TL;DR
This paper explores how Bayesian clustering methods can group users with similar search behaviors in e-commerce. Surprisingly, a simpler parametric model (LDA) outperformed more complex non-parametric ones in recommending relevant search queries, suggesting that sometimes simpler is better for understanding user cohorts.
What Problem Does This Paper Address?
In the booming world of e-commerce, helping customers find what they're looking for is paramount. Product search is a critical component of this, and personalizing search results is key to customer satisfaction. However, understanding individual user search intent is challenging due to the vast diversity in preferences and the sparsity of user interaction data (many users perform few searches). This paper tackles the problem of user cohorting – grouping users with similar search patterns – to better understand and model their collective search interests, thereby addressing the sparsity issue and potentially improving personalization.
Methods
The core of this paper lies in applying and comparing different Bayesian clustering techniques to model user cohorts based on their search query patterns.
- Topic Modeling: Users' search queries are first represented as "documents" of words. Topic models like Latent Dirichlet Allocation (LDA) are used to extract underlying "topics" from these query documents, creating a topical representation for each user.
- Bayesian Clustering:
- Parametric: Latent Dirichlet Allocation (LDA) is used not only for topic modeling but also as a basis for clustering.
- Non-parametric: The paper investigates more sophisticated non-parametric methods:
- Hierarchical Dirichlet Process (HDP) * Dirichlet Process Gaussian Mixture Model (DPGMM) * Distance Dependent Chinese Restaurant Process (ddCRP): This is a novel extension of the Chinese Restaurant Process (CRP) that incorporates a distance function (specifically, the Hellinger distance between LDA topic distributions) to define the probability of users joining existing clusters.
- Evaluation Task: The quality of the learned user cohorts is evaluated using a collaborative query recommendation task. The idea is that if users are grouped into meaningful cohorts, recommending queries popular within a user's cohort should be effective.
Datasets
The experiments were conducted on the Amazon product search dataset described by Qingyao et al. (Ai et al., 2017). This dataset was generated from Amazon product reviews, where user queries were extracted, and the purchased product served as the relevant item. The authors focused on two product categories: Compact Discs (CDs) and Vinyls, and Kindle. They filtered users to include only those with at least 200 queries, resulting in approximately 1300 users for their experiments.
Key Insights and Ideas
- The Sparsity Challenge: The paper highlights that user search data is inherently sparse, making it difficult to model individual user interests effectively. Cohorting users is proposed as a way to overcome this by leveraging shared information within groups.
- LDA's Surprising Efficacy: A key takeaway is that the relatively simple parametric LDA model, when used for clustering, outperformed more complex non-parametric methods like HDP, ddCRP, and LDA+IGMM in the query recommendation task. This suggests that for this specific problem, the added complexity of non-parametric models might not translate to better performance.
- Topic Space as a Basis for Cohorting: Representing users in a topic space derived from their search queries provides a viable and effective way to cluster them.
Key Results
The paper presents results from the collaborative query recommendation task on both the CDs/Vinyls and Kindle datasets. The primary metric reported is Precision@5 and Precision@10.
- CDs and Vinyls Dataset:
- LDA achieved Precision@5 of 0.072863 and Precision@10 of 0.088133.
- HDP achieved Precision@5 of 0.069046 and Precision@10 of 0.079751.
- ddCRP achieved Precision@5 of 0.054440 and Precision@10 of 0.063734.
- LDA + IGMM and LDA + GMM performed significantly worse.
- Kindle Dataset:
- LDA achieved Precision@5 of 0.071934 and Precision@10 of 0.088187.
- HDP achieved Precision@5 of 0.044093 and Precision@10 of 0.058691.
- ddCRP achieved Precision@5 of 0.040030 and Precision@10 of 0.056057.
- LDA + IGMM and LDA + GMM performed significantly worse.
In both datasets, LDA consistently outperformed the other methods, including the more sophisticated non-parametric ones, in terms of precision.
Limitations and Future Work
- LDA's Dominance: The paper acknowledges that while LDA performed best, the performance gap between LDA and the non-parametric methods might be due to the specific implementation or hyperparameter tuning of the latter.
- Future Directions: The authors suggest that the performance of ddCRP could potentially be improved by clustering in the space of neural embeddings rather than traditional topic models. This hints at exploring more modern representation learning techniques for user search behavior. They also hope their work sparks further discussion on applying Bayesian non-parametric models for user cohorting.
Summary updated