[Paper Review] NISER: Normalized Item and Session Representations with Graph Neural Networks

Priyanka Gupta, D. Garg+3 more|arXiv (Cornell University)|Sep 10, 2019

Topic Modeling10 references26 citations

TL;DR

This paper proposes NISER, a graph neural network-based session-based recommendation model that mitigates popularity bias by normalizing item and session graph representations. By constraining the magnitude of learned embeddings, NISER significantly improves performance on long-tail and newly introduced items, outperforming state-of-the-art methods on three benchmark datasets.

ABSTRACT

The goal of session-based recommendation (SR) models is to utilize the information from past actions (e.g. item/product clicks) in a session to recommend items that a user is likely to click next. Recently it has been shown that the sequence of item interactions in a session can be modeled as graph-structured data to better account for complex item transitions. Graph neural networks (GNNs) can learn useful representations for such session-graphs, and have been shown to improve over sequential models such as recurrent neural networks [14]. However, we note that these GNN-based recommendation models suffer from popularity bias: the models are biased towards recommending popular items, and fail to recommend relevant long-tail items (less popular or less frequent items). Therefore, these models perform poorly for the less popular new items arriving daily in a practical online setting. We demonstrate that this issue is, in part, related to the magnitude or norm of the learned item and session-graph representations (embedding vectors). We propose a training procedure that mitigates this issue by using normalized representations. The models using normalized item and session-graph representations perform significantly better: i. for the less popular long-tail items in the offline setting, and ii. for the less popular newly introduced items in the online setting. Furthermore, our approach significantly improves upon existing state-of-the-art on three benchmark datasets.

Motivation & Objective

• To address the popularity bias in GNN-based session-based recommendation models, which favor popular items over less frequent, long-tail items.
• To investigate how the magnitude (norm) of learned item and session-graph representations contributes to this bias.
• To develop a training procedure that normalizes these representations to improve fairness and effectiveness for less popular and newly introduced items.
• To demonstrate improved performance on both long-tail items in offline evaluation and newly introduced items in online settings.

Proposed method

• Introduces a normalization technique that constrains the L2 norm of item and session-graph embedding vectors during training.
• Applies normalization to both item-level and session-level representations learned via graph neural networks.
• Modifies the training objective to encourage balanced representation magnitudes, reducing dominance by high-frequency items.
• Uses graph convolutional networks to encode session interactions as graph-structured data, with normalized embeddings for downstream prediction.
• Employs a contrastive-like regularization in the embedding space to stabilize training and promote diversity in representation norms.
• Maintains compatibility with existing GNN architectures while introducing minimal computational overhead.

Experimental results

Research questions

• RQ1 To what extent does the norm of learned item and session representations contribute to popularity bias in GNN-based session recommendation?
• RQ2 Can normalizing the magnitude of these representations improve recommendation performance for long-tail and newly introduced items?
• RQ3 How does the proposed normalization technique compare to state-of-the-art models in terms of overall recommendation accuracy and fairness?
• RQ4 Does the normalization strategy maintain or improve performance on popular items while significantly boosting performance on less popular items?

Key findings

• NISER significantly improves recommendation performance for long-tail items in offline evaluations, demonstrating reduced popularity bias.
• The model achieves superior performance on newly introduced items in online settings, indicating better generalization to cold-start scenarios.
• NISER outperforms existing state-of-the-art models on three benchmark datasets, confirming its effectiveness across diverse settings.
• Normalization of item and session-graph representations leads to more balanced and diverse recommendations without sacrificing performance on popular items.