Identify and explain bias in recommender systems
Bin Aamir, Saad (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-11-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025112310832
https://urn.fi/URN:NBN:fi:tuni-2025112310832
Tiivistelmä
E-commerce platforms thrive on recommendation systems that act like personal assistance, while guiding customers through a sea of products with ease. But these systems tend to play favorites, which means they prefer popular products and hot-selling items while leaving newer or less popular ones in the shadows. So even if those newly launched or low ranked products are just as good as the top products, they struggle to get noticed. After all, the recommendation system is not recommending them at all in top search results, which results in less or no sales. Apparently, this bias in recommendation system comes because of data bias or model bias. For my thesis, I’m focusing on model bias here.
Fact is that new sellers don’t have ratings or reviews to flaunt. Hence the algorithms stick to what’s already popular and ultimately favoring popular products in recommendations. Meanwhile, the psyche of a buyer is that they don’t consider going down in search rankings, and this is why popular and rated products gets all the sales. So, this thesis digs into the unfortunate bias happening in e-commerce recommendation systems, while exploring how these algorithms favors product rankings and influence customers’ purchase decisions. To highlight this, I’ve built a hybrid recommendation system by analyzing an Amazon dataset, that blends collaborative filtering (SVD) with content-based filtering (Item Similarity – Cosine Similarity). As a result, it becomes clear from this system that bias in recommendation ecommerce system is real and is a real issue for new sellers. In my implementation, I have displayed both graphical data and actual bias figures by analyzing dataset.
So, I’m using Amazon dataset from Kaggle to show how this bias plays out in product recommendations, and I’m using model bias to visualize this issue. In addition to identifying bias, this thesis also provides potential bias mitigation techniques. Goal is to help shape e-commerce platforms that give every seller a fair shake, not just benefitting the top sellers, and open the door to a more balanced online shopping environment.
Fact is that new sellers don’t have ratings or reviews to flaunt. Hence the algorithms stick to what’s already popular and ultimately favoring popular products in recommendations. Meanwhile, the psyche of a buyer is that they don’t consider going down in search rankings, and this is why popular and rated products gets all the sales. So, this thesis digs into the unfortunate bias happening in e-commerce recommendation systems, while exploring how these algorithms favors product rankings and influence customers’ purchase decisions. To highlight this, I’ve built a hybrid recommendation system by analyzing an Amazon dataset, that blends collaborative filtering (SVD) with content-based filtering (Item Similarity – Cosine Similarity). As a result, it becomes clear from this system that bias in recommendation ecommerce system is real and is a real issue for new sellers. In my implementation, I have displayed both graphical data and actual bias figures by analyzing dataset.
So, I’m using Amazon dataset from Kaggle to show how this bias plays out in product recommendations, and I’m using model bias to visualize this issue. In addition to identifying bias, this thesis also provides potential bias mitigation techniques. Goal is to help shape e-commerce platforms that give every seller a fair shake, not just benefitting the top sellers, and open the door to a more balanced online shopping environment.