Introduction
In e-commerce, watchlists are essential tools for users to track items of interest. However, for active users, these lists can become cluttered, making it difficult to monitor price fluctuations, stock availability, or changing personal preferences. This paper introduces Watchlist Recommendation (WLR), a specialized task focused on prioritizing items within a user's existing watchlist.
Unlike standard Sequential Recommendation (SR) which predicts the next item from a static global catalog, WLR faces unique challenges:
- Dynamic Recall Set: The candidate items are restricted to the user's current watchlist, which is constantly evolving.
- Attribute Sensitivity: Users often track items specifically for dynamic attributes like price.
- Data Sparsity: Relying solely on item IDs often leads to overfitting due to the long-tail distribution of items.
Methodology: The Trans2D Model
To address these challenges, the authors propose Trans2D, a Transformer-based architecture that treats user history as a 2D array (sequence $\times$ attributes) rather than a 1D sequence of IDs.
Attention2D
The core innovation is the Attention2D layer, which computes three distinct types of attention scores:
- Point-to-Point ($A^F$): Captures interactions between specific attributes of different items.
- Item-to-Item ($A^I$): Models the overall relationship between items in the sequence.
- Channel-to-Channel ($A^C$): Learns general dependencies between attribute types (e.g., how price trends affect category interest).
Linear2D and Hybrid History
The model utilizes Linear2D layers to maintain attribute-specific semantics throughout the network. Furthermore, it integrates a Hybrid History approach, combining explicit "watchlist clicks" with implicit "recently viewed items" (RVIs) to better capture shifting user intent.
Experimental Setup
The model was evaluated using a large-scale dataset from eBay, comprising two weeks of data from February 2021. The dataset included over 40,000 active users and 5.3 million items, each characterized by 16 attributes such as price, seller-ID, and hierarchical category levels. Trans2D was compared against several state-of-the-art baselines, including GRU4Rec, SASRec, BST, and FDSA.
Key Insights and Results
Trans2D significantly outperformed all baselines. For instance, it achieved a Precision@1 of 43.51%, a substantial improvement over the best baseline (GRU with concatenated attributes) at 39.08%.
Analysis of Model Performance
- Attributes vs. IDs: In sparse environments, attribute-based representations generalize better than item IDs, particularly for "cold-start" items.
- The Value of Late Fusion: By avoiding early pooling or concatenation of attributes, Trans2D allows the attention mechanism to discover fine-grained patterns that are otherwise lost.
- Recency Bias: Sensitivity analysis reveals that the most critical signals come from very recent interactions. Performance gains plateau quickly as sequence length ($N$) and the number of days considered increase.
- Positional Importance: The "Relative Snapshot Position"—how recently an item was added to the watchlist—proved to be a highly predictive feature.
Conclusion
Trans2D demonstrates the power of 2D sequential modeling for recommendation tasks where item attributes are as important as the items themselves. While currently focused on clicks and views in e-commerce, the Attention2D mechanism holds potential for other domains with multi-attribute sequential data. Future work may involve incorporating richer feedback types, such as search queries and purchase history, to further refine intent prediction.
Summary updated