EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
3756. From Personalization to Optimization: A Framework for Enhanced Product Recommendations
Invited abstract in session WC-50: E-Commerce, stream Retail Operations.
Wednesday, 12:30-14:00Room: M2 (building: 101)
Authors (first author is the speaker)
| 1. | ATAKAN YILMAZ
|
| Data Science, Trendyol | |
| 2. | Onur Demiray
|
| Data Science, Trendyol | |
| 3. | Betül Ahat
|
| Data Science, Trendyol | |
| 4. | Ahmet Çınar
|
| Data Science, Trendyol | |
| 5. | Berker Özbek
|
| Data Science, Trendyol |
Abstract
Pointwise ranking models diverge from conventional recommendation systems by tailoring item rankings directly to individual user preferences, thereby offering nuanced recommendations aligned with user tastes. However, overlooking pairwise, listwise and position based factors in ranking output can diminish the overall user experience. To obtain user-specific rankings while considering the whole ranked list, we propose a framework that synergistically leverages machine learning and optimization techniques. Focused on personalizing product recommendations within the Trendyol app, our approach integrates xGBoost Classifier and integer programming to derive item scores and optimize rankings while addressing diversity concerns across categories and brands. Our framework starts with using nearly 200 features to learn user action probabilities. Then, an integer programming model determines rankings of products by maximizing expected purchases while ensuring diversification. This process operates in near real-time, updating personalized rankings every 2-3 minutes for 150,000 users. To tackle scalability challenges, we propose an alternative formulation and decomposition algorithm. Lastly, we introduce an ML-based multi-objective optimization framework to extend our focus to other performance metrics such as position-based utilities and clicks, alongside purchases. Rigorously tested via our effective AB testing system, we comprehensively analyze and discuss our results.
Keywords
- E-Commerce
- Machine Learning
- Programming, Integer
Status: accepted
Back to the list of papers