Microsoft Technology Licensing, LLC

Other

David Stern

Ralf Herbrich

Jurgen Anne Francois Marie Van Gael

Thore Graepel

Tauhid Rashed Zaman

Gilad Lotan

Thore K H Graepel

Ulrich Paquet

Gjergji Kasneci

Thore Kraepel

Philipp Hennig

Horst Cornelius Samulowitz

One or more techniques and/or systems are disclosed for predicting propagation of a message on a social network. A predictive model is trained to determine a probability of propagation of information on the social network using both positive and negative information propagation feedback, which may be collected while monitoring the social network over a desired period of time for information propagation. A particular message can be input to the predictive model, which can determine a probability of propagation of the message on the social network, such as how many connections may receive at least a portion of the message and/or a likelihood of at least a portion of the message reaching respective connections in the social network.

Information propagation probability for a social network

A recommender system may be used to predict a user behavior that a user will give in relation to an item. In an embodiment such predictions are used to enable items to be recommended to users. For example, products may be recommended to customers, potential friends may be recommended to users of a social networking tool, organizations may be recommended to automated users or other items may be recommended to users. In an embodiment a memory stores a data structure specifying a bi-linear collaborative filtering model of user behaviors. In the embodiment an automated inference process may be applied to the data structure in order to predict a user behavior given information about a user and information about an item. For example, the user information comprises user features as well as a unique user identifier.

Recommending items to users utilizing a bi-linear collaborative filtering model

Knowledge corroboration is described. In an embodiment many judges provide answers to many questions so that at least one answer is provided to each question and at least some of the questions have answers from more than one judge. In an example a probabilistic learning system takes features describing the judges or the questions or both and uses those features to learn an expertise of each judge. For example, the probabilistic learning system has a graphical assessment component which aggregates the answers in a manner which takes into account the learnt expertise in order to determine enhanced answers. In an example the enhanced answers are used for knowledge base clean-up or web-page classification and the learnt expertise is used to select judges for future questions. In an example the probabilistic learning system has a logical component that propagates answers according to logical relations between the questions.

Knowledge corroboration

Machine learning techniques may be used to train computing devices to understand a variety of documents (e.g., text files, web pages, articles, spreadsheets, etc.). Machine learning techniques may be used to address the issue that computing devices may lack the human intellect used to understand such documents, such as their semantic meaning. Accordingly, a topic model may be trained by sequentially processing documents and/or their features (e.g., document author, geographical location of author, creation date, social network information of author, and/or document metadata). Additionally, as provided herein, the topic model may be used to predict probabilities that words, features, documents, and/or document corpora, for example, are indicative of particular topics.

Topic models

Many computing scenarios involve the classification of content items within one or more categories. The content item set may be too large for humans to classify, but an automated classifier (e.g., an artificial neural network) may not be able to classify all content items with acceptable accuracy. Instead, the automated classifier may calculate a classification confidence while classifying respective content items. Content items having a low classification confidence may be sent to a human classifier, and may be added, along with the categories identified by the human classifier, to a training set. The automated classifier may then be retrained using the training set, thereby incrementally improving the classification confidence of the automated classifier while conserving the involvement of human classifiers. Additionally, human classifiers may be rewarded for classifying the content items, and the costs of such rewards may be considered while selecting content items for the training set.

Human-assisted training of automated classifiers

Managing a portfolio of experts is described where the experts may be for example, automated experts or human experts. In an embodiment a selection engine selects an expert from a portfolio of experts and assigns the expert to a specified task. For example, the selection engine has a Bayesian machine learning system which is iteratively updated each time an experts performance on a task is observed. For example, sparsely active binary task and expert feature vectors are input to the selection engine which maps those feature vectors to a multi-dimensional trait space using a mapping learnt by the machine learning system. In examples, an inner product of the mapped vectors gives an estimate of a probability distribution over expert performance. In an embodiment the experts are automated problem solvers and the task is a hard combinatorial problem such as a constraint satisfaction problem or combinatorial auction.

Managing a portfolio of experts

Scoring a board configuration for a territory board game is often not straightforward and yet there is a desire to determine such scores quickly and accurately. For example, in the game of GO, determining the score at the end of the game involves assessing whether stones on the board are alive or dead which is a difficult judgment. Given a board configuration, the game is played by a scoring system to obtain a terminal board configuration. This is repeated to obtain a plurality of terminal board configurations from which an assessment can be made as to how likely each board position is to be won by a particular player at the end of the game. The scoring system obtains the terminal board configurations by playing random moves or by making a biased sampling of moves. The biased sampling is made using an evaluation function or in any suitable way. In the game of GO, seki positions are quickly and easily identified. An automated game playing system uses the output of the scoring system to assess when to offer to end a game. The output of the scoring system can also be used to provide hints to players during a game.

Scoring system for games

We describe an apparatus for learning to predict moves in games such as chess, Go and the like, from historical game records. We obtain a probability distribution over legal moves in a given board configuration. This enables us to provide an automated game playing system, a training tool for players and a move selector/sorter for input to a game tree search system. We use a pattern extraction system to select patterns from historical game records. Our learning algorithm learns a distribution over the values of a move given a board position based on local pattern context. In another embodiment we use an Independent Bernoulli model whereby we assume each moved is played independently of other available moves.

Growing community of inventors

Cambridge, United Kingdom

David Stern