1. Popular items are recommended even more frequently than their popularity would warrant

2. The long-tail phenomenon is common in RS data: in most cases, a small fraction of popular items account for most user interactions

3. When trained on such long-tailed data, the model usually gives higher scores to popular items than their ideal values while giving unpopular items lower scores.

Importance of addressing popularity bias in recommender systems

1. It decreases personalization by always recommending popular items that will hurt user experience, especially for the users favouring niche items. [Improved user-level personalisation for users with niche tastes]

2. It decreases the fairness of the recommendation results (Fairness refers to the equitable treatment of all items and user groups, regardless of their popularity or user preferences). [Improve the visibility of a wider range of items]

3. Popular items are not always of high quality. Over-recommending popular items will reduce the visibility of other items even if they are good matches, which is unfair. [Propagation of new shows with high retention, ARPU takes a lot of time]

4. Popular bias will further increase the exposure opportunities of popular items, making popular items even more popular – the collected data for future training becomes more unbalanced. [The loss function becomes dominated by these popular items, potentially leading to biased learning and less effective recommendations for niche or less popular content]

Why Popularity bias is not so bad?

1. User Satisfaction: Popular items are typically popular for a reason - they are often well-liked and enjoyed by a large portion of the user base. Recommending popular items can increase user satisfaction by providing content that aligns with the preferences of the majority of users.

2. Discovery of High-Quality Content: Popular items often have higher quality, making them more likely to be enjoyed by users. Recommending these items can help users discover content that is generally well-received and of high quality.

3. Reduced Risk: Popular items have already been vetted by a large number of users, reducing the risk of recommending content that users may not enjoy. This can lead to fewer instances of dissatisfaction with recommendations and increased user trust in the recommendation system.

4. Engagement and Retention: Recommending popular items can increase user engagement and retention by providing content that users are more likely to enjoy and spend time consuming. This can lead to higher user satisfaction and loyalty to the platform.

Measuring Popularity Bias

1. Popularity Ratio: Calculate the ratio of recommended popular items to less popular items. A high ratio indicates a bias towards recommending popular items.

2. Item Coverage: Evaluate the proportion of unique items in the recommendation list that are popular versus less popular. A low coverage of less popular items indicates a bias towards recommending popular items.

3. Diversity Metrics: Use diversity metrics such as entropy or diversity indices to measure the variety of recommended items. A low diversity score may indicate a bias towards recommending a narrow set of popular items.