GRU4RecBE: A Hybrid Session-Based Movie Recommendation System (Student Abstract)

Michael Potter, Hamlin Liu, Yash Lala, Christian Loanzon, Yizhou SunAAAI 2022 · DOI

Introduction

Session-based Recommendation Systems (SBRSs) are critical components of modern streaming and e-commerce platforms. Unlike traditional recommendation engines that rely on long-term user history, SBRSs predict a user's next action based solely on their current session. This approach is particularly valuable for anonymous users or when privacy concerns limit historical data tracking. However, the limited data available in a single session often leads to sparsity issues.

Most existing SBRS models, such as the classic GRU4Rec, rely exclusively on item IDs. By treating movies as arbitrary numerical identifiers, these models ignore the rich semantic information available in movie descriptions. For instance, if a user watches several psychological thrillers, an ID-only model must learn the relationships between those specific IDs from scratch. GRU4RecBE (GRU4Rec with BERT Embeddings) addresses this by integrating movie plot summaries into the recommendation process, leveraging the idea that "what" a movie is about is as significant as the item's unique identifier.

Methodology

The GRU4RecBE architecture enhances the standard Gated Recurrent Unit (GRU) framework by incorporating a dual-embedding strategy. The model processes each movie in a session through two parallel paths:

  1. ID Embedding Path: A standard latent vector represents the specific movie ID.
  2. Plot Embedding Path: The movie’s IMDb plot summary is processed by a pre-trained BERT model. The representation from the [CLS] token of the final layer is extracted to capture the semantic essence of the narrative.
The GRU4RecBE architecture, illustrating the integration of BERT-derived plot embeddings with traditional movie ID embeddings.

Because BERT embeddings and ID embeddings exist in different vector spaces, the plot embedding is passed through a Projection Layer (an affine layer with a Leaky ReLU activation) to align its dimensions and scale with the ID embedding. These vectors are then merged via element-wise addition before being fed into the GRU, which tracks the evolution of user interest over time. To prevent the model from over-relying on a single data type, the authors employ an alternating optimization strategy, updating weights for plot and ID embeddings in successive epochs.

Experimental Setup

The model was evaluated using two industry-standard benchmarks: MovieLens 1M and MovieLens 20M. To facilitate the hybrid approach, the authors used the IMDbPY tool to map MovieLens IDs to their corresponding IMDb plot summaries. This created a comprehensive dataset combining interaction history with natural language descriptions.

Results and Analysis

GRU4RecBE demonstrated superior performance across all tested metrics, including HIT@k, NDCG@k, and Mean Reciprocal Rank (MRR). It consistently outperformed established models such as BERT4Rec, SASRec, and SR-GNN.

Performance comparison of GRU4RecBE against state-of-the-art models on MovieLens 1M and 20M datasets.

Key findings from the experimental results include:

  • Superior Accuracy: On the MovieLens 20M dataset, GRU4RecBE achieved a HIT@10 of 0.7908, a significant improvement over BERT4Rec (0.7473).
  • Enhanced Ranking: The model showed a notable boost in MRR (0.5063 on ML-20m), indicating that the correct recommendations appear higher in the generated lists.
  • Scalability: The performance gap between GRU4RecBE and its competitors widened as the dataset size increased from 1M to 20M. This suggests that semantic side-information becomes increasingly valuable as user behavior complexity grows.

Conclusion

The success of GRU4RecBE highlights the power of pre-trained language models in recommendation tasks. By leveraging BERT's inherent "knowledge" of language and themes, the system can link movies with similar tropes (e.g., "heist films") even if their ID-based interaction patterns are sparse.

Future work will focus on improving the fusion mechanism—potentially using attention mechanisms to dynamically weight the importance of plot versus ID—and enhancing model interpretability to provide human-readable explanations for recommendations. Additionally, the authors aim to address the computational overhead associated with real-time BERT processing in high-traffic environments.

Summary updated

BibTeX
@inproceedings{potter2022gru,
  author = {Michael Potter and Hamlin Liu and Yash Lala and Christian Loanzon and Yizhou Sun},
  title = {{GRU4RecBE: A Hybrid Session-Based Movie Recommendation System (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21651},
}
Figure