Accurate and scalable social recommendation using mixed-membership stochastic block models

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Predicting user preferences is crucial with vast information growth.
• Existing collaborative filtering models have limitations in handling large datasets and complex user-item interactions.

Purpose of the Study:

• To develop a novel collaborative filtering model for accurate user preference prediction.
• To create a scalable algorithm for inferring overlapping group memberships and predicted ratings.

Main Methods:

• Proposed a collaborative filtering model where users and items can belong to multiple groups simultaneously.
• Developed an expectation-maximization algorithm for inferring model parameters.
• The algorithm's running time scales linearly with the number of observed ratings.

Main Results:

• The model achieves higher accuracy in predicting user ratings compared to current algorithms on large datasets.
• Demonstrated the ability to handle complex user-item relationships beyond linear dependencies.
• Enabled scalable preference prediction for extensive datasets.

Conclusions:

• The proposed model offers a more accurate and scalable approach to collaborative filtering.
• Overlapping group memberships provide a richer representation of user-item interactions.
• This method significantly improves preference prediction in big data scenarios.