DAGNN: Demand-aware Graph Neural Networks for Session-based Recommendation

Xiaofeng ZhangarXiv 2021 · arXiv

TL;DR

This paper introduces DAGNN, a novel approach to session-based recommendation that explicitly models user "demand" to improve predictions. By understanding the underlying motivations behind a user's sequence of actions, DAGNN achieves state-of-the-art performance on real-world datasets, outperforming existing methods.

What Problem Does This Paper Address?

Session-based recommendation systems aim to predict a user's next action (e.g., what item to click or buy) based on their recent browsing history, without relying on explicit user profiles. This is particularly challenging because sessions are often short and users' intentions can be implicit. Existing methods often focus on capturing item sequences or user preferences, but they tend to overlook the crucial aspect of why a user is interacting with these items – their underlying "demand." This demand, which can be thought of as a semantic-level factor driving the session, is what DAGNN seeks to model. For example, a user might be looking for "decoration" items, which could encompass a sofa, table, and curtains, rather than just a single item.

Methods

DAGNN introduces a three-component framework:

  1. Demand Modeling Component: This is the core innovation. It aims to extract the underlying demands from a session. It projects item categories into multiple "demand spaces" and aggregates these representations to form a session-specific demand vector. This component also calculates "demand scores" for both the session and target items, indicating their relevance to each demand.
  2. Demand-aware Item Embedding Component: This component builds upon the demand modeling. It constructs "demand graphs" where edges are weighted by the demand scores. This allows for message passing between items that are related not only by sequence but also by shared demand. A graph neural network (GNN) then learns item embeddings that are aware of these demand contexts. A mutual information loss is also incorporated to align local item embeddings with global session embeddings and item categories with extracted demands.
  3. Demand-driven Recommendation Component: This component takes the learned demand-aware item and session embeddings and uses them to predict the probability of recommending a target item. A prediction layer combines session and item embeddings, and a softmax function outputs the final recommendation probabilities.

Datasets

The paper evaluates DAGNN on two widely adopted real-world datasets:

  • Tmall Dataset: Collected from user purchase records on the Tmall platform.
  • Tafeng Dataset: Collected from transaction data at a Chinese grocery store.

The authors pre-process these datasets by defining sessions based on daily purchases and filtering for session length and item frequency.

Key Insights and Ideas

  • Demand as a Driving Force: The paper strongly posits that user demand is a fundamental driver of session behavior, and explicitly modeling it leads to better recommendations.
  • Multi-faceted Demands: A single session can be influenced by multiple underlying demands, and these demands can shift in importance throughout the session.
  • Demand-weighted Graph Construction: Using demand scores to weight edges in the graph neural network allows for a more nuanced understanding of item relationships, prioritizing connections that are relevant to the user's current intent.
  • Mutual Information for Alignment: The mutual information loss is crucial for ensuring that the learned embeddings effectively capture the relationships between local item information, global session context, and the underlying demands.

Key Results

DAGNN consistently outperforms both baseline and state-of-the-art session-based recommendation methods on both the Tafeng and Tmall datasets. For instance, on the Tafeng dataset, DAGNN achieves an R@20 of 11.6363 and an NDCG@40 of 6.7319, which are improvements over the best performing baseline (SR-GNN*) with R@20 of 11.2129 and NDCG@40 of 6.0464. The paper also demonstrates that incorporating item category information generally improves performance for other models, but DAGNN's demand-aware approach provides a more robust advantage. Ablation studies confirm the effectiveness of each component of DAGNN, with the demand modeling and demand-aware GNN components showing significant impact.

Limitations and Future Work

  • Demand Number Tuning: The optimal number of demands to model can vary between datasets, requiring hyperparameter tuning. The paper notes that the Tmall dataset is less "demand-driven," suggesting that the impact of demand modeling might be dataset-dependent.
  • Computational Cost: While not explicitly detailed as a limitation, the introduction of a demand modeling component and demand graphs likely increases computational complexity compared to simpler models.
  • Future Work: The authors suggest further exploring how to automatically determine the optimal number of demands and investigating more sophisticated demand modeling techniques. They also hint at the potential for DAGNN to be extended to other recommendation scenarios beyond session-based tasks.

Summary updated

BibTeX
@misc{zhang2021dagnn,
  author = {Xiaofeng Zhang},
  title = {{DAGNN: Demand-aware Graph Neural Networks for Session-based Recommendation}},
  year = {2021},
  eprint = {2105.14428},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2105.14428},
}
Figure