CatDive: A simple yet effective method for maximizing category diversity in sequential recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sequential recommendation systems often lack category diversity, leading to repetitive suggestions and reduced user satisfaction.
• Existing methods for category diversification in sequential recommendation typically sacrifice accuracy for diversity.
• Maximizing category diversity is crucial for novelty and serendipity in user recommendations.

Purpose of the Study:

• To propose CatDive, a method for maximizing category diversity in sequential recommendation without sacrificing accuracy.
• To improve user satisfaction by offering a wider range of item categories.
• To establish a new state-of-the-art in category-diversified sequential recommendation.

Main Methods:

• CatDive integrates category information using multi-embedding (item and category embeddings) during training.
• Negative sampling prioritizes users' preferred categories while maintaining high-confidence negatives for accuracy.
• A reranking phase enhances and controls category diversity by prioritizing underrepresented categories via a tunable hyper-parameter.

Main Results:

• CatDive achieves state-of-the-art performance on real-world datasets.
• Demonstrates up to 152.49% higher category diversity at a similar accuracy level compared to competitors.
• Achieves up to 39.73% higher accuracy at a similar category diversity level.

Conclusions:

• CatDive effectively maximizes category diversity in sequential recommendation without accuracy loss.
• The method offers a tunable approach to balance accuracy and diversity.
• CatDive presents a significant advancement in personalized and diverse recommendation systems.