Popularity Bias in Recommendation systems:

Popularity bias in recommender systems is a problem where popular items are recommended more frequently, leading to less popular items being overlooked. This bias can be harmful for both users and item providers. While many studies have looked at this issue, most focus on a static setting, analyzing bias for a single round of recommendations using logged data. However, real-world recommendation processes are dynamic, which raises questions like: How does popularity bias change over time? What unique factors affect bias in dynamic recommendations? How can we address bias in the long term?

Imagine you’re using a movie streaming service like Netflix. The recommendation system suggests movies based on what you’ve watched before and what others with similar tastes have liked. However, the system doesn’t consider how your preferences might change over time.

For instance, let’s say you recently started watching superhero movies and enjoyed them. The recommendation system starts suggesting more superhero movies and fewer of other genres you used to like. However, after a while, your interest in superhero movies wanes, and you start enjoying documentaries instead.

The issue here is that the recommendation system doesn’t adapt to this change in your preferences. It continues to recommend superhero movies based on your past behavior, failing to account for the dynamic nature of your movie preferences in the real world.

To tackle these questions, the authors conducted simulation experiments. They proposed a dynamic debiasing strategy and a novel False Positive Correction method. This method uses false positive signals to reduce bias, and in their experiments, it showed effective performance.

In simpler terms, imagine you’re looking for a book to read. If the recommender system only suggests popular books, you might miss out on hidden gems that you’d love. The authors studied how this bias changes over time and proposed a way to make recommendations fairer, even as trends and preferences shift.

What is popularity bias?

Alright, class, let’s talk about popularity bias in recommender systems! Imagine you’re at a school where the cafeteria always serves the same popular dish, like pizza, every day. Everyone knows about pizza and likes it, so it’s a safe choice for the cafeteria. But what about those students who might really enjoy trying new foods or have dietary preferences that aren’t met by pizza?

This is similar to what happens in recommender systems. They often recommend popular items, like blockbuster movies or hit songs, because they’re safe choices that many people like. However, this can mean that other, less popular items, which some people might really enjoy, are overlooked.

Most studies on popularity bias look at a static setting, like looking at what movies people watch in a month. But this doesn’t consider how people’s tastes change over time, like maybe you’re into superhero movies this month but next month you’re all about documentaries.

So, this paper is like a recipe for understanding popularity bias in a more dynamic way. It wants to figure out how this bias changes over time, what factors affect it, and how we can make recommendations more fair and diverse, just like adding more variety to the school cafeteria menu!

Dynammic Recommendation:

Let us understand it in an easy-to-understand way:

1. Dynamic recommendation can be seen as a loop: This means that recommending things to users is like a circle that keeps going around and around.
2. Users interact with the system by doing things like clicking, viewing, or rating items: This is how users tell the system what they like and don’t like.
3. User-feedback data is used to train a recommendation model: The system uses the information from users to learn what to recommend in the future.
4. Trained model recommends new items: Based on what it learned, the system suggests new things to users.
5. Opportunities for bias to affect this process: There are many ways that the recommendations can be biased.
6. Three key factors that may impact popularity bias:

• Inherent audience size imbalance: Some items are liked by more people than others, even if the recommendations are fair.
• Model bias: The recommendation system itself might make the popularity bias worse by favoring already popular items.
• Closed feedback loop: Since the system learns from what users do, bias can accumulate over time if it keeps recommending the same types of things.

What is done to study popularity bias?

To understand popularity bias in dynamic recommendation, we did three things:

1. Empirical Study: We conducted experiments to see how popularity bias changes over time in dynamic recommendation. We looked at how the three factors we identified affect this bias.
2. Mitigation Methods: We explored ways to reduce popularity bias in dynamic recommendation. We adapted existing methods that were designed for static settings to work in dynamic scenarios. We also came up with a new method called False Positive Correction (FPC) that can be used alongside other methods to reduce bias even more.
3. Experimental Results: Finally, we tested our methods extensively to see how well they worked compared to other methods that are commonly used. We found that our approach was effective in reducing popularity bias in dynamic recommendation.

What has been done earlier to study popularity bias?

We will continue elaborating this in the next part of this blog post.

Thanks for reading!