A Transformer-based Embedding Model for Personalized Product Search

Keping Bi,Qingyao Ai,W. Bruce Croft

DOI: https://doi.org/10.1145/3397271.3401192

2020-05-19

Abstract:Product search is an important way for people to browse and purchase items on E-commerce platforms. While customers tend to make choices based on their personal tastes and preferences, analysis of commercial product search logs has shown that personalization does not always improve product search quality. Most existing product search techniques, however, conduct undifferentiated personalization across search sessions. They either use a fixed coefficient to control the influence of personalization or let personalization take effect all the time with an attention mechanism. The only notable exception is the recently proposed zero-attention model (ZAM) that can adaptively adjust the effect of personalization by allowing the query to attend to a zero vector. Nonetheless, in ZAM, personalization can act at most as equally important as the query and the representations of items are static across the collection regardless of the items co-occurring in the user's historical purchases. Aware of these limitations, we propose a transformer-based embedding model (TEM) for personalized product search, which could dynamically control the influence of personalization by encoding the sequence of query and user's purchase history with a transformer architecture. Personalization could have a dominant impact when necessary and interactions between items can be taken into consideration when computing attention weights. Experimental results show that TEM outperforms state-of-the-art personalization product retrieval models significantly.

Information Retrieval

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to achieve more effective personalized search results in product searches on e - commerce platforms. Although personalization can improve search quality, it is not applicable in all cases. Existing personalized product search techniques usually adopt a unified personalization mechanism for all search sessions, and this method may damage search quality due to data sparsity and the introduction of unnecessary noise. Moreover, most existing methods are unable to dynamically adjust the degree of personalization according to different contexts.

Specifically, the paper points out that current methods have the following limitations:
1. **Fixed coefficients control personalization**: Some methods use fixed coefficients to control the influence of personalization, which results in the same degree of personalization in all search sessions.
2. **Attention mechanism always in effect**: Other methods make personalization always effective through the attention mechanism, but this method may introduce unnecessary noise when personalization is not required.
3. **Limitations of the Zero - Attention Model (ZAM)**: Although ZAM can adaptively adjust the degree of personalization by allowing the query to focus on the zero vector, its maximum degree of personalization can only make the query and user information equally important, and it cannot make personalization dominant in some cases.

To solve these problems, the paper proposes a Transformer - based Embedding Model (TEM), which can dynamically control the degree of personalization. TEM encodes the query and the user's purchase history sequence, enabling personalization to be dominant when necessary, and takes into account the interactions between items when calculating the attention weights. Experimental results show that TEM significantly outperforms existing state - of - the - art methods in personalized product retrieval.