Yahoo! Inc.

Oath Inc.

Excalibur Ip, LLC

Yahoo Assets LLC

Wei Chu

Deepak Agarwal

Belle L Tseng

Seung-Taek Park

Martin Zinkevich

Achint Oommen Thomas

Lihong Li

Todd Beaupre

Bee-Chung Chen

Scott Roy

Raghu Ramakrishnan

Nitin Motgi

Pradheep K Elango

H Scott Roy

Raghunath Ramakrishnan

Zhaohui Zheng

Anirban Dasgupta

Vishwanath Tumkur Ramarao

Jiang Chen

Jessi Dong

Xiaopeng Xi

Jyh-Shin Shue

Jay Pujara

Wei Du

Uminder Singh

Zhenzhen Kou

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

Methods and systems are disclosed that relate to ranking functions for multiple different domains. By way of example but not limitation, ranking functions for multiple different domains may be trained based on inter-domain loss, and such ranking functions may be used to rank search results from multiple different domains so that they may be blended without normalizing relevancy scores.

Methods and systems relating to ranking functions for multiple domains

This disclosure describes systems and methods for selecting and/or ranking web-based content predicted to have the greatest interest to individual users. In particular, articles are ranked in terms of predicted interest for different users. This is done by optimizing an interest model and in particular through a method of bilinear regression and Bayesian optimization. The interest model is populated with data regarding users, the articles, and historical interest trends that types of users have expressed towards types of articles.

Personalized recommendations on dynamic content

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

A set of item-item affinities for a plurality of items is determined based on collaborative-filtering techniques. A set of an item's nearest neighbor items based on the set of item-item affinities is determined. A set of user feature-item affinities for the plurality of items and a set of user features is determined based on least squared regression. A set of a user feature's nearest neighbor items is determined based in part on the set of user feature-item affinities. Compatible affinity weights for nearest neighbor items of each item and each user feature are determined. Based on user features of a user and items a user has consumed, a set of nearest neighbor items are identified as a set of candidate items, and affinity scores of candidate items are determined. Based on the affinity scores, a candidate item from the set of candidate items is recommended to the user.

Determining user preference of items based on user ratings and user features

Content items are selected to be displayed on a portal page in such a way as to maximize a performance metric such as click-through rate. Problems relating to content selection are addressed, such as changing content pool, variable performance metric, and delay in receiving feedback on an item once the item has been displayed to a user. An adaptation of priority-based schemes for the multi-armed bandit problem, are used to project future trends of data. The adaptation introduces experiments concerning a future time period into the calculation, which increases the set of data on which to solve the multi-armed bandit problem. Also, a Bayesian explore/exploit method is formulated as an optimization problem that addresses all of the issues of content item selection for a portal page. This optimization problem is modified by Lagrange relaxation and normal approximation, which allow computation of the optimization problem in real time.

Enhanced matching through explore/exploit schemes

Two items are determined to be similar to each not only based on previous actual user behavior, but also based on the observed relatedness of the characteristics of those two items. A first characteristic and a second characteristic are determined to have some affinity for each other if a high proportion of users who select items having the first characteristics also select items that have the second characteristic, and vice-versa. Two items having characteristics with high affinity for each other are determined to have some similarity to each other, even if very few or no users who selected one of those items ever selected the other of those items. A first item that is determined to be sufficiently similar to second item in this manner may be recommended to a user who has selected the second item as potentially also being of interest to that user.

Predicting item-item affinities based on item features by regression

A set of item-item affinities for a plurality of items is determined based on collaborative-filtering techniques. A set of an item's nearest neighbor items based on the set of item-item affinities is determined. A set of user feature-item affinities for the plurality of items and a set of user features is determined based on least squared regression. A set of a user feature's nearest neighbor items is determined based in part on the set of user feature-item affinities. Compatible affinity weights for nearest neighbor items of each item and each user feature are determined and stored. Based on user features of a particular user and items a particular user has consumed, a set of nearest neighbor items comprising nearest neighbor items for user features of the user and items the user has consumed are identified as a set of candidate items, and affinity scores of candidate items are determined. Based at least in part on the affinity scores, a candidate item from the set of candidate items is recommended to the user.

Content items are selected to be displayed on a portal page in such a way as to maximize a performance metric such as click-through rate. Problems relating to content selection are addressed, such as changing content pool, variable performance metric, and delay in receiving feedback on an item once the item has been displayed to a user. An adaptation of priority-based schemes for the multi-armed bandit problem are used to project future trends of data. The adaptation introduces experiments concerning a future time period into the calculation, which increases the set of data on which to solve the multi-armed bandit problem. Also, a Bayesian explore/exploit method is formulated as an optimization problem that addresses all of the issues of content item selection for a portal page. This optimization problem is modified by Lagrange relaxation and normal approximation, which allow computation of the optimization problem in real time.

Growing community of inventors

Sunnyvale, CA, United States of America

Wei Chu