MIRN: A multi-interest retrieval network with sequence-to-interest EM routing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Retrieval

Background:

• Vector-based retrieval is common for recommendations but struggles with representing diverse user interests using a single vector.
• Existing methods often neglect model scale and speed, leading to high computational costs and inefficient item retrieval due to high-dimensional vectors.

Purpose of the Study:

• To propose a novel lightweight multi-interest retrieval network (MIRN) for efficient and accurate processing of users' multiple interests.
• To address the limitations of single-vector representations and improve the accuracy and diversity of item retrieval.

Main Methods:

• Incorporated sequence-to-interest Expectation Maximization (EM) routing to handle multiple user interests.
• Utilized Capsule networks for multi-interest representation learning, clustering multiple Capsule vectors from user behavior sequences.
• Introduced a composite capsule clustering strategy to reduce model scale and a Capsule-aware module with attention for adaptive learning of user representations.

Main Results:

• The proposed MIRN significantly outperforms state-of-the-art approaches in item retrieval.
• Demonstrated substantial improvements in metric evaluations compared to existing methods.
• Achieved higher accuracy and diversity in item retrieval by effectively representing multiple user interests.

Conclusions:

• MIRN offers an effective and lightweight solution for multi-interest retrieval in recommendation systems.
• The novel approach enhances both the accuracy and efficiency of item retrieval.
• The use of Capsule networks and EM routing provides a robust framework for capturing complex user preferences.