Facebook, Inc.

Meta Platforms, Inc.

David Vickrey

Khalid Bakry El-Arini

Jeffrey William Pasternack

Lars Seren Backstrom

Wayne Kao

Robert William Cathcart

Erich James Owens

Romain Thibaux

Alexey Maykov

Max Gubin

Yigal Dan Rubinstein

Prasoon Mishra

Gregory Matthew Marra

Varun Kacholia

Guy Dassa

Holly Marie Ormseth

Sahil Prakash Thaker

Shengbo Guo

Mahmud Sami Tas

Max Christian Eulenstein

Annie Hsin-Wen Liu

Yuanxuan Wang

Jianfu Chen

Yue Zhuo

Sanjeet Uday Hajarnis

Elad Gerson

Dan Yigal Rubinstein

Austen Norment McDonald

Kritarth Anand

Polina Kuznetsova

Gaurav Shankar

Justin MacLean Coughlin

Joyce Tang

A social networking system generates a newsfeed for a user to view when accessing the social networking system. Candidate stories associated with users of the social networking system are selected and an expected value score for each candidate story is determined. An expected value score is based on the probability of a user performing various types of interactions with a candidate story and a numerical value for each type of interaction. The numerical value for a type of interaction represents a value to the social networking system of the type of interaction. Based on the expected value scores, the candidate stories are ranked and the ranking used to select candidate stories for the newsfeed.

Arranging stories on newsfeeds based on expected value scoring on a social networking system

An online system predicts topics for content items. The online system provides one or more topic labels for a user to apply concurrently while a user is composing a post, in response to requests periodically received from the user's device. A request includes information such as content composed by the user and contextual information. The online system employs machine learning techniques to analyze content composed by a user and contextual information thereby to predict topic labels. Different machine learning models for classifying individual topic labels, identifying relevant topic labels, and/or detecting changes in existing topic predictions are developed. Some machine learning models predict topics for full content and some predict topics for partial content. The online system trains the machine learning models to ensure accurate topic predictions are provided timely. The online system employs various machine learning model training methods such as active training and gradient training.

Automatically tagging topics in posts during composition thereof

An online system, such as a social networking system, monitors user interactions with news feed stories of the social networking system and divides the user interactions into non-content clicks and content clicks. The non-content clicks indicate a user's interest in news feed stories based on user actions such as comments on, likes, shares, and hides the news feed stories. The content clicks indicate a user's interest in news feed stories based on user actions on different specific portions of multimedia content (e.g., videos) in the news feed stories such as playing, fast forwarding. The social networking system trains a model based on the monitored user interactions with news feed stories and uses the trained model to rank news feed stories for presentation to a user. The ranks of news feed stories for a user are determined based on a likelihood that the user would find the story interesting.

Ranking of news feeds of content including consideration of specific content clicks by users

An online system ranks topic-groups for users and presents content items in topic-based feeds. A topic group corresponds to one or more topic(s) and can be used to generate a feed for presenting the content items related to the topic(s). For a particular user, the topic groups are ranked according to the likelihood of the user interacting with content items included in the topic groups. The topic groups are ranked using information of the users and/or users' historical interaction data such as click-based interaction data, post-based interaction data, or engagement-based interaction data. The online system generates and provides a user interface for presenting the topic groups to the client device. Content items that are related to the topic(s) corresponding to the topic group are presented in each topic-based feed such that the user can switch between different topic-based feeds.

Topic ranking of content items for topic-based content feeds

An online system generates a feed of content for a user that includes content items provided by, or otherwise related to, other users who are connected to the user via the online system. The online system supplements the feed with additional content items that are not related to users who are connected to the user but are likely to be of interest to the user. The additional content items may be associated with users who are connected to additional users who are connected to the user, content items having received a threshold amount of interacting by other users, content items provided by users who provided other content with which the user interacted, or have other characteristics. The additional content items and content items associated with users connected to the user are included in one or more selection processes that generate the feed for the user.

Generating a feed of content for a user of an online system including content associated with additional online system users who are not connected to the user via the online system

A social networking system receives messages from users that include links to webpages that designate keywords of the webpage. The social networking system identifies webpages linked by users to generate computer models that predict whether a webpage or message should be associated with particular keywords. The social networking system generates computer models that are trained on example webpages and related keywords linked by users in messages. Prior to generating computer models, the social networking system applies one or more filters to exclude webpages and keywords from consideration. The filters may exclude webpages that have low-reliability, are associated with an excessive number of keywords, or keywords that appear on an insufficient number of domains. After training the computer models, messages composed by users may be analyzed and a keyword predicted for the message, which may be suggested to the user to categorize the message.

Filtering automated selection of keywords for computer modeling

Systems, methods, and non-transitory computer readable media configured to detect access by a user to an original content item relating to a story. At least one of a comments based technique, a token based technique, and a tag based technique is performed on content items. Constraints are applied to identify at least one follow up content item from the content items relating to a development of the story.

Systems and methods for recommending follow up content

A social networking system classifies content items according to their qualities for ranking and selection of content items to present to users within, for example, a newsfeed. Low-quality content items that are unlikely to be interesting or relevant to a user may be distinguished though they may appear to be popular among users in the social networking system. The social networking system identifies within the content items one or more features that are indicators of the quality of the content items. The social networking system can use one or more classifiers to evaluate the content items based on the features, and it can compute a quality metric indicating the quality of a content item based on the result obtained from the classifiers. The quality metric can be used in the ranking and selection of a set of content items to provide to the user.

Content quality evaluation and classification

Systems, methods, and non-transitory computer readable media configured to determine a value of a utility factor associated with a content item corresponding to a link. An optimized utility value relating to an interaction type of an outbound click is determined based on the value of the utility factor. An expected utility score associated with the content item is generated based on the optimized utility value to determine potential presentation of the content item to a user.

Systems and methods for throttling click bait

An online system receives a posted content item from a posting user. The online system labels the posted content item with an audience, the audience being a subset of a group of users having an affinity to a topic of the online system, the subset of the group of users sharing a particular treatment regarding the topic. After identifying an opportunity to present content to a viewing user, the online system selects candidate content items, and scores each candidate content item by determining whether the candidate content item is associated with an audience that includes the viewing user, and if so, modifying the score of the candidate content item to be higher. The online system ranks the candidate content items based on the associated score, selects a subset of the candidate content items based on the associated ranking, and presents the selected subset to the viewing user.

Determining an audience of users to assign to a posted content item in an online system

A social networking system generates stories based on actions of users in the system and provides a newsfeed to users that contain stories that related to one or more of their friends in the system. Although the story ranking algorithm includes a time decay to penalize older stories, stories may actually become stale at different rates. To measure the staleness of a story, the system computes a ratio of a current engagement rate for the story to an average engagement rate for the story. Based on this ratio, the system may filter out stale stories, includes the ratio as a feature in the scoring model, and/or adjust the decay rate.

Determining temporal relevance of newsfeed stories

A social networking system receives messages from users that include hashtags. The social networking system may use a natural language model to identify terms in the hashtag corresponding to words or phrases of the hashtag. The words or phrases may be used to modify a string of the hashtag. The social networking system may also generate computer models to determine likely membership of a message with various hashtags. Prior to generating the computer models, the social networking system may filter certain hashtags from eligibility for computer modeling, particularly hashtags that are not frequently used or that more typically appear as normal text in a message instead of as a hashtag. The social networking system may also calibrate the computer model outputs by comparing a test message output with outputs of a calibration group that includes positive and negative examples with respect to the computer model output.

Filtering automated selection of hashtags for computer modeling

Machine learning models are used for ranking news feed stories presented to users of a social networking system. The social networking system divides its users into different sets, for example, based on demographic characteristics of the users and generates one model for each set of users. The models are periodically retrained. The news feed ranking model may rank news feeds for a user based on information describing other users connected to the user in the social networking system. Information describing other users connected to the user includes interactions of the other users with objects associated with news feed stories. These interactions include commenting on a news feed story, liking a news feed story, or retrieving information, for example, images, videos associated with a news feed story.

Growing community of inventors

Mountain View, CA, United States of America

David Vickrey