Microsoft Technology Licensing, LLC

Boyi Chen

Yang Yang

Zheng Li

Yanbin Jiang

Shaunak Chatterjee

Qi He

Bo Long

Weidong Zhang

Jaewon Yang

Bee-Chung Chen

Kevin Chang

Grace W Tang

Joel Daniel Young

Tie Wang

Yijie Wang

Ying Xuan

David J Stein

Timothy Paul Jurka

Wei Zhuang

Pannagadatta K Shivaswamy

Wenxuan Gao

Paul Taylor Ogilvie

Amit Chandak

Shilpa Gupta

Kexin Nie

Priyanka Gariba

Yunsong Meng

Yen-Jung Chang

Baohua Huang

Varnit Agnihotri

Divya Gadde

Ruoyang Wang

Amit Yadav

Yangchun Luo

Ke Wu

The disclosed embodiments provide a system for streamlining machine learning. During operation, the system determines a resource overhead for a baseline version of a machine learning model that uses a set of features to produce entity rankings and a number of features to be removed to lower the resource overhead to a target resource overhead. Next, the system calculates importance scores for the features, wherein each importance score represents an impact of a corresponding feature on the entity rankings. The system then identifies a first subset of the features to be removed as the number of features with lowest importance scores and trains a simplified version of the machine learning model using a second subset of the features that excludes the first subset of the features. Finally, the system executes the simplified version to produce new entity rankings.

Feature removal framework to streamline machine learning

The disclosed embodiments provide a system for processing data. During operation, the system obtains a feature configuration for a feature. Next, the system obtains, from the feature configuration, an anchor containing metadata for accessing the feature in an environment. The system then uses one or more attributes of the anchor to retrieve one or more feature values of the feature from the environment. Finally, the system provides the one or more feature values for use with one or more machine-learning models.

Framework for managing features across environments

Technologies for monitoring performance of a machine learning model include receiving, by an unsupervised anomaly detection function, digital time series data for a feature metric; where the feature metric is computed for a feature that is extracted from an online system over a time interval; where the machine learning model is to produce model output that relates to one or more users' use of the online system; using the unsupervised anomaly detection function, detecting anomalies in the digital time series data; labeling a subset of the detected anomalies in response to a deviation of a time-series prediction model from a predicted baseline model exceeding a predicted deviation criterion; creating digital output that identifies the feature as associated with the labeled subset of the detected anomalies; causing, in response to the digital output, a modification of the machine learning model.

Machine learning model monitoring

In an example embodiment, a GLMix model is utilized that models viewers and actors of feed items. This allows for random effects of individual viewers and actors to be taken into account without introducing biases. Additionally, in an example embodiment, predictions/recommendations are made more accurate by using three models, which are then combined, instead of a single GLMix model. Each of these models has different granularities and dimensions. A global model may model the similarity between user attributes (e.g., from the member profile or activity history) and item attributes. A per-viewer model may model user attributes and activity history of actors on feed items. A per-actor model may model user attributes and activity history of the viewers of feed items. The per-actor model may therefore, rely on information regarding how and what type of viewers interacted with items acted on by the particular actor.

Personalized per-member model in feed

A news feed system provided with an on-line social network system determines that a news feed is to be constructed for a viewer. The news feed system accesses the viewer's profile and other information associated with the viewer, accesses an inventory of activities that have been identified as potentially of interest to the viewer, and calculates relevance score for each item inventory of activities using the combined predictions methodology. The activities are then arranged for presentation to the viewer via a news feed web page, using respective calculated relevance scores.

Combined predictions methodology

A system and method for personalizing cross session diversity is disclosed. The system receives a member opportunity request. In response, the system generates a list of members in response to the received member opportunity request, wherein the list of members is determined based on member profile data stored at a social networking system. For each member in the generated list of members, the system generates a profile value score based on the stored member profile data. The system ranks the members of the generated list at least in part based on the generated profile value scores. The system then selects one or more members in the list of members based on the ranking of members in the generated list.

Personalized cross session diversity

Growing community of inventors

Sunnyvale, CA, United States of America

Boyi Chen