1922. A Predict-then-Optimize Framework for Traffic Allocation in Local Commerce Recommendation Systems
Invited abstract in session TC-38: Forecasting, prediction and optimization 3, stream Data Science meets Optimization.
Tuesday, 12:30-14:00Room: Michael Sadler LG19
Authors (first author is the speaker)
| 1. | Yani Ji
|
| Tianjin university | |
| 2. | Yang Nan
|
| 美团 | |
| 3. | Ran Ao
|
| 美团 | |
| 4. | Yanting Jin
|
| Meituan | |
| 5. | Ning Zhu
|
| Tianjin University | |
| 6. | Zhu Meng
|
| 7. | Jie Chen
|
| College of Management and Economics, Tianjin University | |
| 8. | YUN LONG
|
| College of Management and Economics, Tianjin University |
Abstract
With the surging trend of users utilizing local commerce platforms to search for nearby in-store services, these online platforms require to perform local searches more efficiently and more accurately. This study introduces a data-driven predict-then-optimize framework for determining the optimal impression structure of supply assortments. Our approach leverages a multilayer perceptron (MLP) to estimate conversion rates by dissecting and modeling the interdependencies amongst user attention distribution, the positional distribution of category proportions, and quality stratification (e.g., price and distance segment distributions). The MLP extracts latent patterns from these multifaceted inputs, while a Gaussian Mixture Model (GMM) further captures the complex interplay among them. Building on these predictions, we formulate a nonlinear programming model (NLP) designed to maximize overall conversion rates and develop a tailored heuristic algorithm that obtains the impression structure of supply assortments. We test our approach on a real-world dataset from the online food service sector. The results reveals that our framework not only enhances the precision of conversion rate predictions but also achieves a 7% improvement in conversion rates over existing strategies, underscoring its practical applicability in enhancing the performance of local commerce recommendation systems.
Keywords
- Machine Learning
- Forecasting
- Optimization Modeling
Status: accepted
Back to the list of papers