranking is relative not absolute; that is, only the ranked objects are considered, rather than the total set of objects.
Initializing Network for Training—Step 220
A suitable architecture is first selected for a feed-forward network.
All weights of the links and thresholds of the nodes are initialized to small random values (for example, in the range specified by [−0.1, 0.1]). These seed values provide a basis for training the neural network.
Calculating Revised Parameter Weights—Step 230
Step 230 is performed for every pair of objects (xp, xq) in the ranked list r. Here (xp, xq) implies that (xp>xq) in the ranked list r. That is, xp is preferred over xq. The repeated calculation in step 230 corresponds to the iterative training steps of the neural network of
A pair of objects (xp, xq) is taken as an initial input. The neural network output is calculated for both objects in the pair individually, for computing the required change in network parameters as described below. To compute the neural network output for a particular object (for example, xp) the input to the neural network is the attribute vector xp, which represents the corresponding object. The corresponding output of the neural network is the resulting utility function ƒ(xp) generated by the neural network.
If ƒ(xp) is less than or equal to ƒ(xq), then the network is considered to have made an error. An error is made because the utility function ƒ(.) implied by the neural network effectively ranks xq above xp.
The error measure is expressed in Equation (1) below
In Equation (1), K is a constant between zero and one; that is, 0<K<1. A typical value of K may be, for example, 0.95. The overall error measure is expressed in Equation (2) below
The described algorithm then computes the required change in weights of the links connecting neurons between successive layers of the neural network, to reduce the error made by the neural network with the current weights (as given by the error measure Equation (2)). The parameters (that is, weights of the links connecting neurons between successive layers) are changed in the opposite direction of the error gradient (gradient descent) so that a movement by a small factor in the parameter space opposite to the gradient direction results in a decrease in the total error (on the pairs of training bids/items).
For the pair of objects (xp, xq), for every i, j, and l, the expression
is computed. This expression represents the required change in weight
of the link connecting the neuron i of layer (l−1) to neuron j of layer l of the neural network of
is computed in accordance with Equation (3) below, which applies if
for the objects
Otherwise, the expression
is equal to zero.
In Equation (3) above, K is a constant between zero and one; that is, 0<K<1. A typical value of K may be, for example, 0.95. K is constant which forces the network output such that
in the converged state where
when the network makes a mistake. The parameter h is a constant that is referred to as “learning rate”. This value, h, is computed in accordance with Equation (4) below.
In Equation (4) above, the value for
is provided by the expression of Equation (3) below.
Further,
and
are respective outputs of the jth neuron in layer l of the network for xp and xq as inputs to the neural network. The output of node j of layer l for an input xp is calculated in accordance with Equation (6) below.
In Equation (6) above,
is the total input to the jth node of layer l from the previous layer, given as
for an input node vj0(p)=xpj in which xp=(xp1, xp2, . . . xpm). Similarly, vjl(q) can be computed for an input xq, for different layers and nodes.
In the above Equations (3) to (6), djl(p) and djl(q) represent error that propagates backwards from the output layer to node j of layer l. In Equation (7) below, the error at a node i of layer l depends on the error of every node k of layer l+1 connected to node i. Thus, to compute the error at a particular node, the topmost layer, (that is the output layer) in first considered and then the errors are computed successively downwards. In other words, the error in the output layer propagates backward down to the input layer. Error value d can be recursively computed in accordance with Equation (7) below.
For the output layer, d1,L(p)=1. In this case, L is the number of layers. Similarly, djl(q) can be computed for an input xq.
The above-described training procedure of step 230 is repeated for all (xp, xq) pairs in the ranked list r.
Verifying Revised Parameter Weights—Step 240
The average change in weights in the neural network is calculated in accordance with Equation (8) below.
In Equation (8) above, N is the total number of instances in which the neural network makes an error in this iteration.
Weights
are updated in accordance with Equation (9) below.
Steps 230 and 240 are both repeated until there is no significant change in
as determined in step 240. That is, until Equation (10) below is satisfied.
In Equation (10), e is a small constant. Alternatively, the repeat-until loop can be repeated a predetermined number of times; that is, for a fixed number of iterations.
EXAMPLESExamples are now described of implementing the error measure given by Equation (1) and (2) using other techniques apart from neural networks (or without any reference to neural networks). Optimization techniques, such as genetic algorithms (GA) or simulated annealing (SA) can be used to minimize the error expressed in Equations (1) and (2). A specific case of using genetic algorithm for minimizing the error measure is described, as provided in Equations (1) and (2).
The utility function can be described as a known parametric form of second order as expressed in Equation (11) below.
In Equation (11) above, i and j denote the indices of the attributes of the items/bids. One can extend this second order form to higher order forms also. However, that does not affect the usage of GA and the error measure in this algorithm.
The GA-based technique operates as described below under points (i) to (vi).
-
- (ii) Define chromosomes that encode the parameters a's. For example, if there are only two attributes, then there exist parameters a0, a1, a2, a12 (in total 4 parameters). Let all the parameters be bounded in −1<a <+1 and each can be encoded in 8 bits such that a maximum precision that can be achieved is 1/256 in the representation of the value of each parameter. Thus, in this example, a chromosome length will be 4 byte (32 bits).
- (iii) Initialize a pool of chromosomes (a pool can consist of a large number of chromosomes depending on the constraints imposed the systems running time and resources). In the above example, one can start with 16 chromosomes in a pool. Initialize each chromosome randomly.
- (iv) Perform the crossover operation on the chromosomes.
- (v) Perform a mutation operation.
- (vi) Evaluate each chromosome to determine the error each chromosome is representing over all the bids/items. Perform selection, such as Roullette-wheel selection so that the chromosomes representing lower error measure (according to Equations (1) and (2)) have a higher probability of selection.
- (vii) Go to the step (ii) and repeat the procedure until the minimum error represented by the pool of chromosome does not decrease any further.
The above algorithm does not involve the use of neural networks. This algorithm can determine the utility function of complex nonlinear form using stochastic optimization. This algorithm can also be used for higher order complex form of utility functions. In a similar manner, simulated annealing algorithm can be used to determine an optimal utility function of known complex form.
Application of Described Techniques
Consider the following application. A buyer wants to buy a digital camera. If the buyer goes to an online department store, he finds there are more than 100 digital cameras. Each digital camera has more than 12 attributes. The aim of the buyer is not to evaluate each of them separately but to evaluate a few of them and on the basis of these rank all the remaining cameras.
The camera attributes considered are price, CCD resolution, memory card included or not, optical zoom, digital zoom, width, height, depth, and weight. Two possible utility functions that a user can have on these cameras are considered, based on the above attributes. These utility functions are:
f(camera)=0.4u(price)+0.3u(CCD-res)·U(mem-card-incl)·U(opt-zoom)·U(digital-zoom)+0.3u(width)·u(height)·u(depth)·u(weight) (ii) Sum of Products
f(camera)=u(price)·u(CCD-res)·u(mem-card-incl)·u(opt-zoom)·u(digital-zoom)·u(width)·u(height)·u(depth)·u(weight) (iii) Product
The effectiveness of the described techniques are tested for both of the above utility functions with random sampling and query based sampling algorithms. Different sizes of training sets, containing either a fully ranked set of cameras or the cameras are partially ranked, can be used. The performance measure that can be used to evaluate the effectiveness of the described techniques are the number of actual top k cameras predicted in the top k positions by the described techniques.
Computer Hardware and Software
The computer software involves a set of programmed logic instructions that are able to be interpreted by the computer system 300 for instructing the computer system 300 to perform predetermined functions specified by those instructions. The computer software can be an expression recorded in any language, code or notation, comprising a set of instructions intended to cause a compatible information processing system to perform particular functions, either directly or after conversion to another language, code or notation.
The computer software is programmed by a computer program comprising statements in an appropriate computer language. The computer program is processed using a compiler into computer software that has a binary format suitable for execution by the operating system. The computer software is programmed in a manner that involves various software components, or code means, that perform particular steps in the process of the described techniques.
The components of the computer system 300 include: a computer 320, input devices 310, 315 and video display 390. The computer 320 includes: processor 340, memory module 350, input/output (I/O) interfaces 360, 365, video interface 345, and storage device 355.
The processor 340 is a central processing unit (CPU) that executes the operating system and the computer software executing under the operating system. The memory module 1050 includes random access memory (RAM) and read-only memory (ROM), and is used under direction of the processor 1040.
The video interface 345 is connected to video display 390 and provides video signals for display on the video display 390. User input to operate the computer 320 is provided from input devices 310, 315 consisting of keyboard 310 and mouse 315. The storage device 355 can include a disk drive or any other suitable non-volatile storage medium.
Each of the components of the computer 320 is connected to a bus 330 that includes data, address, and control buses, to allow these components to communicate with each other via the bus 330.
The computer system 300 can be connected to one or more other similar computers via a input/output (I/O) interface 365 using a communication channel 385 to a network 380, represented as the Internet.
The computer software program may be provided as a computer program product, and recorded on a portable storage medium. In this case, the computer software program is accessed by the computer system 300 from the storage device 355. Alternatively, the computer software can be accessed directly from the network 380 by the computer 320. In either case, a user can interact with the computer system 300 using the keyboard 310 and mouse 315 to operate the programmed computer software executing on the computer 320.
The computer system 300 is described for illustrative purposes: other configurations or types of computer systems can be equally well used to implement the described techniques. The foregoing is only an example of a particular type of computer system suitable for implementing the described techniques.
Overview
Though techniques are described herein in the context of neural networks, other implementations are possible, such as those involving genetic algorithms, as noted above.
In step 440, a determination is made concerning whether the revised implied utility function is satisfactory. If the implied utility function is satisfactory, the implied utility function is accepted in step 450. Otherwise, steps 420 to 440 are repeated until the revised utility function is found to be satisfactory.
CONCLUSIONIn the case of a RFQ problem, a buyer receives a set of responses from his suppliers, the buyer can choose a subset of the received responses and rank these responses. The ranked subset can be presented to the above-described algorithm to train a neural network to represent this utility function. This utility function so computed by the described algorithm can then be used to order the entire set of responses. The buyer is likely to choose a winning responses from amongst the top ranking responses.
Various analogous contexts exist in which the described techniques can also be applied in a similar manner.
The described techniques for learning a utility function over a subset of ranked objects have the following advantages.
No restriction is imposed on interactions between the individual attribute utilities.
No prior knowledge of individual attribute utility functions is required. Neural networks can be used to learn user utility functions from the ranked list.
The objective function can be optimized with computing tools such as decision trees, support vector machines, evolutionary algorithms, Bayesian and belief networks, probabilistic reasoning.
The techniques described herein require only a set of ranked objects to generate a utility function.
Various alterations and modifications can be made to the techniques and arrangements described herein, as would be apparent to one skilled in the relevant art.
Claims
1. A method of providing goods or services, said method comprising:
- requesting a list of goods or services available from a supplier for purchase by a buyer through a computer system;
- associating attributes with said goods or services, wherein said attributes correspond to characteristics of said goods or services and are measured according to a quantifiable standard;
- providing a subset list of said goods or services as a training example to a multi-layered neural network;
- automatically generating a utility function comprising weighted components associated with said attributes based on training performed by the neural network;
- automatically ranking all said goods or services according to said utility function;
- presenting to said buyer a ranked list comprising a pair of the highest ranked goods or services as a representative sample from all said goods or services available from said supplier;
- evaluating said ranked list to determine whether said highest ranked goods or services match a buyer preference corresponding to said attributes of said goods or services;
- submitting the evaluation to said neural network;
- inputting said evaluation into said utility function;
- re-ranking all said goods or services according to said attributes and said evaluation;
- presenting to said buyer a re-ranked list comprising a pair of the highest re-ranked goods or services as a representative sample from all said goods or services available from said supplier;
- re-evaluating said re-ranked list to determine whether said highest re-ranked goods or services match said buyer preference corresponding to said attributes of said goods or services;
- iteratively repeating the submitting process through the re-evaluating process until said re-ranked list matches said ranked list; and
- presenting to said buyer a final ranked list of all said goods or services ranked according to buyer preferences corresponding to said attributes.
2. The method of claim 1, all the limitations of which are incorporated herein by reference, wherein said subset list of said goods or services is substantially less than a total number of said goods or services that are being ranked.
3. The method of claim 2, all the limitations of which are incorporated herein by reference, wherein said subset list of said goods or services is less than one percent of said total number of said goods or services that are being ranked.
4. The method of claim 1, all the limitations of which are incorporated herein by reference, further comprising:
- calculating an error measure associated with said utility function;
- revising utility function in response to said error measure; and
- performing the calculating of said error measure and the revising until said error measure results in a user-defined satisfactory difference between said re-ranked list and said ranked list.
5. The method of claim 1, all the limitations of which are incorporated herein by reference, wherein said quantifiable standard comprises any of a numeric measure and a true/false measure.
6. The method of claim 1, all the limitations of which are incorporated herein by reference, wherein said neural network is a feed-forward neural network.
7. The method of claim 1, all the limitations of which are incorporated herein by reference, wherein said subset list of said goods or services comprises a ranking of said goods or services.
8. The method of claim 7, all the limitations of which are incorporated herein by reference, wherein the ranking in said subset list is performed by said buyer.
9. The method of claim 1, all the limitations of which are incorporated herein by reference, further comprising performing the repeating process a fixed predetermined number of times.
10. The method of claim 1, all the limitations of which are incorporated herein by reference, wherein said utility function is mathematically denoted as ƒ(x),
- wherein
- f ( x ) = a 0 + ∑ i a i x i + ∑ i, j a ij x i x j,
- wherein a is a good or service; x is said pair of the highest ranked goods or services; and i and j are attributes of said goods or services.
11. A program storage device readable by computer, tangibly embodying a program of instructions executable by said computer to perform a method of providing goods or services, said method comprising:
- requesting a list of goods or services available from a supplier for purchase by a buyer through a computer system;
- associating attributes with said goods or services, wherein said attributes correspond to characteristics of said goods or services and are measured according to a quantifiable standard;
- providing a subset list of said goods or services as a training example to a multi-layered neural network;
- automatically generating a utility function comprising weighted components associated with said attributes based on training performed by the neural network;
- automatically ranking all said goods or services according to said utility function;
- presenting to said buyer a ranked list comprising a pair of the highest ranked goods or services as a representative sample from all said goods or services available from said supplier;
- evaluating said ranked list to determine whether said highest ranked goods or services match a buyer preference corresponding to said attributes of said goods or services;
- submitting the evaluation to said neural network;
- inputting said evaluation into said utility function;
- re-ranking all said goods or services according to said attributes and said evaluation;
- presenting to said buyer a re-ranked list comprising a pair of the highest re-ranked goods or services as a representative sample from all said goods or services available from said supplier;
- re-evaluating said re-ranked list to determine whether said highest re-ranked goods or services match said buyer preference corresponding to said attributes of said goods or services;
- iteratively repeating the submitting process through the re-evaluating process until said re-ranked list matches said ranked list; and
- presenting to said buyer a final ranked list of all said goods or services ranked according to buyer preferences corresponding to said attributes.
12. The program storage device of claim 11, all the limitations of which are incorporated herein by reference, wherein said subset list of said goods or services is substantially less than a total number of said goods or services that are being ranked.
13. The program storage device of claim 12, all the limitations of which are incorporated herein by reference, wherein said subset list of said goods or services is less than one percent of said total number of said goods or services that are being ranked.
14. The program storage device of claim 11, all the limitations of which are incorporated herein by reference, further comprising:
- calculating an error measure associated with said utility function;
- revising utility function in response to said error measure; and
- performing the calculating of said error measure and the revising until said error measure results in a user-defined satisfactory difference between said re-ranked list and said ranked list.
15. The program storage device of claim 11, all the limitations of which are incorporated herein by reference, wherein said quantifiable standard comprises any of a numeric measure and a true/false measure.
16. The program storage device of claim 11, all the limitations of which are incorporated herein by reference, wherein said neural network is a feed-forward neural network.
17. The program storage device of claim 11, all the limitations of which are incorporated herein by reference, wherein said subset list of said goods or services comprises a ranking of said goods or services.
18. The program storage device of claim 17, all the limitations of which are incorporated herein by reference, wherein the ranking in said subset list is performed by said buyer.
19. The program storage device of claim 11, all the limitations of which are incorporated herein by reference, further comprising performing the repeating process a fixed predetermined number of times.
20. The program storage device of claim 11, all the limitations of which are incorporated herein by reference, wherein said utility function is mathematically denoted as ƒ(x), wherein f ( x ) = a 0 + ∑ i a i x i + ∑ i, j a ij x i x j, wherein a is a good or service; x is said pair of the highest ranked goods or services; and i and j are attributes of said goods or services.
Type: Application
Filed: Dec 5, 2007
Publication Date: May 1, 2008
Applicant: International Business Machines Corporation (Armonk, NY)
Inventors: Jayanta Basak (New Delhi), Manish Gupta (New Delhi)
Application Number: 11/950,490
International Classification: G06N 3/08 (20060101); G06F 15/18 (20060101);