Positive-Unlabeled Learning in Implicit Feedback from Data Missing-Not-At-Random Perspective.

Area of Science:

• Machine Learning
• Recommender Systems
• Data Science

Background:

• The absence of explicit negative labels is a common problem in Computer Vision (CV), Natural Language Processing (NLP), and Recommender Systems (RSs).
• Existing negative sample completion methods in CV, NLP, and RSs often struggle with the Missing-Not-At-Random (MNAR) data nature and unmeasured confounders, impacting model robustness.
• These limitations necessitate advanced techniques for accurate prediction in implicit feedback scenarios.

Purpose of the Study:

• To address the challenge of missing negative labels in recommender systems (RSs) with implicit feedback.
• To propose a novel framework that improves the robustness of RS models against unmeasured confounders.
• To provide a theoretical analysis of existing methods and introduce a more reliable approach.

Main Methods:

• Formulating the RS prediction task with implicit feedback as a positive-unlabeled (PU) learning problem.
• Developing a two-phase debiasing framework involving exposure status imputation and a doubly robust estimator.
• Incorporating a robust deconfounding method to mitigate the effects of unmeasured confounders, addressing limitations of propensity-based approaches.

Main Results:

• Theoretical analysis demonstrates the bias in existing propensity-based methods when unmeasured confounders are present.
• The proposed doubly robust estimator combined with robust deconfounding effectively mitigates the impact of unmeasured confounders.
• Extensive experiments on three real-world datasets validate the effectiveness of the proposed methods.

Conclusions:

• The proposed PU learning formulation and debiasing framework offer a robust solution for recommender systems with implicit feedback.
• The novel approach effectively handles missing negative labels and unmeasured confounders, enhancing model reliability.
• This work advances the state-of-the-art in recommender systems by providing a more principled and effective method for handling data biases.