Yahoo Assets LLC

Yufeng Ma

Kostas Tsioutsiouliklis

Donghyun Kim

Rao Shen

Liuqing Li

Yu Wang

In an example, in response to a request for content, content items may be identified. Content item scores associated with the content items may be determined using a content scoring model. A first ranked list may be generated based upon the content item scores using an exploration model. The first ranked list is indicative of rankings of the content items. A content item selection interface may be generated based upon the first ranked list. The content item selection interface may include selectable objects, corresponding to the content items, arranged according to rankings indicated by the first ranked list.

System and method for ranking content items

The disclosed systems and methods provide a novel framework that provides mechanisms for a Deep & Cross Network (DCN) framework that performs distilled deep prediction for personalized stream ranking on portal websites. The disclosed framework is scalable to satisfy the much more stringent latency and computational requirements required by current network operating environments. The disclosed framework is able to dynamically evaluate and leverage live traffic on network sites in order to provide, update and maintain current recommendations for users as they traverse to a portal and when they navigate within the portal. The disclosed framework implements a DCN model(s) that is capable of being compressed into a model size for a unified optimization within a live traffic environment by combining knowledge distillation and model compression techniques. The disclosed framework is built as a light-weight deep learning model that can be served in production and perform on par with large models.

Computerized system and method for distilled deep prediction for personalized stream ranking

One or more computing devices, systems, and/or methods for debiasing training data based upon information seeking behaviors are provided. Users associated with a set of training data are segmented into information seeking behavior groups corresponding to varying degrees of information seeking behaviors of the users. Biases for the information seeking behavior groups may be estimated based upon information seeking behaviors of users within the information seeking behavior groups. The training set of data is debiased using the biases to generate a debiased training set of data. A model may be trained to perform a task based upon the debiased training set of data.

Debiasing training data based upon information seeking behaviors

In an example, first entities are extracted from user profiles. Second entities are extracted from content information associated with content item. User-associated metrics associated with the first entities are determined based upon the user profiles and/or content events. First vector representations of the first entities and second vector representations of the second entities are processed to generate an attention distribution array. Each value of the attention distribution array represents, for a user interested in an entity of the first entities, a proportion of (i) entity-specific activity, of the user, related to an entity of the second entities relative to (ii) an entirety of activity of the user. An inferred activity distribution array is generated by applying the user-associated metrics to the attention distribution array. A filtered subset of activity distribution values is generated by pruning values from the inferred activity distribution array. Transmission of content is controlled using the filtered subset of activity distribution values.

System and method for feature determination and content selection

The present teaching relates to personalized content recommendation. A webpage is contrasted for a user having a plurality of slots each of which is to be allocated with a content item. For each of the plurality of slots, a plurality of content items in a plurality of types of content are accessed. For each of the plurality of types of content, a personalized score is predicted for each content item in the type of content, wherein the personalized score is obtained based on a trained model trained. A recommended content item of the type of content is selected based on personalized scores. An overall recommended content item is selected and allocated to a slot based on criteria associated with the personalized scores of the recommended content items and a business rule. The webpage with the plurality of slots allocated with content items is provided to the user.

Growing community of inventors

Sunnyvale, CA, United States of America

Yufeng Ma