Multidomain time series recommendation
Events are captured from a plurality of source domains by a plurality of entities for a plurality of different physical items. The events are captured as time series data to generate a unified machine learning embedding. The unified machine learning embedding is used for predictions in different domains of a plurality of prediction target domains.
1 . A method, comprising:
capturing events from a plurality of source domains by a plurality of entities for a plurality of different physical items as time series data;
generating a unified machine learning embedding including by:
transforming the captured events into an encoder input space;
processing the transformed events using a temporally-masked encoder that merges the transformed events with a hidden client embedding to determine a weighted masked average; and
passing an output of the temporally-masked encoder through a gated update to generate the unified machine learning embedding; and
using the unified machine learning embedding for predictions in different domains of a plurality of prediction target domains.
2 . The method of claim 1 , wherein capturing the events from the plurality of the source domains includes creating an update event for a specific source domain, wherein the update event is associated with a source domain identifier, a client identifier, a timestamp, and a payload.
3 . The method of claim 2 , wherein the update event corresponds to a profile update, an updated client survey, a viewed item, a purchased item, a returned item, an item added to a virtual shopping bag, or a rating submitted for an item.
4 . The method of claim 2 , wherein the payload of the update event is based on the specific source domain.
5 . The method of claim 2 , wherein the payload includes one or more of the following: an item identifier, a sold status identifier, a price rating, a price change, a fit rating, or a style rating.
6 . The method of claim 2 wherein the payload includes one or more of the following: a birthday, a weight value, a height value, a foot size, a shirt size, a bottom size, a cup size, a color preference, a pattern preference, a fabric preference, one or more preferred brands, one or more brands to avoid, or an occupation.
7 . The method of claim 1 , wherein using the unified machine learning embedding for the predictions in the different domains of the plurality of prediction target domains includes creating a target event, wherein the target event is associated with a target prediction domain identifier, a client identifier, and a payload, and wherein the payload identifies one or more prediction outcomes.
8 . The method of claim 7 , wherein the target event is further associated with a timestamp.
9 . The method of claim 1 , wherein a loss function is configured for each of the different domains of the plurality of prediction target domains.
10 . The method of claim 1 , further comprising sorting the captured events by time.
11 . The method of claim 1 , wherein using the unified machine learning embedding for the predictions in the different domains of the plurality of prediction target domains includes receiving a plurality of items and predicting corresponding client-item scores for each of the items of the received plurality of items.
12 . The method of claim 11 , further comprising displaying information related to one or more of the items of the received plurality of items based on the corresponding predicted client-item scores, wherein the information is provided for one of the different domains of the plurality of prediction target domains.
13 . The method of claim 1 , further comprising receiving a prediction request, wherein the prediction request is associated with one or more of the following: a selection of an item category, a selection to view an arranged outfit, a selection to view a custom item arrangement, a request for a custom item arrangement, or a selection to view a pseudo-random item.
14 . A system, comprising:
one or more processors; and
a memory coupled to the one or more processors, wherein the memory is configured to provide the one or more processors with instructions which when executed cause the one or more processors to:
capture events from a plurality of source domains by a plurality of entities for a plurality of different physical items as time series data;
generate a unified machine learning embedding including by causing the one or more processors to:
transform the captured events into an encoder input space:
process the transformed events using a temporally-masked encoder that merges the transformed events with a hidden client embedding to determine a weighted masked average; and
pass an output of the temporally-masked encoder through a gated update to generate the unified machine learning embedding; and
use the unified machine learning embedding, generate predictions in different domains of a plurality of prediction target domains.
15 . The system of claim 14 , wherein capturing the events from the plurality of the source domains includes creating an update event, wherein the update event is associated with a source domain identifier, a client identifier, a timestamp, and a payload.
16 . The system of claim 15 , wherein the update event corresponds to one or more of the following: a profile update, an updated client survey, a viewed item, a purchased item, a returned item, an item added to a virtual shopping bag, or a rating submitted for an item.
17 . The system of claim 14 , wherein generating predictions in the different domains of the plurality of prediction target domains includes creating a target event, wherein the target event is associated with a target prediction domain identifier, a client identifier, and a payload, and wherein the payload identifies one or more prediction outcomes.
18 . The system of claim 14 , wherein generating predictions in the different domains of the plurality of prediction target domains includes receiving a plurality of items and predicting client-item scores for each of the items of the received plurality of items.
19 . The system of claim 14 , wherein the memory is further configured to provide the one or more processors with instructions which when executed cause the one or more processors to receive a prediction request, wherein the prediction request is associated with one or more of the following: a selection of an item category page, a selection of an arranged outfit page, a selection to view a custom item arrangement, a request for a custom item arrangement, or a selection to view a pseudo-random item.
20 . A computer program product, the computer program product being embodied in a non-transitory computer readable storage medium and comprising computer instructions for:
capturing events from a plurality of source domains by a plurality of entities for a plurality of different physical items as time series data;
generating a unified machine learning embedding including by:
transforming the captured events into an encoder input space;
processing the transformed events using a temporally-masked encoder that merges the transformed events with a hidden client embedding to determine a weighted masked average; and
passing an output of the temporally-masked encoder through a gated update to generate the unified machine learning embedding; and
using the unified machine learning embedding for predictions in different domains of a plurality of prediction target domains.