Manage Product Compatibility with Machine Learning

Context

Research Paper: https://arxiv.org/pdf/2206.13749

Year: 2022

Models: LLM, Decision Tree

Topic: Product compatibility based on structured and unstructured data

Keywords: #LLM #DecisionTree #RecommenderSystem

Introduction

The paper “Adaptive Multi-view Rule Discovery for Weakly-Supervised Compatible Products Prediction” by Rongzhi Zhang, Rebecca West, Xiquan Cui, and Chao Zhang presents an innovative method for predicting product compatibility on e-commerce platforms. This document, published at the 28th ACM SIGKDD Conference, addresses the challenges of product compatibility prediction due to heterogeneous data and the lack of manually labeled training data.

Background and Problem Statement

Importance of Compatibility Prediction

Predicting product compatibility is crucial for improving user experience on e-commerce platforms. Accurate recommendations of compatible products enable consumers to make informed and confident purchases, which increases customer satisfaction and potentially boosts sales.

Challenges

1. Heterogeneous Data: Products may have both structured (product attributes) and unstructured (textual descriptions) data.
2. Lack of Manually Labeled Data: The process of manually labeling data is expensive and time-consuming. The available data is often noisy and incomplete.

Methodology

The paper proposes AMRule, a multi-view rule discovery framework that combines structured and unstructured data to improve product compatibility prediction.

Core Principles

1. Weakly Supervised Learning: Utilizes user behavioral data (co-purchase) as a source of supervision to generate weakly labeled instances.
2. Adaptive Rule Discovery: Uses a boosting strategy to discover new rules by focusing on hard instances where the current model makes errors.
3. Multi-view Integration: Discovers rules from both structured attributes and unstructured product descriptions.

Methodology Steps

1. Rule Discovery from Structured Attributes:

   • Decision trees are used to generate rules based on product attributes.
   • Rules are created by combining categorical and numerical attributes to capture complex relationships between products.

2. Rule Discovery from Unstructured Descriptions:

   • Utilizes pre-trained language models (LLM) to generate rules based on textual data.
   • Product descriptions are used to complement the missing information from structured attributes.

3. Rule Annotation and Matching:

   • Proposed rules are evaluated by human annotators to ensure their quality.
   • Validated rules are used to label new instances, thus expanding the training dataset.

4. Learning and Ensemble Modeling:

   • A weighted ensemble model is used where each weak model contributes to the final model based on its error rate.
   • This approach improves overall performance by integrating complementary models.

Experimental Results

The experiments were conducted on real-world datasets from The Home Depot, covering four product categories: Lighting, Appliances, Bath, and Tools.

Performance

• AMRule outperformed baseline methods by an average of 5.98%.
• The boosting strategy and multi-view rule discovery showed significant improvement over static methods and iterative approaches.

Component Analysis

1. Multi-view Rules: Integrating rules based on structured attributes and unstructured descriptions provided substantial performance gains.
2. Adaptive Rule Discovery: The boosting strategy helped target hard instances and improved the quality of proposed rules.

Case Study

A concrete example on power tools and batteries illustrated how AMRule discovers and proposes rules based on product attributes and descriptions, demonstrating the effectiveness of the method.

Conclusion

AMRule proposes an innovative approach to product compatibility prediction by using multi-view rules and an adaptive boosting strategy. This method not only improves the performance of weakly supervised models but also offers better interpretability of predictions through rule discovery.