Disentangling Interest and Conformity for an Adversarial Learning-Based Personalized Recommendation System

Welcome to the code repository for the DICA project.

🚀 About the Project

The DICA model addresses the challenges of popularity bias in recommendation systems, which often overemphasize popular items

This project aims to:

• Rank learning with DICE

  We utilize the DICE mechanism to learn rankings, which employs disentangled embeddings of conformity and interest.

---

• Adversarial learning

  Additionally, adversarial learning is employed to promote fair recommendations between popular and unpopular items.

---

Data

We used the Movielens-10M and Netflix datasets. To convert explicit feedback data into implicit feedback data, we used only the data with a rating of 5 as an indicator of positive feedback.

Dataset	User	Item	Interaction
Movielens-10M	37,962	4,819	1,371,473
Netflix	32,450	8,432	2,212,690

---

Experiment result

We experienced a slight decrease in performance in terms of accuracy metrics(Recall, HR, NDCG), but achieved significant improvements in fairness metrics(RSP, REO). Notably, we achieved an over 10% improvement in the RSP metric.

This graph shows the interaction score distribution for combinations of user and item groups, categorized by popularity. The overlap of all four graph types indicates that the proposed model fairly recommends items from both popular and less popular groups to both user types.