Querying features based on user actions in online systems

摘要

Online systems, for example, social networking systems store features describing relations between entities represented in the online system. The information describing the features is represented as a graph. The online system maintains a cumulative feature graph and an incremental feature graph. Feature values based on recent user actions are stored in the incremental graph and feature values based on previous actions are stored in the cumulative graph. Periodically, the information stored in the incremental feature graph is merged with the information stored in the cumulative feature graph. The incremental graph is marked as inactive during the merge and information based on new user actions is stored in an active incremental feature graph. If a request for feature information is received, the feature information obtained from the cumulative feature graph, inactive incremental feature graph and the active incremental feature graph are combined to determine the feature information.

主权项

1. A computer-implemented method comprising:
maintaining, by an online system, a cumulative feature store storing feature values determined from user actions performed before a time point; maintaining, by the online system, an incremental feature store storing feature values determined from user actions performed after the time point, the maintaining comprising updating feature values of the incremental feature store responsive to receiving information describing user actions; receiving a request for a feature value, the request identifying a user and a feature; receiving a first partial result from the cumulative feature store, the first partial result determined from user actions of the type performed by the user before the time point; receiving a second partial result from the incremental feature store, the second partial result determined from user actions of the type performed by the user after the time point; determining a weighted combination comprising the first partial result and the second partial result, wherein the first partial result is weighted by a decay factor; and returning the weighted combination as the requested feature value.