TL;DR
The FCI-GNN model improves session-based recommendations by explicitly fusing information from potential candidate items with both local session data and global transition patterns. By using a triple-layered attention approach—global, session, and candidate-level—the model better captures the intent of users with diverse interests, outperforming state-of-the-art baselines on major e-commerce and music datasets.
What Problem Does This Paper Address?
Session-based recommendation (SBR) is the "cold start" challenge of the web: how do you recommend the next item to an anonymous user with no login history? Most modern approaches use Graph Neural Networks (GNNs) to model the sequence of clicks as a graph, looking at transitions within the current session and across all historical sessions (global context).
However, the authors identify a key gap: current models often ignore the "candidate items" themselves during the feature extraction phase. If a user has broad interests (e.g., looking at a computer but also interested in furniture), a model that only looks at the current session's items might narrow its focus too much. The authors argue that by "activating" the user's diverse interests through the lens of candidate items, the recommendation becomes more accurate and personalized to the immediate intent.
Methods
The authors propose the FCI-GNN (Furnish Charging Information Graph Neural Network). Its architecture consists of five main components:
- Global-Level Item-Embedding: It builds a global graph from all sessions to break the assumption of sequence independence. It uses a session-aware attention mechanism to pull relevant transition data from other users' journeys.
- Session-Level Item-Embedding: This layer focuses on the current session graph, using attention to learn the complex relationships (in-edges, out-edges, and self-loops) between items the user just clicked.
- Position Embedding: Interestingly, the model uses reverse location coding. Since the most recent items in a session are usually the strongest predictors of the next click, reverse coding ensures the model prioritizes the "end" of the session regardless of its total length.
- Candidate-Level Item-Embedding: This is the core novelty. It uses a target attention mechanism to calculate the relevance of the user's historical actions in the current session relative to each potential candidate item.
- Prediction Layer: Finally, it integrates these three levels of representation to calculate the probability of the next click.
Datasets
The model was evaluated on three standard industry benchmarks:
- Diginetica: Transaction data from the CIKM Cup 2016.
- Tmall: Anonymous purchase data from the Tmall e-commerce platform.
- Nowplaying: A collection of music listening habits from Twitter users.
The authors preprocessed these by removing sessions with only one item and sessions with fewer than five items to ensure enough context for the GNN to learn.
Key Insights and Ideas
- The "Candidate Activation" Hypothesis: The most striking idea is that candidate items shouldn't just be the output of a model; they should be part of the input context. By comparing session items to candidate items via attention, the model "reminds" itself of the user's potential broader interests.
- Reverse Position Matters: In many sequences, the first item is less relevant than the last. By using reverse position embedding (counting backward from the current item), the model more effectively weights the most recent—and usually most relevant—user actions.
- Global vs. Local Balance: The paper reinforces that while global transitions (what everyone else does) are helpful, they can introduce noise. The attention mechanism acts as a filter to ensure only relevant global patterns are fused into the current session's representation.
Key Results
FCI-GNN outperformed 10 baseline models (including popular ones like SR-GNN, GCE-GNN, and DHCN) in 10 out of 12 evaluation metrics.
- Accuracy Gains: On the Tmall dataset, FCI-GNN outperformed the GCE-GNN baseline by 1.86% in MRR@10 (Mean Reciprocal Rank) and 1.49% in MRR@20.
- Ranking Precision: The improvement in MRR (which measures how high the correct item is in the list) was more significant than the improvement in P@K (Precision), suggesting the model is particularly good at ranking the "true" next item near the very top of the list.
- Ablation Success: Removing the candidate-level layer (FCI-GNN-T) led to a noticeable drop in performance, confirming that fusing candidate information is indeed the "secret sauce."
Limitations and Future Work
- Noisy Long Sessions: The authors noted that on the Nowplaying dataset (which has longer average sessions), the model's performance was slightly more varied. They suspect that longer sessions contain more "noise" (random clicks) that can confuse the attention mechanism.
- Computational Complexity: GNN-based models are resource-intensive. The authors acknowledge that these models take a long time to train and consume significant memory, suggesting that future work should focus on reducing model complexity without sacrificing accuracy.
- Hyperparameter Sensitivity: The model's performance is sensitive to the number of neighbors chosen in the global graph (optimally 12) and the embedding dimensions (optimally 100). Finding the "sweet spot" for different types of datasets remains a manual task.
Summary updated