PERSONALIZATION FROM SEQUENCES AND REPRESENTATIONS IN ADS
Aspects of the disclosure provide a computer-implemented method for generating personalized results. The method includes identifying a set of user actions by a specific user within a sliding window of time, generating a first representations for the set of user actions using an encoder component of a personalization module, generating a second representation for the set of user actions using a pretrained representations component of the personalization module, generating a third representation for the set of user actions using a learned representations component of the personalization module, using the personalization module to combine the first representation, second representation and the third representation to generate a short-term personalized representation for the specific user, and providing a set of results for display to the user based on the short-term personalized representation.
1 . A computer-implemented method comprising:
identifying, by one or more processors of a server computing device, a set of user actions by a specific user within a sliding window of time;
generating, by the one or more processors, a first representation for the set of user actions using an encoder component of a personalization module;
generating, by the one or more processors, a second representation for the set of user actions using a pretrained representations component of the personalization module;
generating, by the one or more processors, a third representation for the set of user actions using a learned representations component of the personalization module;
using, by the one or more processors, the personalization module to combine the first representation, second representation and the third representation to generate a short-term personalized representation for the specific user; and
providing, by the one or more processors, a set of results for display to the specific user based on the short-term personalized representation.
2 . The method of claim 1 , wherein the set of user actions include one or more of search queries, item favorites, listing views, items added to a cart of the specific user, or one or more past purchases.
3 . The method of claim 1 , further comprising:
inputting the short-term personalized representation into one or more personalized downstream models in order to generate a value; and
ranking the set of results based on the value, and wherein the ranked set of results is provided for display to the specific user.
4 . The method of claim 3 , wherein the one or more personalized downstream models includes a first model that generates a predicted probability that a particular listing will be clicked.
5 . The method of claim 4 , wherein the one or more personalized downstream models further include a second model that generates a predicted conditional probability that a good or service represented by a listing will be purchased.
6 . The method of claim 1 , wherein the personalization module is implemented as a Tensorflow Keras layer.
7 . The method of claim 1 , further comprising determining a length of the sliding window based on a location of the specific user.
8 . The method of claim 1 , further comprising determining a length of the sliding window based on a type of listing selected by the specific user within the sliding window.
9 . The method of claim 1 , wherein the sliding window is no more than 1 hour.
10 . The method of claim 1 , wherein the set of user actions is limited in number according to a maximum sequence length.
11 . The method of claim 1 , wherein the encoder component includes a transformer encoder.
12 . The method of claim 11 , wherein the encoder component is implemented as an importable Keras layer which encodes sequences of listings.
13 . The method of claim 1 , wherein the pretrained representations component is configured to encode sequences of user actions within the sliding window.
14 . The method of claim 1 , wherein the pretrained representations component is configured to encode sequences of search queries within the sliding window as text representations.
15 . The method of claim 14 , wherein the text representations are Skip-gram text representations.
16 . The method of claim 1 , wherein the pretrained representations component is configured to encode sequences of listing identifiers within the sliding window as multimodal representations.
17 . The method of claim 1 , wherein the pretrained representations component is configured to encode sequences of listing identifiers within the sliding window as visual representations.
18 . The method of claim 1 , wherein the pretrained representations component is configured to encode sequences of listing identifiers within the sliding window as Skip-gram listing representations.
19 . The method of claim 1 , wherein the learned representations component is configured as a look-up table.
20 . A computer system configured to generate personalized results, the computer system comprising:
memory configured to store a set of user actions for a specific user within a sliding window of time; and
one or more processors operatively coupled to the memory, the one or more processors being configured to:
generate a first representation for the set of user actions using an encoder component of a personalization module;
generate a second representation for the set of user actions using a pretrained representations component of the personalization module;
generate a third representation for the set of user actions using a learned representations component of the personalization module;
use the personalization module to combine the first representation, second representation and the third representation to generate a short-term personalized representation for the specific user; and
provide a set of results for display to the specific user based on the short-term personalized representation.