Abstract
Personalized product search aims to resolve query ambiguity by leveraging user history. However, traditional models often conflate various product attributes into a single feature vector, introducing noise and failing to capture specific attribute-level preferences. This paper introduces the Multi-Attribute Interest (MAI) learning model, which separately profiles attribute-specific interests and dynamically fuses them based on learned contribution weights. Experimental results on large-scale datasets demonstrate that MAI significantly outperforms existing state-of-the-art methods.
Introduction
When a user searches for a term like "MAC," the intent could range from an Apple computer to a specific brand of cosmetics. Personalized search engines attempt to disambiguate these queries by analyzing the user's historical interactions. A significant challenge in this domain is that user preferences are often tied to specific product attributes—such as brand, category, or product name—rather than the product as a whole. Furthermore, these preferences are dynamic; a user might be a brand loyalist for high-end electronics but a price-sensitive shopper for household staples.
Existing models typically treat all product information as a unified "soup" of features. This approach ignores the fact that different attributes drive decisions for different users at different times. By mixing these signals, models introduce noise that can obscure the user's true intent.
Methodology
The proposed Multi-Attribute Interest (MAI) model addresses these limitations by breaking the interest-learning process into three distinct stages:
1. Attribute-centered Interest Profiling
Instead of creating a single general user profile, MAI builds separate profiles for each key attribute (e.g., a "Brand Profile," a "Category Profile," and a "Name Profile"). It utilizes a Transformer encoder to capture the temporal evolution of a user’s interest in each specific attribute based on their short-term history.
2. Attribute-aware Interest Profiling
To maintain a holistic view, this module concatenates all attribute features into a unified vector. This allows the model to capture vital cross-attribute correlations—for instance, recognizing that a preference for a specific brand may be highly dependent on the product category.
3. Dynamic Interest Fusion
The "secret sauce" of the MAI model is the dynamic fusion of these profiles. The model calculates contribution weights for each attribute by analyzing the stability and affinity of historical attribute representations. If a user’s recent history shows a consistent preference for a specific brand, the model increases the weight of the "Brand" attribute in the final ranking score.
Additionally, the paper introduces Attribute-centered Query Representations. Rather than using raw query text, the model reformulates the query to emphasize terms related to specific attributes, ensuring a better match between the user's current search and their historical attribute-level preferences.
Experimental Setup
The MAI model was evaluated against several baselines, including HEM, ZAM, and the previous state-of-the-art, TEM. Testing was conducted on two major benchmarks:
- JDsearch: A massive dataset from JD.com containing real-world search logs and user interactions.
- Amazon Dataset: Specifically the "Electronics" and "CDs & Vinyl" sub-datasets, filtered to ensure sufficient profiling data for each user and item.
Results and Discussion
The MAI model consistently outperformed all baselines across both datasets. On the JDsearch benchmark, MAI achieved a 4.31% improvement in Mean Reciprocal Rank (MRR) and a 1.84% improvement in Normalized Discounted Cumulative Gain (NDCG) over the TEM model.
Ablation Studies
Ablation experiments revealed that the Attribute-centered (AC) module is the most critical innovation. Removing this module led to the most significant performance degradation, confirming that isolating attribute-level interests is key to de-noising user history.
Qualitative Analysis of User Interests
Visualizing the learned contribution weights provides insight into how the model adapts to different users. As shown in the heatmap analysis, the model can distinguish between users who are primarily driven by product names and those who are brand-centric.
The study also found that a "short-term history" window of approximately 15 items provides the optimal balance. Shorter windows lack sufficient context, while longer windows introduce irrelevant historical noise that can dilute the signal of current intent.
Conclusion and Future Work
The MAI model represents a significant step forward in personalized product search by recognizing and modeling the dynamic importance of different product attributes. By isolating these "anchors" of user interest, the model effectively de-noises historical data to provide more accurate search results.
Future research will focus on the "switching problem," where users abruptly change their attribute focus (e.g., a brand-loyalist suddenly prioritizing price for a new category). Additionally, the authors aim to develop more explicitly interpretable weighting strategies to better explain the model's decision-making process in real-time.
Summary updated