International Business Machines Corporation

Yahoo! Inc.

Oath Inc.

Nec Laboratories America, Inc.

Zynga Inc.

Yahoo Assets LLC

Belle L Tseng

Zon-Yin Shae

John Richard Smith

Ferdinand Hendriks

Xiping Wang

Anirban Dasgupta

Achint Oommen Thomas

Wei Chu

Martin Zinkevich

Ching-Yung Lin

Liang Zhang

Yi Ke Wu

Vishwanath Tumkur Ramarao

Lihong Li

Jyh-Shin Shue

Ya Zhang

Lyndon Kennedy

Srinivas Vadrevu

Xiaodan Song

Roelof Van Zwol

Maxim Gurevich

Sharat Narayan

Markus Weimer

Gordon Guo-Zheng Sun

Young Maeng

Nicolas Torzec

Fuad Elias Doany

Shanmugasundaram Ravikumar

Larry Paul Heck

Xin Li

Milind R Naphade

Christian Alexander Lang

Yun Chi

Bo Long

Chalapathy Venkata Neti

Jean-Marc Langlois

Apostol Ivanov Natsev

Kunal Punera

David Mark Ciemiewicz

Xiaopeng Xi

Jay Pujara

Sudarshan Lamkhede

Hugh William Adams, Jr

Chung Kei Wong

Benjamin Graham

Koji Hino

Keke Chen

Arun A Prakash

Rongqing Lu

Giridharen Iyengar

Hung-Kuo (James) Chu

Software for online active learning receives content posted to an online stream at a website. The software converts the content into an elemental representation and inputs the elemental representation into a probit model to obtain a predictive probability that the content is abusive. The software also calculates an importance weight based on the elemental representation. And the software updates the probit model using the content, the importance weight, and an acquired label if a condition is met. The condition depends on an instrumental distribution. The software removes the content from the online stream if a condition is met. The condition depends on the predictive probability, if an acquired label is unavailable.

Online active learning in user-generated content streams

Embodiments are directed towards multi-level entity classification. An object associated with an entity is received. In one embodiment the object comprises and email and the entity comprises the IP address of a sending email server. If the entity has already been classified, as indicated by an entity classification cache, then a corresponding action is taken on the object. However, if the entity has not been classified, the entity is submitted to a fast classifier for classification. A feature collector concurrently fetches available features, including fast features and full features. The fast classifier classifies the entity based on the fast features, storing the result in the entity classification cache. Subsequent objects associated with the entity are processed based on the cached result of the fast classifier. Then, a full classifier classifies the entity based on at least the full features, storing the result in the entity classification cache.

Anti-spam transient entity classification

Embodiments are directed towards generating a unified user account trustworthiness system through user account trustworthiness scores. A trusted group of user accounts may be identified for a given action by grouping a plurality of user accounts into tiers based on a trustworthiness score of each user account for the given action. The tiers and/or trustworthiness scores may be employed to classify an item, such as a message as spam or non-spam, based on input from the user accounts. The trustworthiness scores may also be employed to determine if a user account is a robot account or a human account. The trusted group for a given action may dynamically evolve over time by regrouping the user accounts based on modified trustworthiness scores. A trustworthiness score of an individual user account may be modified based on input received from the individual user account and input from other user accounts.

User trustworthiness

Techniques are provided for broadcasting history or coarse representations or both of stroke media and other media. Levels are created for media, and the levels are periodically broadcast. The levels segment the media and allow the media to be examined with details that range from coarse to fine or examined with varying importance or both. The periodic broadcast of the levels provides history of the media. In particular, for stroke data, levels of detail are determined through a certain method. Each level of detail is periodically broadcast. The levels of detail may be shown by themselves or combined at a receiver to create segmented views of the stroke data. A user can select a coarse view of the stroke data to quickly see the stroke history. Additionally, techniques are provided for quickly and repeatedly accessing data elements in already existing data structures. Generally, a description is created that describes data elements in a preexisting data structure. One or more links are created, from the description, which allow a user to directly access particular data elements for the preexisting data structure.

Periodic broadcast and location of evolving media content with application to seminar and stroke media

Software for supervised learning extracts a set of pixel-level features from each source image in collection of source images. Each of the source images is associated with a thumbnail created by an editor. The software also generates a collection of unique bounding boxes for each source image. And the software calculates a set of region-level features for each bounding box. Each region-level feature results from the aggregation of pixel values for one of the pixel-level features. The software learns a regression model, using the calculated region-level features and the thumbnail associated with the source image. Then the software chooses a thumbnail from a collection of unique bounding boxes in a new image, based on application of the regression model. The software uses a thumbnail received from an editor instead of the chosen thumbnail, if the chosen thumbnail is of insufficient quality as measured against a scoring threshold.

Image cropping using supervised learning

A method of generating a time managed challenge-response test is presented. The method identifies a geometric shape having a volume and generates an entry object of the time managed challenge-response test. The entry object is overlaid onto the geometric shape, such that the entry object is distributed over a surface of the geometric shape, and a portion of the entry object is hidden at any point in time. The geometric shape is rotated, which reveals the portion of the entry object that is hidden. A display region on a display is identified for rendering the geometric shape and the geometric shape is presented in the display region of the display.

Time managed challenge-response test

Software for supervised learning extracts a set of pixel-level features from each source image in collection of source images. Each of the source images is associated with a thumbnail created by an editor. The software also generates a collection of unique bounding boxes for each source image. And the software calculates a set of region-level features for each bounding box. Each region-level feature results from the aggregation of pixel values for one of the pixel-level features. The software learns a regression model, using the calculated region-level features and the thumbnail associated with the source image. Then the software chooses a thumbnail from a collection of unique bounding boxes in a new image, based on application of the regression model.

Embodiments are directed towards clustering cookies for identifying unique mobile devices for associating activities over a network with a given mobile device. The cookies are clustered based on a Bayes Factor similarity model that is trained from cookie features of known mobile devices. The clusters may be used to determine the number of unique mobile devices that access a website. The clusters may also be used to provide targeted content to each unique mobile device.

Clustering cookies for identifying unique mobile devices

Various users' navigational behaviors relative to search results presented by a search engine are monitored. URLs that are visited and revised queries that are submitted after the submission of an original query are placed within a trail that begins with the original query. These trails are grouped based on the original queries with which they begin. For each trail group, a set of URLs that frequently occur in that group's trails, and a set of revised queries that frequently occur in that group's trails, are determined. These frequently occurring elements are mapped to the original queries with which all the trails in the corresponding trail group begin. In response to subsequent submissions of the same original query, the search engine ensures that URLs and revised queries that are mapped to the original query are prominently displayed on the search results pages that are initially returned in response to those submissions.

Automated user behavior feedback system for whole page search success optimization

Embodiments presented herein provide methods, systems and computer program products for determining a count of network users. One method identifies one or more login access requests, from one or more server logs. Each of the one or more login access requests comprises a login cookie, and a user identifier. The method then forms one or more connected components based on the one or more login access requests. Each connected component comprises one or more user identifiers, wherein each of the one or more user identifiers is associated with one or more login cookies. The method finally determines a count of login users based, at least in part, on the user identifiers in the one or more connected components.

System and method for counting network users

According to an example embodiment, a method comprises executing instructions by a special purpose computing apparatus to, for labeled source domain data having a plurality of original labels, generate a plurality of first predicted labels for the labeled source domain data using a target function, the target function determined by using a plurality of labels from labeled target domain data. The method further comprises executing instructions by the special purpose computing apparatus to apply a label relation function to the first predicted labels for the source domain data and the original labels for the source domain data to determine a plurality of weighting factors for the labeled source domain data. The method further comprises executing instructions by the special purpose computing apparatus to generate a new target function using the labeled target domain data, the labeled source domain data, and the weighting factors for the labeled source domain data, and evaluate a performance of the new target function to determine if there is a convergence.

System and method for cross domain learning for data augmentation

Growing community of inventors

Forest Hills, NY, United States of America

Belle L Tseng