TL;DR
This paper tackles the challenge of automatically categorizing e-commerce products using only their titles, a crucial step for improving search and recommendations. They developed a multi-class classification system that achieves an impressive .88 f-score on a large dataset, outperforming previous methods by leveraging neural network embeddings.
What Problem Does This Paper Address?
Online shopping platforms like Amazon, eBay, and Google rely heavily on product taxonomies to organize millions of items. This organization is essential for effective search, product recommendations, and duplicate removal. However, sellers often struggle to manually categorize their products across different platforms, leading to inconsistencies and significant labor. Existing automated methods often require extensive product descriptions, but this paper focuses on the more challenging task of classifying products using only their titles, aiming to alleviate human effort and improve categorization accuracy.
Methods
The authors propose a multi-class classification approach to assign product titles to categories within a hierarchical taxonomy. Their core methodology involves extracting various features from product titles:
- Lexical Features: Standard unigram and bigram features are used as a baseline.
- Mutual Information Dictionary: They create a custom dictionary by calculating the mutual information between words and product categories. This helps identify words strongly associated with specific categories.
- LDA Topics: Latent Dirichlet Allocation (LDA) is employed to uncover latent topics within product titles, capturing semantic meaning beyond individual words.
- Neural Network Embeddings: This is their key innovation. They use a feedforward neural network (similar to Word2Vec) to learn dense vector representations (embeddings) for words and phrases within product titles. For a whole title, they average the embeddings of its constituent words.
These features are then fed into two multi-classification algorithms: One-Against-All (OAA) and Error Correcting Tournament (ECT).
Datasets
The authors used a dataset of 445,408 manually annotated product titles from Yahoo's shopping platform. This dataset is organized into a rich product taxonomy with 319 distinct categories spread across 6 hierarchical levels. The data was split into a training set of 353,809 examples and a test set of 91,599 examples.
Key Insights and Ideas
- Title-only classification is feasible and valuable: The paper demonstrates that accurate product categorization can be achieved using only titles, which is a significant step towards automating a labor-intensive process.
- Neural network embeddings are powerful: The study highlights that dense vector representations learned through neural networks are the most effective features for this task, outperforming traditional lexical and topic-based features.
- Predicted categories can be more descriptive: Interestingly, the authors found that their model sometimes predicts categories that are more specific and fine-grained than those provided by human annotators, suggesting potential for improving existing taxonomies.
Key Results
The best-performing system, utilizing neural network embeddings with the One-Against-All (OAA) classifier, achieved an f-score of .88. This significantly outperforms other feature combinations, such as unigrams/bigrams (.72 OAA), MI Dictionary (.85 OAA), and LDA Topics (.79 OAA). The Error Correcting Tournament (ECT) classifier generally performed slightly lower than OAA across all feature sets.
Limitations and Future Work
The authors acknowledge that their model sometimes predicts categories that are semantically similar but not an exact match to the gold standard. They also note that their predicted categories can be more specific than human annotations, which, while potentially beneficial, means they are penalized in an exact match evaluation. Future work could explore how to better integrate these more specific predictions or refine the evaluation metrics to account for semantic similarity. The paper also suggests that further exploration of feature combinations and classification algorithms could yield even better results.
Summary updated