Gradient-Based Multiple Robust Learning Calibration on Data Missing-Not-at-Random via Bi-Level Optimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Recommendation systems (RS) are crucial for digital platforms, but suffer from missing not at random (MNAR) rating data due to user self-selection.
• This data bias leads to inaccurate rating predictions in RS.
• Existing methods like Doubly Robust (DR) and Multiple Robust (MR) learning offer debiasing but have limitations in calibration.

Purpose of the Study:

• To propose a novel gradient-based calibrated multiple robust learning method for enhancing recommendation system performance.
• To address the miscalibration of imputed errors and propensity scores in existing Multiple Robust (MR) methods.
• To improve the accuracy and reliability of rating prediction in the presence of MNAR data.

Main Methods:

• Developed a gradient-based calibrated multiple robust learning framework.
• Employed bi-level optimization to determine weights and coefficients for propensity and imputation models within the MR framework.
• Integrated differentiable expected calibration error into the objective function for direct optimization of model calibration quality.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art baselines.
• Experiments on three real-world datasets validate the effectiveness of the new approach.
• The method successfully enhances debiasing performance and reliability in rating prediction.

Conclusions:

• The gradient-based calibrated multiple robust learning method effectively tackles MNAR data challenges in recommendation systems.
• The proposed approach offers improved accuracy and reliability in rating prediction.
• This work advances debiasing techniques for recommendation systems by focusing on model calibration.