Smarter User Recommendations with ATRank
Alibaba’s New Model for Effective User Behavior Prediction
This article is part of the Academic Alibaba series and is taken from the paper entitled “ATRank: An Attention-Based User Behavior Modeling Framework for Recommendation” by Chang Zhou, Jinze Bai, Junshuai Song, Xiaofei Liu, Zhengchao Zhao, Xiusi Chen, and Jun Gao, accepted by the 2018 Conference of the Association for the Advancement of Artificial Intelligence. The full paper can be read here.
User behavior modeling is a tough science. There are many possible actions a user can take, and a large number of contexts in which those actions can take place.
With user behavior being so difficult to predict, the challenge for retailers is making smart recommendations based on accurate predictions. How can corporations accurately analyze and interpret user behavior, use that information to make smart predictions, and turn those predictions into effective recommendations?
Many models rely on manually extracted, aggregated data. The issue with this approach is that when the extraction is done manually, human oversight may lead to an extraction which fails to fully represent the data. As an alternative, recurrent neural network (RNN)-based methods have recently been trialed to provide overall embeddings of a user behavior sequence. However, these methods can only provide limited information, or aggregated memories of user behavior.
For downstream applications working with incomplete data, it’s difficult to preserve data integrity and avoid introducing noise from unrelated user behaviors.
To resolve these issues, engineers at Alibaba have put forward the ATRank model. This model considers the range of user behaviors and the correlations between those behaviors.
ATRank is an attention-based user behavior modeling framework. User behaviors pass through various elements within the modeling framework shown above, each element performing a specific function:
· Raw feature spaces: Classify behaviors into behavior groups
· Behavior embedding spaces: Embed raw features of user behaviors
· Latent semantic spaces: Create connections and comparisons between behaviors
· Self-attention layer: Captures relationships between behaviors in each semantic space
· Vanilla attention: Produces the final context vector with respect to the embedding vector which is to be predicted
The ATRank model has proven highly effective. In tests against similar systems using an Amazon dataset, ATRank outperformed all of its competitors, as shown in the following table:
A score closer to 1 indicates greater prediction accuracy. More test results are available in the full paper.
The full paper can be read here.
Alibaba Tech
First hand, detailed, and in-depth information about Alibaba’s latest technology → Search “Alibaba Tech” on Facebook
