TL;DR
This paper introduces DBPMaN, a novel deep learning model that predicts user click-through rates (CTR) by analyzing sequences of user behaviors as "behavior paths." By matching these paths to historical ones, DBPMaN captures deeper user intent than traditional methods, leading to significant improvements in CTR and advertising revenue on the Meituan platform.
What Problem Does This Paper Address?
Click-through rate (CTR) prediction is crucial for e-commerce platforms like Meituan Takeout, aiming to show users the most relevant items and thus increase engagement and revenue. While existing models often focus on individual user interactions or short sequences, they tend to overlook the underlying decision-making process. The authors argue that a user's sequence of actions (browsing, clicking, ordering) forms a "behavior path" that carries significant psychological clues about their intent. The challenge is to effectively model these behavior paths, which are often sparse, noisy, and difficult to match accurately with historical patterns to predict future clicks.
Methods
The core of the proposed solution is the Deep Behavior Path Matching Network (DBPMaN). It tackles the challenges of behavior path modeling through three key modules:
- Path Enhancing Module (PEM): This module addresses noise and sparsity. It uses a dynamic activation mechanism to focus on more influential behaviors within a path and employs contrastive learning to augment behavior path representations, making them more robust.
- Path Matching Module (PMM): This module handles the accurate matching of behavior paths. It implements a two-level matching strategy: first, it identifies the top-k most similar historical behavior paths to the current one, and then it matches these paths with candidate items to estimate CTR.
- Path Augmenting Module (PAM): This module further refines behavior path embeddings using contrastive learning, pulling embeddings from similar paths closer together.
The overall network takes various features (item, user, behavior sequence) as input, embeds them, and then passes them through PEM, PMM, and PAM to generate a final CTR prediction.
Datasets
The paper evaluates DBPMaN on two datasets:
- Taobao: A public dataset containing 10 days of user interactions.
- Meituan: An industrial dataset collected from the Meituan Takeout App, featuring 14 days of interactions from 100 million users. This dataset is noted for its scale and relevance to the food delivery domain.
Key Insights and Ideas
- Behavior Paths as Decision Trails: The central idea is that sequences of user actions aren't just random events but form "paths" that reveal underlying decision-making psychology. This is a departure from point-to-point interaction modeling.
- Two-Level Matching for Precision: The PMM's two-level approach (path-to-path similarity first, then path-to-candidate matching) is crucial for filtering relevant historical context and improving prediction accuracy.
- Contrastive Learning for Robustness: Using contrastive learning to augment behavior paths helps overcome sparsity and noise, leading to more reliable representations.
- Domain-Specific Performance: The model shows particularly strong improvements on the Meituan dataset, suggesting that the behavior path concept is highly relevant in domains with repetitive user interactions (like food delivery).
Key Results
- Offline Performance: DBPMaN significantly outperforms state-of-the-art CTR models like DeepFM, xDeepFM, DIN, DIEN, and CAN on both the Taobao and Meituan datasets, achieving higher AUC and RelaImpr scores. For instance, on the Meituan dataset, DBPMaN achieved an AUC of 0.6812, a notable improvement.
- Online A/B Test: Deployed on the Meituan food delivery platform, DBPMaN resulted in a 1.6% improvement in CTR and a 1.8% improvement in advertising revenue compared to the baseline model.
- Ablation Study: The results indicate that the Path Matching Module (PMM) is the most critical component, with its removal leading to the largest performance drop.
Limitations and Future Work
- Path Length Sensitivity: The paper notes that the performance can be sensitive to the chosen behavior path length (
l), with different lengths (e.g.,l=8vs.l=16) showing slightly varied results in online tests. - Computational Cost: While not explicitly detailed as a limitation, the complexity of modeling behavior paths and the two-level matching might introduce higher computational overhead compared to simpler models.
- Future Directions: The authors suggest that DBPMaN opens avenues for exploring user decision-making psychology through behavior path modeling. Further research could delve deeper into understanding these psychological aspects and potentially applying this approach to other recommendation scenarios.
Summary updated