2846. Robust Product Recommendations via Graph-Driven Co-Occurrence Analysis and Adaptive User Clustering
Invited abstract in session WA-34: Advancements of OR-analytics in statistics, machine learning and data science 6 , stream Advancements of OR-analytics in statistics, machine learning and data science.
Wednesday, 8:30-10:00Room: Michael Sadler LG10
Authors (first author is the speaker)
| 1. | Serhat Mirkan Acar
|
| Data Science, Trendyol | |
| 2. | ATAKAN YILMAZ
|
| Data Science, Trendyol | |
| 3. | Egemen Karabiyik
|
| Data Science, Trendyol |
Abstract
We propose a recommendation framework that infers user similarity from implicit interactions—such as views, favorites, and orders—while carefully reducing the skew introduced by highly popular products. Unlike conventional methods that treat every product identically, our approach applies a weighted formula to discount widely popular items, thereby emphasizing more distinctive user interactions. In addition to examining individual products, we build an undirected graph based on product pairs, assigning less influence to product pairs co-interacted by a large number of users. This refined strategy yields a more accurate measure of user alignment, enabling the construction of a user-user graph where edges capture these similarity scores. A label propagation process then groups closely related users into coherent clusters, and personalized recommendations are generated by suggesting items within each cluster that users have not yet encountered. Rigorously evaluated through our effective AB testing system, we comprehensively analyze and discuss the impact of our approach, demonstrating its potential to enhance personalization and enrich the user experience.
Keywords
- Algorithms
- Analytics and Data Science
- E-Commerce
Status: accepted
Back to the list of papers