| CPC G06N 3/08 (2013.01) | 20 Claims |
|
1. A system, comprising:
a training system implemented by one or more hardware processors with associated memory, wherein the memory stores program instructions that when executed on the one or more hardware processors cause the training system to:
train a multitask neural network (MNN) to perform a plurality of machine learning tasks to predict a plurality of attributes of entities, wherein the plurality of machine learning tasks is selected for performing a downstream machine learning task, and wherein the MNN comprises:
an encoder layer configured to:
receive a time-ordered sequence of events associated with an entity, wherein the entity includes multiple members associated with a user account;
generate, based at least in part on the time-ordered sequence of events, a member composition of the user account that indicates classes of the multiple members in the user account; and
generate a fixed-size representation of the entity based at least in part on the time-ordered sequence and the member composition of the user account; and
a decoder layer including a plurality of decoders configured to:
receive the fixed-size representation of the entity; and
perform the machine learning tasks based at least in part on the fixed-size representation of the entity to output the plurality of attributes of the entity;
wherein the training trains the encoder layer to encode signals of the plurality of attributes in fixed-size representations of entities for use by the downstream machine learning task;
after the MNN is trained, store the trained encoder as part of an entity representation generator; and
train a second machine learning model to perform the downstream machine learning task, wherein second machine learning model is trained using entity representations generated by the entity representation generator that includes the signals of the plurality of attributes associated with the plurality of machine learning tasks.
|