Microsoft Technology Licensing, LLC

Ajith Muralidharan

Ankan Saha

Shaunak Chatterjee

Shipeng Yu

Yan Gao

Yiping Yuan

Deepak Agarwal

Yijie Wang

David Benjamin Liu

Yi Zhen

Yiming Ma

Sandor Nyako

Pratik Daga

Matthew Walker

Nirav Nalinbhai Shingala

Matthew Hsing Hung Walker

Swetha Nagabhushan Karthik

Florian Raudies

Jiaqi Ge

Zhongen Tao

Ying Han

Bethany J Wang

Xuexin Ren

Viral Gupta

Padmini Jaikumar

Preetam Nandy

Miao Cheng

Gargi Harish Bhandari

Ivan Lopez Moreno

Adriel Fuad

Rakesh Malladi

Yingkai Hu

Ruiqi Wang

Prakruthi Prabhakar

In an example embodiment, tensor train decompositions are used to create large, personalized layers that are efficient for segment modeling. More particularly, rather than performing learning on an input matrix of training data that contains all segments, and then crossing this matrix with a vector for a particular segment, the matrix is mapped to an N-dimensional tensor, where each of the dimensions corresponds to one of the properties used to compose the segment, which can then be approximated by tensor train decomposition to enable efficient training and scoring.

Segment modeling for machine learning using tensor train decompositions

Methods, systems, and computer programs are presented for selecting notifications based on an affinity score between a content generator and a viewer of the content. One method includes capturing interactions of content generators with notifications, received by the content generators, associated with viewer responses to creator-generated content items. The method further includes training a machine-learning model based on the interactions, and detecting a first set of notifications, for a first content generator, associated with interactions of a set of viewers to first-content generator content. The ML model calculates an affinity score between the first content generator and each viewer, and the set of first notifications are ranked based on the affinity scores of the first content generator and the viewer associated with each notification. A set of second notifications is selected based on the ranked first notifications; and generating notifications are generated, for the first content generator, for the selected set of second notifications.

Leveraging affinity between content creator and viewer to improve creator retention

Technologies for unseen notification handling are described. Embodiments select an initial set of notifications, provide the selected initial set of notifications to a client device, store seen notifications in a first data store, maintain sent but unseen notifications in a second data store that is an in-memory online data store, retrieve a set of the sent but unseen notifications from the second data store, create a list of unseen notifications by combining the retrieved set of sent but unseen notifications with a set of unsent and unseen notifications, generate a set of relevance scores for the list of unseen notifications, create a new version of the list of unseen notifications based on the new set of relevance scores, and provide the new version of the list of the unseen notifications to the client device.

Read-time relevance-based unseen notification handling

Methods, systems, and computer programs are presented for scheduling user notifications to maximize short-term and long-term benefits from sending the notifications. One method includes an operation for identifying features of a state used for reinforcement learning. The state is associated with an action to decide if a notification to a user is to be sent and a reward for sending the notification to the user. Further, the method includes capturing user responses to notifications sent to users to obtain training data and training a machine-learning (ML) algorithm with reinforcement learning based on the features and the training data to obtain an ML model. Additionally, the method includes receiving a request to send a notification to the user, and deciding, by the ML model, whether to send the notification based on a current state. The notification is sent to the user based on the decision.

User-notification scheduling

Technologies for generating dynamic notification content for notification messages using a machine learned model are provided. The disclosed techniques include identifying an event related to a particular user, where the event has a particular notification type that represents a subject type of the event. Based on the particular notification type of the event, a set of candidate headline and call-to-action combinations corresponding to the particular notification type are identified. Using the machine learned model, scores are calculated for each headline and call-to-action combination in the set of candidate headline and call-to-action combinations. One or more particular headline and call-to-action combinations from the set of candidate headline and call-to-action combinations are selected based upon the scores calculated for each combination of the set of candidate headline and call-to-action combinations. A notification message is generated for the event that includes the one or more particular headline and call-to-action combinations selected.

Machine learning driven dynamic notification content

Methods, systems, and computer programs are presented for providing a user experience that facilitates navigation among different topics and articles on a social network. One method includes an operation for identifying a hierarchy of topics, each topic corresponding to a respective subject, where the hierarchy defines relationships between the topics. A first topic page for a first topic is presented in a user interface in the social network. The first topic page includes articles and first options for navigating to topic pages of topics related to the first topic. The method further includes detecting a selection of a first article. In response to detecting the selection, a first article page for the first article is presented in the user interface. The first article page includes details of the first article and second options for navigating to topic pages of topics related to the first article.

Social network navigation based on content relationships

A system, a machine-readable storage medium storing instructions, and a computer-implemented method as described herein are directed to a Personalized Article Engine that generates respective prediction models for each article in a plurality of candidate articles in a social network system. The Personalized. Article Engine generates a respective article score according to each article's prediction model and at least one feature of a target member account. The Personalized Article Engine generates a plurality of output scores based on combining each respective article score with a corresponding article's global model score. The Personalized Article Engine ranks the output scores to identify a subset of candidate articles relevant to the target member account.

Per-article personalized models for recommending content email digests with personalized candidate article pools

A system, a machine-readable storage medium storing instructions, and a computer-implemented method as described herein are directed to a Fast Ranker Engine that identifies global model features present in an article in a social network service. The Fast Ranker Engine assembles respective fixed vectors based on at least one member account feature and each coefficient that corresponds to a present global article feature of the global model. The Fast Ranker Engine generates a transformation feature(s) for a prediction model of the article based on the respective fixed vectors.

Per-article personalized model feature transformation

Technologies for determining whether to send notification messages, from different sources, to a target user are provided. The disclosed techniques include receiving a first notification event from a first notification service and receiving a second notification event from a second notification service. The first and second notification services are different services. Using a machine-learned model to assign a first score to the first notification event and a second score to the second notification event. Based on the first score, a determination is made to generate a first notification message for the first notification event. The first notification message is then sent to a target user. Based on the second score, a determination is made not to generate a second notification message for the second notification event.

Multi-objective, multi-input, efficient decoupling holistic platform for communication selection

Techniques for identifying and delivering notifications of user-generated content to network-limited users are provided. In one technique, for each selected target entity that has a limited network, one or more topics associated with the target entity are identified and the target entity is assigned to one or more entity-topic buckets for the identified topics. For each selected content item, one or more topics associated with the content item are identified and the content item is assigned to one or more content-topic buckets for the identified topics. The entity-topic buckets are matched to the content-topic buckets, resulting in assigning, for each selected target entity, zero or more content items to that target entity. For each target entity that is assigned one or more content items based on the matching, a notification is generated and transmitted over a computer network to a computing device of the target entity.

Out-of-network notifications of network-transmitted content items

In an example embodiment, a machine learned model is used to determine whether to send a notification for a feed object to a user. This machine learned model is optimized not just based on the likelihood that the notification will cause the user to interact with the feed object, but also the likely short-term and long-term impacts of the user interacting with the feed object. This machine learned model factors in not only the viewer's probability of immediate action, such as clicking on a feed object, but also the probability of long-term impact, such as the display causing the viewer to contribute content to the network or the viewer's response encouraging more people to contribute content to the network. As such, the machine learned model is optimized not just on notification interactivity but also on feed objects interactivity.

Joint optimization of notification and feed

The present disclosure describes various embodiments of methods, systems, and machine-readable mediums which help determine a user's likely affinity for consuming content (such as an article) in a particular language presented (or to be presented) in a heterogeneous feed of a social network.

Determining viewer language affinity for multi-lingual content in social network feeds

The present disclosure describes various embodiments of methods, systems, and machine-readable mediums which help determine a user's likely affinity for an article presented (or to be presented) in a heterogeneous feed of a social network.

Determining viewer affinity for articles in a heterogeneous content feed

Technologies for determining whether to send a notification to an entity is provided. Disclosed techniques include receiving entity features describing attributes related to observed entity sessions. A set of entity-specific session features values may be generated from the received entity features. A session-quality prediction model may be generated using the set of entity-specific session feature values. The session-quality prediction model may determine an expected session score for a new entity session for an entity, where the expected session score describes a level of interaction for the new entity session. A notification may be received for a particular entity. The session-quality prediction model may be used to determine the expected session score for a new entity session for the particular entity. A determination may be made as to whether a notification should be sent to the particular entity based upon the expected session score for the new entity session.

Driving high quality sessions through optimization of sending notifications

Disclosed are systems, methods, and non-transitory computer-readable media for automated profile image generation based on scheduled video conferences. A profile image generation system generates, based on image data captured during a first video conference, a first facial feature data set for a first identified face identified from the image data. The first facial feature data set includes numeric values representing the first identified face. The profile image generation system calculates, based on the first facial feature data set and historic facial feature data sets generated from image data captured during previous video conferences, a first value indicating a likelihood that the first identified face is a first meeting participant that participated in the first video conference. The profile image generation system determines that the first value meets or exceeds a threshold value, and in response, determines that the first identified face is the first meeting participant.

Growing community of inventors

Sunnyvale, CA, United States of America

Ajith Muralidharan