Microsoft Technology Licensing, LLC

Microsoft Corporation

Tie-Yan Liu

Wei-Ying Ma

Hang Li

Bin Gao

Tao Qin

Taifeng Wang

Lei Yang

Lei Qi

Hong-Jiang Zhang

Jun Xu

Yuting Liu

Yu-Ting Liu

Hua-Jun Zeng

Bo Feng

Zhi Chen

Yunbo Cao

Minghao Liu

Xiaodong Fan

Wei Ting Chen

Zhike Kong

Jinhui Yuan

Ming-Feng Tsai

Zhe Cao

Qiwei Ye

Congkai Sun

Qin Tao

Geng Xiubo

Qiankun Zhao

Huai-Yuan Yang

Qi Meng

In implementations of the present disclosure, a solution for optimization of a learning network in an equivalent class space is provided. In this solution, base paths running through layers of a learning network are determined. Each node utilizes an activation function with a scaling invariant property to process an input from a node of a previous layer, each base path comprises a single node in each layer, and processing in the base paths is linearly independent from each other. A combined value of parameters associated with nodes in each base path is updated. A parameter associated with a node is used to adjust an input obtained from a node of a previous layer. Values of parameters associated with nodes in the base paths are updated based on updated combined values of parameters. Through this solution, optimization efficiency can be improved and more accurate optimized values of parameters are achieved.

Optimization of neural network in equivalent class space

In some examples, a machine learning system may use morphological knowledge to enhance a deep learning framework for learning word embedding. The system may consider, among other things, morphological similarities between and among words in a learning process so as to handle new or rare words, edit distances, longest common substring similarities, morpheme similarities, and syllable similarities as morphological knowledge to build a relation matrix between or among words. The system may apply the deep learning framework to query classification, web search, text mining, information retrieval, and natural language processing tasks, for example. The system may accomplish such tasks with relatively high efficiency and speed, while utilizing less computing resources as compared to other systems.

Learning word embedding using morphological knowledge

In implementations of the subject matter described herein, each token for containing an element in the training data is sampled according to a factorization strategy in training. Instead of using a single proposal, the property value of the target element located at the token being scanned is iteratively updated one or more times based on a combination of an element proposal and a context proposal. The element proposal tends to accept a value that is popular for the target element independently of the current piece of data, while the context proposal tends to accept whenever the property value that is popular in the context of the target data or popular for the element itself. The proposed modeling training approach can converge in a quite efficient way.

Mixed proposal based model training system

Many search engines attempt to understand and predict a user's search intent after the submission of search queries. Predicting search intent allows search engines to tailor search results to particular information needs of the user. Unfortunately, current techniques passively predict search intent after a query is submitted. Accordingly, one or more systems and/or techniques for actively predicting search intent from user browsing behavior data are disclosed herein. For example, search patterns of a user browsing a web page and shortly thereafter performing a query may be extracted from user browsing behavior. Queries within the search patterns may be ranked based upon a search trigger likelihood that content of the web page motivated the user to perform the query. In this way, query suggestions having a high search trigger likelihood and a diverse range of topics may be generated and/or presented to users of the web page.

Active prediction of diverse search intent based upon user browsing behavior

Techniques for determining user information needs and selecting data based on user information needs are described herein. The present disclosure describes extracting topics of interests to users from multiple sources including search log data and social network website, and assigns a budget to each topic to stipulate the quota of data to be selected for each topic. The present disclosure also describes calculating similarities between gathered data and the topics, and selecting top related data with each topic subject to limit of the budget. A search engine may use the techniques described here to select data for its index.

User information needs based data selection

A method and system for augmenting a training set used to train a classifier of documents is provided. The augmentation system augments a training set with training data derived from features of documents based on a document hierarchy. The training data of the initial training set may be derived from the root documents of the hierarchies of documents. The augmentation system generates additional training data that includes an aggregate feature that represents the overall characteristics of a hierarchy of documents, rather than just the root document. After the training data is generated, the augmentation system augments the initial training set with the newly generated training data.

Augmenting a training set for document categorization

Some implementations provide techniques for selecting web pages for inclusion in an index. For example, some implementations apply regularization to select a subset of the crawled web pages for indexing based on link relationships between the crawled web pages, features extracted from the crawled web pages, and user behavior information determined for at least some of the crawled web pages. Further, in some implementations, the user behavior information may be used to sort a training set of crawled web pages into a plurality of labeled groups. The labeled groups may be represented in a directed graph that indicates relative priorities for being selected for indexing.

Page selection for indexing

A method and system for generating a search result for a query of hierarchically organized documents based on retrieval of subtrees that are key resources for topic distillation is provided. The retrieval system may identify documents relevant to a query using conventional searching techniques. The retrieval system then calculates a subtree feature for subtrees that have an identified document as their root. After the retrieval system calculates the subtree feature for the subtrees, the retrieval system may generate a subtree relevance score for each subtree based on its subtree feature. The retrieval system may then order the identified documents based on their corresponding subtree relevances.

Topic distillation via subsite retrieval

A method and system is provided for calculating importance of documents based on transition probabilities from a source document to a target document based on looking ahead to information content of target documents of the source document. A look-ahead importance system generates transition probabilities of transitioning between any pair of source and target documents based on analysis of links to target documents of the source document. The system may calculate the transition probabilities based on the number of links on documents a look-ahead distance away. The system then solves for the stationary probabilities of the transition probabilities. The stationary probabilities represent the importance of the documents.

Look-ahead document ranking system

The present invention provides methods for improving a ranking model. In one embodiment, a method includes the step of obtaining queries, documents, and document labels. The process then initializes active sets using the document labels, wherein two active sets are established for each query, a perfect active set and an imperfect active set. Then, the process optimizes an empirical loss function by the use of the first and second active set, whereby parameters of the ranking model are modified in accordance to the empirical loss function. The method then updates the active sets with additional ranking data, wherein the updates are configured to work in conjunction with the optimized loss function and modified ranking model. The recalculated active sets provide an indication for ranking the documents in a way that is more consistent with the document metadata.

Directly optimizing evaluation measures in learning to rank

Efficient search query clustering using tripartite graphs may enable a search engine developer to model information needs of users while expending less computing resources. The efficient clustering of search queries may involve multiple computing devices receiving a subgraph of a multi-partite graph that encompasses search queries, as well as receiving a global center vector table that includes cluster center entries for query clusters. At each computing device, the received global center vector table may be filtered to eliminate one or more cluster center entries that are irrelevant to the search queries. Subsequently, the search queries may be clustered into the query clusters by at least using the filtered global center vector table at each of the computing devices. In some instances, one or more comparisons between search queries and the cluster center entries in the global center vector table during the clustering may be eliminated.

Efficient query clustering using multi-partite graphs

Method for creating a graph representing web browsing behavior, including receiving web browsing behavior data from one or more web browsers; adding a node on the graph for each web page listed in the web browsing behavior data; adding a first link connecting two or more nodes on the graph, wherein the first link representing a hyperlink for accessing a webpage; calculating an amount of time in which each web page is being accessed; determining a number of units of time in the calculated amount of time; adding one or more virtual nodes to the graph based on the number of units of time; and adding a second link connecting two or more virtual nodes on the graph, wherein the second link representing a virtual hyperlink for accessing a webpage.

Calculating a webpage importance from a web browsing graph

Ranking documents based on a series of web graphs collected over time is provided. A ranking system provides multiple transition probability distributions representing different snapshots or times. Each transition probability distribution represents a probability of transitioning from one document to another document within a collection of documents using a link of the document. The ranking system determines a stationary probability distribution for each snapshot based on the transition probability distributions for that snapshot and the stationary probability distribution of the previous snapshot. The stationary probability distributions represent a ranking of the documents over time.

Ranking documents based on a series of document graphs

Embodiments for caching and accessing Directed Acyclic Graph (DAG) data to and from a computing device of a DAG distributed execution engine during the processing of an iterative algorithm. In accordance with one embodiment, a method includes processing a first subgraph of the plurality of subgraphs from the distributed storage system in the computing device. The first subgraph being processed with associated input values in the computing device to generate first output values in an iteration. The method further includes storing a second subgraph in a cache of the device. The second subgraph being a duplicate of the first subgraph. Moreover, the method also includes processing the second subgraph with the first output values to generate second output values if the device is to process the first subgraph in each of one or more subsequent iterations.

Data caching for distributed execution computing

Systems, methods, and devices for sorting and processing various types of graph data are described herein. Partitioning graph data into master data and associated slave data allows for sorting of the graph data by sorting the master data. In another embodiment, promoting a data bucket having a first data bucket size to a data bucket having a second data bucket size greater than the first data bucket size upon reaching a memory limit allows for the reduction of temporary files output by the data bucket.

Growing community of inventors

Beijing, China

Tie-Yan Liu