Learning From the Future: Light Cone Modeling for Sequential Recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Recommender Systems

Background:

• Sequential recommendation models typically use only past user behavior, limiting their ability to predict future interests.
• Directly using future behavior for training causes data leakage, compromising model integrity.
• Understanding item transition patterns is crucial for accurate sequential recommendations.

Purpose of the Study:

• To propose a novel method for sequential recommendation that leverages future user behavior without data leakage.
• To introduce a new model that captures both past and future item transition dynamics.
• To enhance the modeling of item transition patterns using geometric structures.

Main Methods:

• Sequential graphs were developed to represent item transition relationships, termed 'light cones'.
• A bidirectional sequential graph convolutional network (BiSGCN) was proposed to encode past and future light cones for item representation learning.
• Manifold translating embedding (MTE) was introduced to model item transition patterns within Riemannian manifolds.

Main Results:

• The proposed BiSGCN model demonstrated superior performance compared to existing methods in sequential recommendation tasks.
• Incorporating future information, learned through collaborative behaviors, significantly improved recommendation accuracy.
• Learning item transitions in Riemannian manifolds using MTE further enhanced the model's ability to capture complex geometric structures.

Conclusions:

• Leveraging future user behavior, indirectly learned, is a critical advancement for sequential recommendation systems.
• The BiSGCN model effectively integrates past and future information for robust item representation.
• Riemannian manifold modeling offers a promising direction for capturing intricate item transition dynamics in sequential data.