Abstract: A neural network device includes a crossbar grid including first metal lines running in a first direction and second metal lines running transversely to the first metal lines and being electrically isolated from the first metal lines. An array of cross-over elements is included. Each cross-over element is connected between a first metal line and a second metal line. The cross-over elements each include a floating gate transistor device having a floating node. The floating node is configured to store a programmable weight value.

Abstract: An arithmetic apparatus used for a neural network includes a plurality of digital-time conversion circuits connected in series and a time-digital conversion circuit connected to a last digital-time conversion circuit in the series. Each of the digital-time conversion circuits is configured to delay a first input time signal by a variable amount, delay a second input time signal by a fixed amount, and output the delayed first and second input time signals respectively as either first and second output time signals or second and first output time signals, in accordance with the input data. The time-digital conversion circuit is configured to generate a digital output signal by comparing first and second output time signals from the last digital-time conversion circuit.

Abstract: Operations of computing devices are managed using one or more deep neural networks (DNNs), which may receive, as DNN inputs, data from sensors, instructions executed by processors, and/or outputs of other DNNs. One or more DNNs, which may be generative, can be applied to the DNN inputs to generate DNN outputs based on relationships between DNN inputs. The DNNs may include DNN parameters learned using one or more computing workloads. The DNN outputs may be, for example, control signals for managing operations of computing devices, predictions for use in generating control signals, warnings indicating an acceptable state is predicted, and/or inputs to one or more neural networks. The signals enhance performance, efficiency, and/or security of one or more of the computing devices. DNNs can be dynamically trained to personalize operations by updating DNN weights or other parameters.

Abstract: A neuromorphic memory circuit including a memory cell with a programmable resistive memory element. A postsynaptic capacitor builds up a leaky integrate and fire (LIF) charge. An axon LIF pulse generator activates a LIF discharge path from the postsynaptic capacitor through the resistive memory element when the axon LIF pulse generator generates axon LIF pulses. A postsynaptic comparator compares the capacitor voltage to a threshold voltage and generates postsynaptic output pulses when the capacitor voltage passes the threshold voltage. The postsynaptic output pulses include a postsynaptic firing characteristic dependent on a frequency of the axon LIF pulses. A refractory circuit prevents the postsynaptic comparator from generating additional postsynaptic output pulses until a refractory time passes since a preceding postsynaptic output pulse. A training circuit adjusts the postsynaptic firing characteristic to match a target firing characteristic.

Abstract: A method is implemented by a network device to classify encrypted data traffic. The method identifies characteristics of the encrypted data traffic that have been modeled where network anomalies have been injected into the encrypted data traffic to provide additional traffic characteristics that enable categorization. The method receives the encrypted data traffic, applies an encrypted traffic categorization model to the received encrypted data traffic to determine a first categorization identification, injects an anomaly into the encrypted data traffic where the first categorization identification is not within a precision threshold, applies the encrypted traffic categorization model to monitored encrypted traffic after injection of the anomaly to determine a second categorization identification, and applies the second categorization identification where the second categorization identification is within the precision threshold.

Abstract: Petroleum Analytics Learning Machine (or PALM) system is a machine learning based, “brutally empirical” analysis system for use in all upstream and midstream oil and gas operations. PALM system optimizes exploration, production and gathering from at least one well of oil and natural gas fields to maximize production while minimizing costs. Normalized data are processed to determine clusters of correlation in multi-dimensional space to identify a machine learned ranking of importance weights for each attribute. Predictive and prescriptive optimization on the normalized data is performed utilizing unique combinations of machine learning and statistical algorithm ensembles. The unstructured textual data are classified to correlate with optimal production to capture the dynamics of at least one or more wells of oil and natural gas fields and to provide categorization results from labeled data sets to identify patterns.

Abstract: Boltzmann machines are trained using an objective function that is evaluated by sampling quantum states that approximate a Gibbs state. Classical processing is used to produce the objective function, and the approximate Gibbs state is based on weights and biases that are refined using the sample results. In some examples, amplitude estimation is used. A combined classical/quantum computer produces suitable weights and biases for classification of shapes and other applications.

Abstract: Social media networks have become a primary source for news and opinions on topics ranging from sports to politics. Sentiment analysis is typically constrained to two classes—positive and negative. A computing system is herein described for building a multi-sentiment multi-label model for electronic data that uses emojis as class labels. The electronic messages are classified into six sentiment classes. The computing system collects and creates a large corpus of clean and processed training data with emoji-based sentiment classes using little-to-no manual intervention. A threshold-based formulation is used to assign one or two class labels (multi-label) to an electronic message. The multi-sentiment multi-label model produces a desirable cross validation accuracy.

Abstract: There are provided system and method of classifying defects in a semiconductor specimen. The method comprises: upon obtaining by a computer a Deep Neural Network (DNN) trained to provide classification-related attributes enabling minimal defect classification error, processing a fabrication process (FP) sample using the obtained trained DNN; and, resulting from the processing, obtaining by the computer classification-related attributes characterizing the at least one defect to be classified, thereby enabling automated classification, in accordance with the obtained classification-related attributes, of the at least one defect presented in the FP image.

Abstract: An apparatus for deciding a simplification policy for a neural network is provided. The deciding apparatus has a plurality of artificial neurons, a receiving circuit, a memory, and a simplifying module. The plurality of artificial neurons are configured to form an original neural network. The receiving circuit receives a set of sample for training the original neural network. The memory is used for recording a plurality of learnable parameters for the original neural network. After the original neural network has been trained with the set of sample, the simplifying module abandons a part of neuron connections in the original neural network based on the learnable parameters recorded by the memory. The simplifying module accordingly decides the structure of a simplified neural network and a plurality of learnable parameters for the simplified neural network.

Abstract: A method and apparatus may include receiving a signal from a motor. The signal is received while the motor is operating. The method also includes performing a pre-processing of the signal. The method also includes inputting the signal to a 1D convolutional neural network. The method also includes detecting a fault of the motor based on the output of the neural network.

Abstract: A neuromorphic memory circuit including a programmable resistive memory element, an axon LIF pulse generator to generate an axon LIF pulse, a back propagation pulse generator to generate a back propagation pulse, a postsynaptic capacitor configured to build up a forward propagation LIF charge over time, and a presynaptic capacitor configured to build up a back propagation LIF charge over time. A first transistor activates a first discharge path from the postsynaptic capacitor through the programmable resistive memory element when the axon LIF pulse generator generates the axon LIF pulse. A second transistor activates a second discharge path from the presynaptic capacitor through the programmable resistive memory element when the back propagation pulse generator generates the back propagation pulse.

Abstract: An expert recommendation method, system, and non-transitory computer readable medium, include a topic extraction circuit configured to extract a topic of a user input message in real-time, an expert recommending circuit configured to recommend a list including a plurality of experts based on the extracted topic, and an expert ranking circuit configured to order the experts on the list of experts based on an expert rank factor.

Abstract: A method and system are provided to calculate a future behavioral data and identify a relative causal impact of external factors affecting the data. Behavioral data and data for one or more external factors are harvested for a first time period. New behavioral data is harvested for a second time period. New data for the second time period is harvested. Based on a second training algorithm, a forecast time series value of a future behavioral data for a third time period that is after the second time period is calculated. A relative causal impact between each external factor and the predicted time series value of the behavioral data, for the third time period, is identified.

Abstract: A mechanism is provided in a computing device configured with instructions executing on a processor of the computing device to implement a question answering system, for answer scoring based on a combined informativity and specificity score. The question answering system, executing on the processor of the computing device and configured with a question answering machine learning model, generates a set of candidate answers for a user-generated input question. For each given candidate answer in the set of candidate answers, an informativity and specificity scorer of the question answering system determines a specificity value of each term in the given candidate answer based on a position of the term in a taxonomy data structure and determining a specificity score of the given candidate answer based on the specificity value of the terms in the given candidate answer.

Abstract: A first and second blending profile may be created for a set of question answering pipelines. A set of test answer data may be generated for a first answering pipeline. The test answer data may be generated based on a set of test question and using an answer key associated with the test questions. Based on the test answer data, a first blending profile can be created for the first answering pipeline. Using the set of test questions and a second answer key, another set of test answer data may be generated. This set may be generated for the second answering pipeline. Using this second answering pipeline test answer data, a second blending profile can be generated for the second answering pipeline. Each blending profile may have metadata about a confidence of each pipeline.

Abstract: An answer to a question may selected from answers from a set of answering pipelines. Question answer data can be generated for a question, using a first answering pipeline. Another set of question answer data can be generated for the second question, using the second answering pipeline. The question answer data can include answers and confidence values for each answer. Using a weighting formula and a blending profile for the first answering pipeline, a vote weight can be determined for an answer with the highest confidence value. The same weighting formula and a second blending profile may be used to determine a vote weight for another answer with the highest confidence value. An answer to the question may be selected from the answers, based on the overall highest vote weight.

Abstract: In one example of the disclosure, a subject message for a display caused by a subject software application is obtained. A prediction model is utilized to identify the subject message as a first type message or a second type message. The model is a model determined based upon a set of target words determined by imposition of a set of rules upon a set of user facing messages extracted from a set of software applications, wherein each of the extracted messages was classified post-extraction as a first type message or a second type message. A communication identifying the subject message as the first type message or the second type message is provided.

Abstract: A method may include selecting a particular entity from a knowledge graph when a level of connectivity between entities in the knowledge graph that are neighbors to the particular entity is above a certain level and determining whether the particular entity is in a character string.

Abstract: Predictive regression models are widely used in different domains such as life sciences, healthcare, pharma etc. and variable selection, is employed as one of the key steps. Variable selection can be performed using random or exhaustive search techniques. Unlike a random approach, the exhaustive search approach, evaluates each possible combination and consequently, is a computationally hard problem, thus limiting its applications. The embodiments of the present disclosure perform i) parallelization and optimization of critical time consuming steps of the technique, Variable Selection and Modeling based on the Prediction (VSMP) ii) its applications for the generation of the best possible predictive models using input dataset (e.g., Blood Brain Barrier Permeation data) and iii) business impact of predictive models that are requires the selection of larger number of variables.

Abstract: There is disclosed a computer implemented method of generating a training object for training a machine learning algorithm (MLA). The method comprises: acquiring a digital training document to be used in the training; transmitting the digital training document to a plurality of assessors, transmitting further including indicating a range of possible labels for the assessors to assess from, the range of possible labels including at least a first possible label and a second possible label; obtaining from each of the plurality of assessors a selected label to form a pool of selected labels; generating a consensus label distribution based on the pool of selected labels, the consensus label distribution representing a range of perceived labels for the digital training document and an associated probability score for each of the perceived labels; and training the machine learning algorithm using the digital training document and the consensus label distribution.

Abstract: A method for detecting a bottleneck in a system includes receiving a graph, wherein a node represents a software module and an edge represents a communication channel between software modules, monitoring selected resources for each software module in comparison to available resources, monitoring a ratio of a bandwidth consumed on a communication channel versus available bandwidth, traversing the graph for identifying a source software module whose produced amount of output is below the amount of output needed by the software module that is in idle mode, and analyzing a resource consumption of the identified source software module to identify a lacking amount of resource for the identified source software module. A computer system and computer program product corresponding to the above method are also disclosed herein.

Abstract: Systems and methods are provided for performing multi-objective optimizations with a relatively large number of objectives to which optimization is to be performed. The objectives of the optimization problem may be partitioned to two or more subsets (e.g., overlapping or non-overlapping subsets) of objectives, and partial optimization(s) may be performed using a subset or combination of subsets of the objectives. One or more of the partial optimizations may use one or more pareto-optimized chromosomes from a prior partial optimization. A final full optimization may be performed according to all of the objectives of the optimization problem and may use one or more chromosomes of any preceding partial optimization as a starting point for finding a final solution to the optimization problem. Any variety of processes may be employed to mitigate archive explosion that may be associated with relatively large objective sets.

Abstract: Embodiments of the present invention provide a system for a managing entity to automatically provide alerts based on tail event analysis. The system may receive input data in real time from vendor data feeds, social media data feeds, and a tail event ledger. The system may then automatically populate surveys, transmit the surveys to responders, and receive survey results from the responders. The survey results may be transmitted to specialists that return a predicted tail event outcome. This predicted tail event outcome is then automatically transmitted to partners, or decision makers, that provide action steps for responding to the predicted tail event outcome. The system may then continuously monitor the input data, identify an indicator of an occurrence of the tail event, and then automatically transmit the action steps to appropriate parties.

Abstract: According to one embodiment, a method for assessing whether a first site possesses a selected characteristic, the method comprising training, using a machine-learning process, a classifier to determine, based on web site data corresponding to one or more known web sites, whether the first web site possesses the selected characteristic, wherein the one or more known web pages comprise web pages known to possess the selected characteristic and web pages known not to possess the selected characteristic.

Abstract: Embodiments associate a relevant semantic data type (e.g., date) with incoming raw data (e.g., a column of digits) which lacks metadata. Assignment of semantic data type is inferred from a plurality of features. A first step determines a first feature comprising success rate in converting the raw data into various semantic data types. Then, alignment between observed/reference distributions of other features (e.g., data first digit, data length) is determined per-semantic data type. Total scores for each semantic data type are calculated from the combined features, and used as a basis for ranking the semantic data types. The total scores may reflect a weighting of the various features. In a second step, top-ranked semantic data types may be further differentiated from one another by applying additional features. User feedback regarding accuracy of semantic data type assignment, may be incorporated into training data used to modify the feature reference distributions.

Abstract: A method to predict whether a particular fungus will be observed by a scout checking on a particular crop at a particular location on a particular date. The method using a set of previous scouting reports and weather information for the particular location, including weather information for a period of time before the particular date.

Abstract: A method for estimating a number of occupants in a region comprises receiving a time series of sensor values detected over a period of time by a motion sensor sensing motion in the region. A spread parameter indicative of the spread of the sensor values is determined. The number of occupants in the region is estimated based on the spread parameter.

Abstract: For a plurality of sensors, a particular sensor is indicated as a target sensor and the other sensors as input sensors. A regression model is trained using historical data from the plurality of related sensors. The trained regression model is applied to the target sensor to generate a predicted target sensor value. A difference between an actual target sensor value and the predicted target sensor value is calculated. A probability of difference for the calculated difference between the actual target sensor value and the predicted target sensor value is compared against a threshold value.

Abstract: The disclosure relates to technology for determining a root cause of anomalous behaviors in networks. First indicators (KQIs) are categorized into first groups (states) and second indicators (KPIs) are categorized into second groups. A conditional probability is estimated by calculating a probability that the second indicators will result in degradation of the first indicators based on historical data using association rule learning. The second indicators having the conditional probability associated with degradation of the first indicators are mapped to a corresponding one of the first groups in a probabilistic network structure based on a detected degradation of the first indicators in the historical data.

Abstract: A provenance method, system, and non-transitory computer readable medium for a plurality of eidetic systems having logs, include a log-segment clustering circuit configured to crawl the logs of each of the eidetic systems to cluster segments across the logs of temporally correlated events into clustered segments, a probabilistic interleaving circuit configured to analyze the correlated segments to interleave an order of processes in the logs and assign a probability to the order of the processes occurring, and a probabilistic linearization circuit configured to create a probability tree which includes a total probability that a process in the clustered segments causes a next process in the clustered segments until an end of the temporal event of the clustered segments for each of the interleaved order of processes interleaved by the probabilistic interleaving circuit.

Abstract: System and method for analyzing driver behavior based on telematics data are disclosed. In an example, a probability of a user driving a vehicle is computed and a risk score is generated to develop at least one driver profile based on the probability. Further, routes taken by said user driving said vehicle are clustered to generate enhanced driver profile and using the clustered output to develop dynamic intelligent contexts for each said route and adding contextual intelligence messages to customize said risk score. Furthermore, the routes taken by the said user in real time are predicted. In addition, a missing route is identified through imputation of missed routes to compute annualized mileage, and a missing distance is imputed in an analysis of at least one trip of the driver in the vehicle. Also, independent trips are stitched based on at least one recommendation from an analytics engine.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for optimizing content presentation.

Abstract: In one embodiment, a method includes sending, through a communications network, several volumes of notifications corresponding to a first notification type to multiple users and several volumes of notifications corresponding to a second notification type to multiple users. The method further determines visitation impacts of the volumes of notifications of the first and second notification types and trains a machine-learning model based on the visitation impacts. The machine-learning model generates an assessment of a likelihood of interaction by a recipient user with each of the notifications.

Abstract: A computer-implemented method includes a processor extracting data and metadata from a process model, where the process is comprised of activities and the metadata is associated with each activity. The processor generates at least one user story for at least one activity, where the at least one user story includes an estimate attribute reflecting a predicted timeframe for completion of at least a portion of the at least one activity. The processor updates the model to reflect the at least one user story and displays the updated model as a project plan in a project management interface on a computing resource. The processor assigns a resource to the at least one user story and dynamically updates the estimate attribute of the at least one user story to reflect a new predicted timeframe, calculates impacts to the process and displays the impacts and the new predicted timeframe in the project management interface.

Abstract: Techniques are described for adaptive and augmented decision making by an artificial intelligence (AI) engine, such as an engine that employs machine learning techniques. A decision-making process may be executed to make a decision regarding operations of the organization, and the AI engine may be employed to analyze the various aspects of a decision and determine a risk level associated with the decision. The risk level may be a combination of the probability of a negative outcome and a magnitude of loss that may occur due to a negative outcome. The automated process may also determine a confidence level that indicates a degree of confidence in the determined risk level. Risk and confidence may be independent values. Implementations may enable risk mitigation by providing a risk estimate with higher confidence than traditional methods.

Abstract: On the basis of a difference between first distribution regarding target task learning data as a plurality of learning data which belongs to a first category of a first task as a target task and second distribution regarding a plurality of learning data which belongs to the first category of source task learning data as learning data which belongs to the first category of a second task different from the first task, a transformation parameter for transforming the source task learning data is adjusted, the source task learning data is transformed based on the adjusted transformation parameter, and a classifier regarding the first task is generated based on the transformed source task learning data and the target task learning data.

Abstract: Systems, technologies and techniques for generating prospective legal strategies are disclosed. The system and technologies employ data mining, natural language processing and machine learning approaches to generate prospective legal strategies. The system and technologies analyze given case facts (i.e., background facts, event type such as accident, injury, malpractice, discrimination) and provide a rich set of insights that assist in formulating effective legal arguments and strategies.

Abstract: A network of multitier user's devices (mobile units) that performs household chores and functions as a mobile general purpose autonomous intelligent machines, having some primary functions and a plurality of additional secondary functions. The primary functions include one or more of vacuum cleaning, lawn mowing, typical drone functionalities, the additional autonomous functions include wardrobe management, premises safety, sentry job, rendering personalized music, self-recharging from typical electric outlet, self-learning to predict the users' behaviors and planning and scheduling the functionalities (decision making) accordingly, providing AI (Artificial Intelligence) assistance and support, Day-In-Life recording and support and so forth.

Abstract: A learning agent is disclosed that receives data in sequence from one or more sequential data sources; generates a model modelling sequences of data and actions; and selects an action maximizing the expected future value of a reward function, wherein the reward function depends at least partly on at least one of: a measure of the change in complexity of the model, or a measure of the complexity of the change in the model. The measure of the change in complexity of the model may be based on, for example, the change in description length of the first part of a two-part code describing one or more sequences of received data and actions, the change in description length of a statistical distribution modelling, the description length of the change in the first part of the two-part code, or the description length of the change in the statistical distribution modelling.

Abstract: A method, computer program product, and computer system for identifying, by a computing device, a plurality of content from at least one source. A first portion of the plurality of content may be categorized in a first feed category based on a first probabilistic model. A second portion of the plurality of content may be categorized in a second feed category based on the first probabilistic model. User feedback may be received to change the categorization of a first content of the first portion of the plurality of content in the first feed category to the second feed category. A second probabilistic model may be generated based upon, at least in part, the user feedback. The categorization of a second content of the first portion of the plurality of content in the first feed category may be reorganized to the second feed category based upon, at least in part, the second probabilistic model.

Abstract: Methods and systems of using machine learning to create a trusted model that improves the operation of a computer system controller are provided herein. In some embodiments, a machine learning method includes training a model using input data, extracting the model, and determining whether the model satisfies the trust-related constraints. If the model does not satisfy the trust-related constraint, modifying at least one of: the model using one or more model repair algorithms, the input data using one or more data repair algorithms, or a reward function of the model using one or more reward repair algorithms, and re-training the model using at least one of the modified model, the modified input data, or the modified reward function. If the model satisfies the trust-related constraints, providing the model as a trusted model that enables a computer system controller to perform system actions within predetermined guarantees.

Abstract: Systems and methods for automated evaluation system routing are described herein. The system can include a memory, which can include a model database and a correlation database. The system can include a first user device and a second user device. The system can include at least one server. The at least one server can: receive a response communication from the user device; generate an initial evaluation value according to an AI model; determine a correlation between the initial evaluation value and evaluation range data; accept the initial evaluation value when the correlation exceeds a threshold value; and route the response communication to the second user device for generation of an elevated evaluation value when the correlation does not exceed the threshold value.

Abstract: Systems and methods for determining video infrastructure delivery problems using machine learning are presented. In an aspect, a system includes a reception component configured to receive information regarding videos streamed by the system to devices, wherein the information identifies video playback events at the devices and rebuffer events respectively associated with the video playback events. The system further includes a quality component configured to identify features related to quality of the playback events at the devices based on the information, and an analysis component configured to determine probabilities of occurrence of the re-buffer events based on different combinations of the features, and determine weighted values for each of the features that reflect their contribution to the probabilities of occurrence of the re-buffer events based on the different combinations of the features.

Abstract: The subject disclosure is directed towards a real-time or near real-time sentiment monitoring service. A set of rules such as keywords and data sources to crawl is provided to the monitoring service, which crawls the sources to obtain sentiment-related data for an entity, such as a corporation or product. Content items may be selected from the crawled data, and/or the data may be analyzed to provide results. The results may be displayed, such as on a content page, to quickly view the public's sentiment regarding the entity. The rules may be dynamically modified by a user or collaborating users to tune monitoring of the entity as desired, e.g., to obtain more relevant results.

Abstract: A method for determining a reservation booking of a chair location positioned in an outdoor environment, the method implemented by a computer processor operating on a set of instructions stored in memory to: receive a reservation request from a user over a communication network, the reservation request including a unique identifier for an establishment providing the outdoor environment and a reservation date/time; send a map of the outdoor environment over the communication network including a plurality of chair locations, each chair of the plurality of chairs having a respective indicator indicating whether said each chair is available or booked for the reservation time, the map including exterior features of the outdoor environment positioned in relation to the plurality of chair locations; receive a selection over the communication network of a desired chair location displayed on the map, the desired chair selected from the plurality of chair locations, and send the reservation booking over a communication

Abstract: The present disclosure is directed to systems and methods for managing reservations. A computer-implemented method for use with a system for managing reservations includes: transmitting, via a vendor transmission apparatus of a vendor device, availability data for a vendor to a display apparatus of a user device, the display apparatus configured to display information on the user device; receiving, via a user transmission apparatus of the user device, reservation request data for the vendor; determining, via a reservation data server coupled to the vendor device and the user device, if there is an available seat at the vendor having the vendor device; charging, via the reservation data server, the user device a variable fee amount in order to reserve the available seat, the variable fee amount being able to be modified by the vendor device; and delivering, via the vendor transmission apparatus, confirmation to the display apparatus of the user device that the available seat has been reserved.

Abstract: A computer implemented method for predicting dynamic pricing for a product or service is disclosed. The method comprises: receiving, at a server, transaction data for a plurality transactions relating to the product or service, the transaction data comprising transaction information indicating attributes of the transaction and booking information indicating attributes of the product or service; storing the transaction data on the server as a data set; calculating, in a derived variable calculation module of the server, derived variables for the data set from the transaction information and/or the booking information; generating, in a model generation module of the server, a model for the price of the product or service as a function of the derived variables; and predicting, in a price prediction module of the server, a price for the product or service using the model.

Abstract: A service voucher processing method, device, and system, and a storage medium are disclosed. The method includes: obtaining information about a candidate service, and loading a virtual identifier of the candidate service in a graphical interface; initiating, if a target service reserved by the reservation instruction is detected, a reservation request; obtaining a service voucher corresponding to the target service based on permission corresponding to the target service, and activating the service voucher when an instruction for using the target service is detected; performing authentication on the permission, a verification request carrying permission information, and a serial communication method being used for the verification request; and providing the permission when the authentication passes, determining that the target service is used, and cancelling the permission.

Abstract: The present invention discloses an information processing device, etc., for constructing as appropriate prediction model even when there exist a plurality of customers having different demand tendencies. An information processing device pertaining to the present invention includes: a means for dividing a plurality of contract units into a discretional number of groups on the basis of a feature corresponding to each of elements that could affect demand in the contract units, the feature varying with the time series of the demand; and a means for constructing, for each of the groups, a prediction model that represents the demand, and outputting the constructed prediction model.