BLOGS

Decoding ‘Cd-HRNN Content Driven HRNN to Improve Session-Based Recommendation System’ with Brijraj and Sonal

By Brijraj Singh, Research Scientist and Sonal Dabral, Data Mining Intern At Sony Research India
26th July 2023
Brijraj Singh and Sonal Dabral summarize their paper that was recently accepted at the IJCNN conference in Australia, a premier conference in the area of neural networks theory, analysis and applications.

Abstract

The increasing popularity of digital entertainment systems has made personalization a key factor for success in the industry. Recommendation systems, particularly for videos and movies, are crucial in this regard. However, many existing systems are implicit feedback recommendation systems that use indirect signals to infer user preferences, such as user actions (e.g. clicks, views, purchases) or interactions with items (e.g. listening to a song, watching a movie).

The challenge lies in the limited information and uncertainty present in user behaviour, making it difficult to predict user interests and preferences. In previous research, Recurrent Neural Networks (RNNs) have shown to be efficient in predicting the next item in a session, based on past item click sequences, but their effectiveness is limited when only relying on click sequences as input data. In this paper, we extend the Hierarchical RNN (HRNN) architecture for generating recommendations by combining information about session clicks and item content, such as item ids and item description respectively. The Bidirectional Encoder Representations from Transformers (BERT) architecture is applied for generating feature vectors from text descriptions of the items. Our model has been extensively tested on the benchmark dataset MovieLens 1m, and has demonstrated superiority over state-of-the-art (SOTA) session-based recommendation systems (SBRS) models. Our experimental results establish the efficacy of using content information along with item ids for recommendation.

Introduction

While dealing with the user-item matrix in the context of recommendation systems, the use of item ids presents a challenge, since the algorithm expects only the exact same item-id to locate a pattern. This does not make much sense in the real world, since similar content such as sequels of a movie/show can have entirely different item ids and a supervised learning algorithm cannot identify them as a pattern. This prevents the algorithm from creating and capturing relevant patterns and results in sub-par recommendation performance.

To overcome this challenge, we proposed a mechanism of considering item description along with item-ids and applied it to Hierarchical Recurrent Neural Network (HRNN), and observed an improved recommendation performance. HRNN is a derivative of RNN models that works with sequential data and captures multiple folds of user behaviour: (a) user behaviour in a particular session and (b) user’s holistic behaviour on the platform over all the time. HRNN achieves this with two hierarchies of RNN units to capture intra-session behaviour + inter-session behaviour.

Methodology:

Figure 1 Architecture of Cd-HRNN

Definition 1: A session is defined as sequential activities of a user’s interaction with the platform. if be the time stamps of a user’s interaction with the platform, then a session is defined as activities at timestamps such that features extracted at one session is different from any other sessions. where is a features extractor of the items interacted during those time stamps:

Table 1 Comparative Performance of the Proposed Method with Other Models

Results:

In our experiments, we observed that the proposed model performance exceeded that obtained by other methods we compared it with. When the item description was added, the recommendation performance improved significantly as compared to the base HRNN model.
Table I shows the accuracy of various algorithms in a recommendation task as measured by Recall and MRR (Mean Reciprocal Rank) at 20. The results demonstrate the inefficiency of popularity-based algorithms on this particular dataset. Algorithms based on similarity perform well. Item Session KNN outperforms the POP method on both metrics. The RNN-based algorithms outperform non-deep learning methods in terms of recall and MRR, regardless of the cut-off. HRNN exhibits better performance compared to other baselines. The ability of HRNN to model cross-session dynamics greatly contributes to the overall quality of the recommendations. Cd-HRNN surpasses the HRNN baseline in terms of recall and maintains a good MRR score. The proposed content driven HRNN model performs significantly better than the HRNN baseline. This confirms the effectiveness of incorporating content information to enhance the accuracy of recommendations.

Conclusion:

Previous research in the field of SBRS has mainly concentrated on using user click history, with limited attention given to incorporating item information. Our contribution is utilizing movie embedding information via transformer models to improve the performance of these systems.

Our proposed method has been demonstrated to outperform existing state-of-the-art feature-agnostic recommendation systems. This result shows the effectiveness and efficiency of our proposed approach, highlighting its superiority compared to other methods in the field. These findings suggest that our proposed method can be a valuable contribution to the area of feature-agnostic recommendation systems and can help improve the accuracy of these systems.

In our future work, we plan to include more item content or user information in the embedding process to further enhance item representation. Additionally, incorporating attention mechanisms alongside feature information could be a promising direction for improving model performance and interpretability.
Skip to content