Abstract
Traditional recommender systems often optimize for a single target behavior, such as purchases, while ignoring the rich sequence of intermediate actions users take. This paper introduces Neural Multi-Task Recommendation (NMTR), a framework designed to leverage multiple cascading behaviors (e.g., view $\rightarrow$ add-to-cart $\rightarrow$ purchase). By modeling the ordinal relationship between these actions through a multi-task learning lens, NMTR significantly improves recommendation accuracy, particularly for users with sparse purchase histories.
Introduction
In e-commerce environments, users interact with items through a variety of behaviors: clicking, viewing, adding to carts, and finally purchasing. While the "purchase" is the primary KPI, it is often the rarest event, leading to severe data sparsity. Existing multi-behavior models like Collective Matrix Factorization (CMF) often fail because they assign different embeddings to the same user for different behaviors or treat behaviors as independent, parallel actions. NMTR addresses these flaws by enforcing a shared user-item embedding space and explicitly modeling the sequential "cascade" of user intent.
The NMTR Framework
The authors propose a neural collaborative filtering approach that captures the semantics of behavioral sequences. The architecture is built on four key pillars:
- Shared Embedding Layer: NMTR maintains a single, static embedding vector for each user and item across all behaviors. This reflects the intuition that a user's underlying preferences are constant, regardless of the specific interaction type.
- Separated Interaction Functions: Above the embeddings, the model employs Neural Collaborative Filtering (NCF) units to learn distinct interaction functions for each behavior (e.g., GMF or NeuMF).
- Cascaded Predictions: To capture the ordinal relationship (View $\rightarrow$ Cart $\rightarrow$ Purchase), the prediction output of a lower-level behavior is fed into the prediction function of the subsequent higher-level behavior.
- Multi-Task Learning (MTL): The model optimizes a joint loss function across all behaviors simultaneously. This allows the abundant signal from low-level behaviors (like views) to refine the shared embeddings, directly benefiting the prediction of sparse target behaviors.
Experimental Results
The framework was evaluated on two real-world datasets: Beibei (maternal/infant products) and Tmall (IJCAI-15 challenge).
Performance and Convergence
NMTR, specifically using the NeuMF interaction function (NMTR-NeuMF), achieved state-of-the-art results. It outperformed the best multi-behavior baseline (MC-NeuMF) by 6.08% in Hit Ratio (HR@100) on the Beibei dataset and by 30.76% on the Tmall dataset. As shown in the training curves, NMTR-NeuMF demonstrates stable convergence and superior ranking performance compared to standard NeuMF and other multi-behavior variants.
Resilience to Data Sparsity
A key strength of NMTR is its performance for "cold" users. For users with only 5 to 8 purchase records, NMTR-NeuMF maintained a performance lead of over 11% in HR@100 over the best baselines. The gap between NMTR and traditional models actually increases as the target behavior data becomes sparser, proving the value of the cascading auxiliary signals.
Key Insights
- Signal Quality: Interestingly, "view" data was found to be more beneficial for purchase prediction than "add-to-cart" data. While carts are highly correlated with purchases, views provide a broader map of user interests and explicit "non-purchases."
- Joint Training: The authors demonstrated that simultaneous multi-task training is essential. Sequential training (training one behavior and freezing it before the next) fails to allow the models to mutually reinforce the shared embedding space.
Conclusion
NMTR provides a robust solution for multi-behavior recommendation by respecting the natural hierarchy of user actions. While the current model assumes a strict ordinal cascade, future work aims to adapt the framework for partial-order behavior graphs and incorporate temporal dynamics through Recurrent Neural Networks (RNNs).
Summary updated