Twitter, Inc.

Yue Lu

Parag Agrawal

Utkarsh Srivastava

Sandeep Pandey

Mike Jahr

Michael Jahr

Jeremy Ginsberg

Ke Zhou

Cheng Li

Sridhar Vishwanathan Iyer

Qicheng Ma

Pengjun Pei

Van Lam

Lanbo Zhang

Bingjun Sun

Zhiyong Xie

Qi Li

Dong Wang

Gary Lam

Feng Zhuge

Shuanghong Yang

Yong Wang

Ziqi Huang

Xiaojiang Guo

In some examples, a computing device includes at least one processor and at least one module, operable by the at least one processor to receive, from a client device of a user, a request for one or more advertisements to display at the client device with a set of messages. The set of messages is associated with the user in a social network messaging service. The at least one module may be further operable to determine a probability that the user will select a candidate advertisement using a machine learning model based on point-wise learning and pair-wise learning. The at least one module may be further operable to determine, based on the probability that the user will select the candidate advertisement, a candidate score for the candidate advertisement, determine that the candidate score satisfies a threshold, and send, for display at the client device, the candidate advertisement.

Click-through prediction for targeted content

A real-time messaging platform allows advertiser accounts to pay to insert candidate messages into the message streams requested by account holders. To accommodate multiple advertisers, the messaging platform controls an auction process that determines which candidate messages are selected for inclusion in a requested account holder's message stream. Selection is based on a bid for the candidate message, the message stream that is requested, and a variety of other factors that vary depending upon the implementation. The process for selection of candidate messages generally includes the following steps, though any given step may be omitted or combined into another step in a different implementation: targeting, filtering, prediction, ranking, and selection.

Pre-filtering of candidate messages for message streams in a messaging platform

An example system includes prediction workers, training workers, and a parameter server. The prediction workers store a local copy of a machine-learned model and run the mode exclusively in serving mode. The training workers store a local copy of a machine-learned model and a local snapshot and run the local copy exclusively in training mode and compare the local model or state to the snapshot after training to send delta updates to the parameter server after training. The parameter server aggregates received delta updates into a master copy of the model, sends the aggregated updates back to training workers and provides two types of updates; a real-time update based on a comparison of the master model with a local snapshot, and a full update. The real-time update occurs at least an order of magnitude more frequently than the full update and includes a subset of the weights in the model.

Efficient online training for machine learning

Real time messaging platform

Pre-filtering in a messaging platform

Growing community of inventors

Redwood City, CA, United States of America

Yue Lu