Modeling Dynamic Missingness of Implicit Feedback for Recommendation.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Implicit feedback is crucial for collaborative filtering but contains missing not at random (MNAR) data, conflating negative and unknown user preferences.
• Existing models address missing data using static 'exposure' variables but neglect temporal dynamics essential for understanding user behavior.
• Temporal dependencies among items significantly influence missing data patterns, impacting the accuracy of recommendation systems.

Purpose of the Study:

• To address the limitations of static models in handling MNAR implicit feedback.
• To introduce a novel approach for modeling the temporal dynamics of missing data in recommendation systems.
• To improve the learning of user preferences by incorporating time-varying missingness patterns.

Main Methods:

• Proposed a latent variable, 'user intent,' to capture temporal changes in item missingness.
• Utilized a hidden Markov model to represent the dynamic process of user intent and item exposure.
• Integrated the dynamic missingness model with matrix factorization (MF) for enhanced recommendations.
• Explored constraints for compact and interpretable user intent representations.

Main Results:

• The proposed framework effectively captures dynamic item missingness.
• Incorporating user intent and temporal dynamics into matrix factorization significantly improves recommendation performance.
• Experiments on real-world datasets show superior results compared to state-of-the-art recommender systems.
• The method provides a more nuanced understanding of user preferences by accounting for temporal feedback patterns.

Conclusions:

• Modeling temporal dynamics of missing data, via 'user intent,' is crucial for accurate recommendations.
• The hidden Markov model approach effectively captures these dynamics, outperforming static methods.
• This work advances collaborative filtering by addressing the complexities of MNAR implicit feedback in a temporal context.