[Paper Review] Improving the Diversity of Top-N Recommendation via Determinantal Point Process.

Laming Chen, Guoxin Zhang+1 more|arXiv (Cornell University)|Sep 15, 2017

Stochastic Gradient Optimization Techniques25 references16 citations

TL;DR

This paper proposes a DPP-based recommendation method that enhances top-N diversity by modeling item repulsion, enabling tunable diversity control and efficient inference via an acceleration algorithm. Evaluated on public datasets, it achieves a superior accuracy-diversity trade-off compared to state-of-the-art baselines.

ABSTRACT

Recommender systems take the key responsibility to help users discover items that they might be interested in. Many recommendation algorithms are built upon similarity measures, which usually result in low intra-list diversity. The deficiency in capturing the whole range of user interest often leads to poor satisfaction. To solve this problem, increasing attention has been paid on improving the diversity of recommendation results in recent years. In this paper, we propose a novel method to improve the diversity of top-$N$ recommendation results based on the determinantal point process (DPP), which is an elegant model for characterizing the repulsion phenomenon. We propose an acceleration algorithm to greatly speed up the process of the result inference, making our algorithm practical for large-scale scenarios. We also incorporate a tunable parameter into the DPP model which allows the users to smoothly control the level of diversity. More diversity metrics are introduced to better evaluate diversification algorithms. We have evaluated our algorithm on several public datasets, and compared it thoroughly with other reference algorithms. Results show that our proposed algorithm provides a much better accuracy-diversity trade-off with comparable efficiency.

Motivation & Objective

• Address the low intra-list diversity common in similarity-based recommendation systems.
• Improve user satisfaction by capturing a broader range of user interests in top-N recommendations.
• Develop a practical, scalable method for diversity-aware recommendation in large-scale settings.
• Introduce a tunable parameter to allow users to control the desired level of diversity in recommendations.
• Provide enhanced evaluation metrics to better assess diversification performance in recommendation systems.

Proposed method

• Employ the Determinantal Point Process (DPP) to model negative correlation (repulsion) between recommended items, promoting diversity.
• Design an acceleration algorithm to significantly speed up DPP inference, enabling scalability to large-scale recommendation scenarios.
• Incorporate a tunable hyperparameter in the DPP kernel to control the trade-off between recommendation accuracy and diversity.
• Use a DPP-based sampling mechanism that selects diverse item sets while maintaining relevance to user preferences.
• Integrate the DPP model into a recommendation pipeline that leverages user-item interaction data to compute item similarity and DPP weights.
• Apply the DPP model to generate top-N recommendations by optimizing for both relevance and diversity through determinantal point selection.

Experimental results

Research questions

• RQ1 Can DPP-based modeling effectively improve the diversity of top-N recommendation lists while preserving recommendation accuracy?
• RQ2 How does the proposed acceleration algorithm impact the inference speed of DPP-based recommendations in large-scale settings?
• RQ3 To what extent can the tunable parameter in the DPP model allow users to control the level of diversity in recommendations?
• RQ4 How does the proposed method compare to existing diversity-aware recommendation algorithms in terms of accuracy and diversity trade-offs?
• RQ5 Do the introduced diversity metrics provide a more reliable evaluation of diversification performance than standard metrics?

Key findings

• The proposed DPP-based method achieves a significantly better accuracy-diversity trade-off compared to baseline algorithms on multiple public datasets.
• The acceleration algorithm reduces inference time substantially, making the DPP-based recommendation practical for large-scale applications.
• The tunable parameter enables smooth control over diversity levels, allowing customization based on user preferences.
• The new diversity metrics provide a more nuanced and reliable evaluation of recommendation diversity than conventional metrics.
• Empirical evaluation confirms that the method maintains high recommendation accuracy while substantially increasing list diversity.
• The method outperforms existing state-of-the-art algorithms in both diversity and accuracy, demonstrating its effectiveness and practicality.