Microsoft Technology Licensing, LLC

Chenxi Lin

Jian Wang

Zheng Chen

Benyu Zhang

Hua-Jun Zeng

Lei Li

HuaJun Zeng

Qiaoling Liu

Min Wu

Lei Ji

Bing Jiang Sun

Jie Han

Ke Tang

Gui-Rong Xue

Guirong Xue

Methods for determining a predictive rating are disclosed. In an embodiment, an active user is compared to a set of clusters. One or more of the clusters are determined to be most similar to the active user. From the one or more clusters, K users are determined to be most similar to the active user. Prior ratings for an item by the K users may be used to predict a rating for the item for the active user.

Cluster-based scalable collaborative filtering

A method and system for use in information retrieval includes, for each of a plurality of terms, selecting a predetermined number of top scoring documents for the term to form a corresponding document set for the term. When a plurality of terms are received, optionally as a query, the system ranks, using an inverse document frequency algorithm, the plurality of terms for importance based on the document sets for the plurality of terms. Then a number of ranked terms are selected based on importance and a union set is formed based on the document sets associated with the selected number of ranked terms.

Efficient retrieval algorithm by query term discrimination

A scalable two-pass scalable probabilistic latent semantic analysis (PLSA) methodology is disclosed that may perform more efficiently, and in some cases more accurately, than traditional PLSA, especially where large and/or sparse data sets are provided for analysis. The improved methodology can greatly reduce the storage and/or computational costs of training a PLSA model. In the first pass of the two-pass methodology, objects are clustered into groups, and PLSA is performed on the groups instead of the original individual objects. In the second pass, the conditional probability of a latent class, given an object, is obtained. This may be done by extending the training results of the first pass. During the second pass, the most likely latent classes for each object are identified.

Scalable probabilistic latent semantic analysis

Described is an efficient retrieval mechanism that quickly locates documents (e.g., corresponding to online advertisements) based on query term discrimination. A topmost subset (e.g., two) of search terms is selected according to their ranked importance, e.g., as ranked by inverted document frequency. The topmost terms are then used to narrow the number of rows of an inverted query index that are searched to find document identifiers and associated scores, such as computed offline by a BM25 algorithm. For example, for each document identifier of each important term, a fast search within each of the narrowed subset of rows (that also contain that document identifier) may be performed by comparing document identifiers to jump a pointer within each other row, followed by a binary search to locate a particular document. The scores of the set of particular documents may then be used to rank their relative importance for returning as results.

Described is a technology by which blocks of web pages may be selected, such as for building a user-personalized web page containing selected blocks. A selection mechanism, such as a browser toolbar add-on, provides a user interface for selecting blocks, and records information about selected blocks. A block tracking mechanism (e.g., a daemon program) uses the information to locate selected blocks of the web pages, including when the web page containing the block is updated with respect to content and/or layout. The block tracking mechanism may update a local gadget that when invoked, such as by browsing to a particular web page, which shows updated versions of the block on a personalized web page. Blocks may be efficiently located by processing trees representing web pages into reduced trees, and then by performing a minimum distance mapping algorithm on the reduced trees.

Block tracking mechanism for web personalization

Described is a behavioral targeting technology for online advertising, by which an original attribute is uniformly expanded. Users that meet an original attribute are aggregated into a mid-result used to determine similarity relative to candidate attribute types. The most similar candidate attributes are selected for the expanded attribute. A URL/URL pattern suggestion technology is provided, with similarity computed from users/URLs visited by the users. URLs are separated into URL tree nodes, for calculating the number of users who have visited each URL and the number of users who have visited the URL on a sub-tree whose root is the node. URL/URL patterns are output based on similarity. Domains are also suggested based on user-visits. Similarities between pairs of domains may be computed (e.g., offline), with an output for a given domain provided in based on its similarity with each other domain.

User segment suggestion for online advertising

Extraction of semantic information and the generation of semantic attributes allows for improved organization and management of data. Semantic attributes are automatically generated and eliminate the need for manual entry of attribute information. A semantic file network may further be constructed based on similarities between files that are based on the semantic attribute information. Semantic links representing a semantic relationship may be built between similar or relevant files. In addition, user operations and user operation patterns may also be considered in building the file network. Semantic attributes and information may further facilitate browsing the file systems as well as improve the accuracy and speed of queries.

Adaptive grouping in a file network

System and method for exploring a semantic file network

A method of creating a personal home page containing information of interest assembled from various web sites. The method includes the partitioning of web pages into web blocks. Users may collect various web blocks from different web pages and utilize those web blocks to define the dynamic personal homepage. In addition, the web blocks may be tracked to update content in the personal home page based on corresponding changes in the original web page.

Creating home pages based on user-selected information of web pages

The invention provides a method of interactively crawling data records on a web page. Users may select various data records of interest on a web page to generate templates to search for similar data items on the same web page or on different web pages. A tree matching algorithm may be used to compare and extract data matching the generated template.

Growing community of inventors

Beijing, China

Chenxi Lin