Remembering past and predicting future: a hybrid recurrent neural network based recommender system.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Traditional recommender systems (RS) often assume static user preferences and are reactive to taste changes.
• Existing RS primarily focus on candidate item generation, neglecting the crucial candidate ranking phase.
• The limitations of static and reactive models hinder effective personalized recommendations over time.

Purpose of the Study:

• To develop a novel recommender system (RS) that addresses the limitations of traditional static and reactive models.
• To explore and integrate both candidate generation and candidate ranking phases for improved recommendation quality.
• To propose a multi-objective RS considering short-term prediction, long-term prediction, diversity, and popularity bias.

Main Methods:

• Exploiting both sequential and non-sequential user behavior patterns for future trajectory prediction.
• Utilizing recurrent neural networks (RNNs) and nearest neighbors approaches for short-term and long-term predictions.
• Introducing a novel candidate ranking method to mitigate recommendation entanglement and improve user experience.

Main Results:

• Achieved a hit rate of 58% and short-term prediction success of 71% on the MovieLens (ML) 1M dataset.
• Successfully handled diversity and item popularity, achieving success rates of 59.22% and 34.28% respectively.
• Demonstrated superior performance compared to traditional methods across multiple datasets.

Conclusions:

• The proposed multi-objective recommender system effectively predicts dynamic user behavior and improves recommendation quality.
• Integrating candidate generation and ranking, alongside diverse objectives, leads to more accurate and engaging recommendations.
• This approach offers a significant advancement in personalized recommendation systems by addressing previously overlooked multi-objective considerations.