Microsoft Technology Licensing, LLC

Ebay, Incorporated

Samaneh Abbasi Moghaddam

Marco Pennacchiotti

Thomas Normile

SeyedMohsen Jamali

Vipul C Dalal

Xiaowen Zhang

Revant Kumar

Jerry Louis

Girish Kathalagiri Somashekariah

Xiaoqing Wang

Vinay Praneeth Boda

Jiuling Wang

Chih Cheng Paul Yuan

Lachlan Green

Ali Abbasi

Techniques are provided for using machine learning techniques to learn embeddings for content items. In one technique, training data is used to learn embeddings for each attribute value of multiple attribute values of multiple content items, embeddings for each attribute value of multiple attribute values of multiple entities, and weights for a set of contextual features. In response to receiving a content request, a content item that is associated with one or more targeting criteria that are satisfied based on the content request is identified. A first set of embeddings for the content item are identified, a requesting entity that initiated the content request is identified along with a second set of embeddings for the requesting entity, and a set of feature values for the set of contextual features is identified. The content item is selected based on the sets of embeddings, the set of feature values, and the weights.

Embedded learning for response prediction in content item relevance

Methods, systems, and apparatus for accessing a set of feedback items, identifying a candidate feedback item from the set of feedback items using a lexical pattern, generating a gist phrase that summarizes the candidate feedback item, and causing display of a user interface on a client device, the user interface including the gist phrase.

Mining textual feedback

In an example embodiment, a skip logic using downsampling is applied to negative signals on a training data set fed to a machine-learning algorithm to train a machine-learned model. By downsampling the negatively labeled pieces of training data, the technical problem encountered in biasing the machine-learned model towards negative cases is overcome.

Down-sampling of negative signals used in training machine-learned model

Techniques for using online engagement footprints for video engagement prediction are provided. In one technique, events are received from multiple client devices, each event indicating a type of engagement of a video item from among multiple types of engagement. One or more machine learning techniques are used to train a prediction model that is based on the events and multiple features that includes the multiple types of engagement. In response to receiving a content request, multiple entity feature values are identified for a particular entity that is associated with the content request. Two or more of the entity feature values correspond to two or more of the types of engagement. A prediction is generated based on the entity feature values and the prediction model. The prediction is used to determine whether to select, from candidate content items, a particular content item that includes particular video.

Using online engagement footprints for video engagement prediction

In an example embodiment, a debugging tool is provided that includes logging functionality to allow a machine learned model administrator to replay machine learned recommendation model executions in order to identify points of error, without the scaling difficulties that would be involved in logging all features used in every recommendation.

Debugging tool for recommendation systems

In an example embodiment, cohort-based generalized linear mixed effect model (GLMIX) training is performed to identify patterns across cohorts of users, rather than slicing across all users blindly without accounting for common characteristics of users. Thus, rather than performing GLMIX training at just the finest granular level (e.g., user-level and job-level) or the highest level (global level), a 'medium' level of granularity is used to train the GLMIX model at cohort-level.

Cohort-based generalized linear mixed effect model

Methods, systems, and apparatus for predicting a status of a transaction are described. Feature data related to one or more transactions is collected and a subset of features is selected for use in predicting the status of the transaction. A model is trained using the collected feature data that corresponds to the selected features, and the model is applied to feature data of a selected transaction to generate a probability of the selected transaction attaining one or more defined statuses. Mitigating or preventive actions are performed based on the generated probability.

Predicting a status of a transaction

Methods, systems, and apparatus for mining feedback are described. A set of one or more lexical patterns associated with one or more of a suggestion and a defect report are determined and the set of one or more lexical patterns are matched against a plurality of feedback items to generate a distance learning training set. A distance learning technique is applied to the distance learning training set to generate a distance learning model and the distance learning model is used to identify one or more candidate feedback items of the plurality of feedback items, each of which is one or more of a candidate suggestion and a candidate defect report.

Growing community of inventors

Sunnyvale, CA, United States of America

Samaneh Abbasi Moghaddam