Microsoft Technology Licensing, LLC

University of California

Alexandros Ntoulas

Stelios Paparizos

Sreenivas Gollapudi

John Christopher Shafer

Samuel Ieong

Wei Liu

Christopher Walter Anderson

Ajay Nair

Samira Tasharofi

Naga Srinivas Vemuri

Krishnaram Kenthapadi

Robert Wagner

Marc Alexander Najork

Mark Steven Manasse

Livia Polanyi

Nina Mishra

Tao Cai

Dennis Craig Fetterly

Yue Lu

Panayiotis Tsaparas

Yingfeng Oh

Nirav Nalinbhai Shingala

Rohan Rajiv

Krishnaram N G Kenthapadi

Kirill Sergeyevich Lebedev

Aleksandra Korolova

Xiaoxu Wu

Pei Qun Yu

Shuai Ding

Christopher D Erbach

Pallavi Agarwal

Junghoo Cho

Petros Zerfos

Techniques for optimizing content item delivery for installations or activations of a mobile application are provided. In one technique, a machine-learned model is trained based on multiple training instances that individually indicate whether an entity performed a particular action relative to a mobile application. In response to receiving a content item request from a third-party content delivery exchange, it is determined whether a client device that initiated the content item request has activated a particular application. In response to determining that the client device has not activated the particular application, multiple feature values of the content item request are identified. Based on inputting the feature values into the model, a score is generated that indicates a likelihood that an entity of the client device will perform the particular action relative to the particular application. Based on the score, a content item is transmitted over a network to the client device.

Optimizing content item delivery for installations of a mobile application

Techniques for efficiently detecting similarity among electronic content items are provided. A vector is generated for each of multiple content items and is used to assign its corresponding content item to a group among multiple groups. A set of content items that is assigned to a particular group of the plurality of groups is identified. For each pair of content items in the set, a similarity score is generated and used to determine whether to classify the pair as similar to each other. Similarity data is updated if to identify the pair if it is determined to classify the pair of content items as similar to each other. The similarity data associates a first content item with a second content item. The second content item is prevented from being presented to a particular user for a period of time after the first content item is presented to the particular user.

Content item similarity detection

Techniques for integrating two server systems are provided. A first server system receives, from a first computing device, provider data generated by a third-party tracking service. The first server system stores association data that associates the provider data with one or more content items. After storing the association data, the first server system receives, from a second computing device, a message that includes a first identifier and that indicates an event and a content item. In response to receiving the message, the first server system determines whether to report the event to the third-party tracking service. If so, the first server system determines a second identifier based on the first identifier, retrieves, based on the association data and the content item, the provider data that is associated with the content item, and transmits, over a network to the third-party tracking service, the second identifier and the provider data.

Server-to-server integration of tracking events

A structured query may specify attribute values for attributes. An estimate of the number of items that will match the structured query if it is applied to a structured database is determined. If the estimated number of items is below a threshold, the structured query may be relaxed to form new candidate structured queries. The number of candidate queries may be determined based on a desired running time. Each of the candidate structured queries may be determined by changing one or more attribute values of the attributes of the structured query. Estimates of the number of items each of the candidate structured queries will match is determined, and the candidate structured query that has the highest matching estimation is used to query the database. The matching results may be output.

Relaxation for structured queries

User generated reviews and scores associated with the reviews may be used to train a review scoring model with textual features of the reviews. The review scoring model may be used to predict scores for newly received reviews. One or more constraints based on social networking application data associated with an author of a review may be used to adjust the predicted score of the review.

Assessing quality of user reviews

Described is answering online web queries from a structured data store, such as a products database. An online web search query is modified into an evaluation expression for accessing a structured data store (e.g., a database, XML file, a flat table) to find search results. A query is matched to a pattern, which is used to generate the expression, as well as determine to which structured data store the query is routed. Tokens (e.g., words) in the query are processed against a dictionary of token classes (sets of tokens) and patterns (sets of token classes) to annotate the query to map the query to the matching pattern. A translation process generates the expression based on translation hints that correspond to the matching pattern. Also described is offline mining of query logs to generate the dictionaries.

Answering web queries using structured data sources

Described is releasing output data representing a search log, in which the data is suitable for most data mining/analysis applications, but is safe to publish by preserving user privacy. The search log is processed such that a query is only included if a sufficient count of that query is present; noise may be added. User contributions that are considered may be limited to a maximum number of queries. The output may indicate how often (possibly plus noise) that each query appeared. Other output may comprise a query-action graph, a query-inaction graph and/or a query-reformulation graph, with nodes representing queries and nodes representing actions, inactions or reformulations (e.g., clicked URLs, skipped URLs, or selected related queries), and edges between nodes representing action, skip or selection counts (possibly plus noise). The output may correspond to the top results/related queries returned from a search.

Synopsis of a search log that respects user privacy

Documents are replicated among servers comprising a search engine based on the value of each document by approximating its value as one of the top search results for one or more exemplary queries. Documents are allocated among servers comprising a search engine by calculating a relevance value for each document and then distributing the documents evenly to the servers. A subset of servers are selected from among a plurality of servers comprising a search engine using term-based, server-specific histograms reflecting the number of instances of the term in each document allocated to each server, and then selecting servers to service a query based on the documents on those servers.

Indexing for limited search server availability

Compression of extensive, rule-based grammars used to facilitate search queries is provided herein. Rule-based grammars include a list of rules that each comprise a sequence of token classes. Each token class is a logical grouping of tokens, and each token is a string of characters. A grammar is parsed to identify rules and token classes. Unimportant token classes are identified and sets of unimportant token classes are merged to generated merged token classes. A compressed grammar is generated by substituting the merged token classes into the grammar for corresponding unimportant token classes used to generate the merged token classes.

Grammar compression

Compression of extensive, rule-based grammars used to facilitate search queries is provided herein. Rule-based grammars includes a list of rules that each comprise a sequence of token classes. Each token class is a logical grouping of tokens, and each token is a string of characters. A grammar is parsed to identify rules and token classes. Unimportant token classes are identified and sets of unimportant token classes are merged to generated merged token classes. A compressed grammar is generated by substituting the merged token classes into the grammar for corresponding unimportant token classes used to generate the merged token classes.

Evaluating content includes receiving content, analyzing the content for web spam using a content-based identification technique, and classifying the content according to the analysis. An index of analyzed contents may be created. A system for evaluating content includes a storage device configured to store data and a processor configured to analyze content for web spam using content-based identification techniques.

Using content analysis to detect spam web pages

A method and system for autonomously downloading and indexing Hidden Web pages from Websites includes the steps of selecting a query term and issuing a query to a site-specific search interface containing Hidden Web pages. A results index is then acquired and the Hidden Web pages are downloaded from the results index. A plurality of potential query terms are then identified from the downloaded Hidden Web pages. The efficiency of each potential query term is then estimated and a next query term is selected from the plurality of potential query terms, wherein the next selected query term has the greatest efficiency. The next selected query term is then issued to the site-specific search interface using the next query term. The process is repeated until all or most of the Hidden Web pages are discovered.

Growing community of inventors

Mountain View, CA, United States of America

Alexandros Ntoulas