Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Text-Based Product Matching -- Semi-Supervised Clustering Approach

Alicja Martinek, Szymon Łukasik, Amir H. Gandomi

TL;DR

This work tackles product matching across feeds by reframing the task as a semi-supervised constrained clustering problem using Deep Embedded Clustering (IDEC). It constructs a 5-feature textual similarity vector from fuzzy string metrics and distances, and enforces Must Link and Can't Link constraints to guide clustering toward two distinct groups ($2$ clusters) corresponding to matching vs. non-matching pairs. Across the Skroutz dataset and additional camera-related corpora, IDEC with carefully chosen constraints consistently outperforms baselines like k-means, XGBoost, and DeepMatcher, achieving a peak $F_1$ score around 0.917 and high Rand Index, indicating strong pairwise agreement with ground truth. The results suggest that semi-supervised, text-centric product matching can reduce labeling requirements while delivering robust performance on real-world, noisy product data, with potential for broader deployment and extensions to other constrained clustering algorithms and richer feature sets.

Abstract

Matching identical products present in multiple product feeds constitutes a crucial element of many tasks of e-commerce, such as comparing product offerings, dynamic price optimization, and selecting the assortment personalized for the client. It corresponds to the well-known machine learning task of entity matching, with its own specificity, like omnipresent unstructured data or inaccurate and inconsistent product descriptions. This paper aims to present a new philosophy to product matching utilizing a semi-supervised clustering approach. We study the properties of this method by experimenting with the IDEC algorithm on the real-world dataset using predominantly textual features and fuzzy string matching, with more standard approaches as a point of reference. Encouraging results show that unsupervised matching, enriched with a small annotated sample of product links, could be a possible alternative to the dominant supervised strategy, requiring extensive manual data labeling.

Text-Based Product Matching -- Semi-Supervised Clustering Approach

TL;DR

This work tackles product matching across feeds by reframing the task as a semi-supervised constrained clustering problem using Deep Embedded Clustering (IDEC). It constructs a 5-feature textual similarity vector from fuzzy string metrics and distances, and enforces Must Link and Can't Link constraints to guide clustering toward two distinct groups ( clusters) corresponding to matching vs. non-matching pairs. Across the Skroutz dataset and additional camera-related corpora, IDEC with carefully chosen constraints consistently outperforms baselines like k-means, XGBoost, and DeepMatcher, achieving a peak score around 0.917 and high Rand Index, indicating strong pairwise agreement with ground truth. The results suggest that semi-supervised, text-centric product matching can reduce labeling requirements while delivering robust performance on real-world, noisy product data, with potential for broader deployment and extensions to other constrained clustering algorithms and richer feature sets.

Abstract

Matching identical products present in multiple product feeds constitutes a crucial element of many tasks of e-commerce, such as comparing product offerings, dynamic price optimization, and selecting the assortment personalized for the client. It corresponds to the well-known machine learning task of entity matching, with its own specificity, like omnipresent unstructured data or inaccurate and inconsistent product descriptions. This paper aims to present a new philosophy to product matching utilizing a semi-supervised clustering approach. We study the properties of this method by experimenting with the IDEC algorithm on the real-world dataset using predominantly textual features and fuzzy string matching, with more standard approaches as a point of reference. Encouraging results show that unsupervised matching, enriched with a small annotated sample of product links, could be a possible alternative to the dominant supervised strategy, requiring extensive manual data labeling.
Paper Structure (13 sections, 4 equations, 5 figures, 3 tables)

This paper contains 13 sections, 4 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Product matching
  4. Transforming textual data into numerical features
  5. Clustering
  6. Evaluation metrics
  7. Proposed Algorithm
  8. Experimental Settings and Results
  9. Constraints impact
  10. Comparison with other methods
  11. Other datasets
  12. Class distribution impact
  13. Conclusion

Figures (5)

  • Figure 1: Example of Must Link (solid line) and Can't Link (dashed line) constraints
  • Figure 2: Impact of increasing the amount of Must Link Constraints
  • Figure 3: Impact of increasing the amount of Can't Link Constraints
  • Figure 4: Impact of increasing the amount of 1-1 pairs in Must Link Constraints
  • Figure 5: Impact of increasing the amount of matching pairs in the dataset