Abstract
Learning attribute applicability of products in the Amazon catalog (e.g., predicting that a shoe should have a value for size, but not for battery-type) at scale is a challenge. The need for an interpretable model is contingent on (1) the lack of ground truth training data, (2) the need to utilise prior information about the underlying latent space and (3) the ability to understand the quality of predictions on new, unseen data. To this end, we develop the MaxMachine, a probabilistic latent variable model that learns distributed binary representations, associated to sets of features that are likely to co-occur in the data. Layers of MaxMachines can be stacked such that higher layers encode more abstract information. Any set of variables can be clamped to encode prior information. We develop fast sampling based posterior inference. Preliminary results show that the model improves over the baseline in 17 out of 19 product groups and provides qualitatively reasonable predictions.
Introduction
In a massive e-commerce catalog like Amazon's, products are defined by a vast array of attributes. However, attribute applicability is highly context-dependent: a t-shirt requires a "size" but not a "battery type," while a book requires a "page count" but not a "waist size." Ensuring high data quality is a significant challenge; missing attributes hinder search and comparison, while irrelevant "hallucinated" attributes clutter the user experience. This problem is characterized by a lack of ground truth, extreme scale and sparsity, and a critical need for interpretability so that engineers can validate model predictions before deployment.
The MaxMachine Model
To address these challenges, researchers from Amazon and the University of Oxford developed MaxMachine, a generative model extending Boolean Matrix Factorization (BMF). In standard BMF, a binary data matrix is decomposed into latent features representing sets of attributes that co-occur. MaxMachine introduces several key innovations:
- The Max-Operation: Unlike standard Boolean disjunction, the model employs a
maxoperation in its likelihood function. This allows latent dimensions to "compete" to explain an observation. The most accurate predictor for a specific data point "wins," allowing the model to handle heteroscedastic noise across diverse product types. - Hierarchical Structure: The model incorporates product categories (e.g., "Shoes," "Electronics") as priors. A stacked architecture allows higher layers to encode abstract category information, which then informs lower-level attribute assignments.
- Fast Inference: By utilizing binary states and the decoupling effect of the max-operation, the authors implemented a sampling-based inference method capable of processing hundreds of thousands of data points in minutes.
Experimental Setup
The model was evaluated using a snapshot of the Amazon German marketplace, stratified into 19 product clusters. Training was performed on 500 randomly sampled products per type, as binary attribute data exhibits high redundancy. Performance was measured using a "held-out" approach, where a percentage of known product-attribute pairs were treated as unobserved during training to test predictive accuracy.
Results and Discussion
MaxMachine demonstrated both quantitative and qualitative success. It achieved a ROC-AUC of 94% on held-out data for clothing products, outperforming the product-type frequency baseline (93%). Across the entire catalog, it surpassed the baseline in 17 out of 19 product clusters.
The model's interpretability is a primary strength. By using binomial priors to maintain sparsity, the latent dimensions naturally form "attribute bundles"—logical groupings like "clothing dimensions" (waist-style, inseam-length) or "book metadata."
Qualitatively, the model correctly assigned high applicability probabilities to specific attributes even when they were missing from the raw data. For example, it identified "cup-size" as applicable to "bra" products while maintaining near-zero probability for unrelated categories, demonstrating a nuanced understanding of product domains.
Conclusion
MaxMachine provides a scalable, interpretable solution for predicting attribute applicability in large-scale catalogs. Future work will explore non-parametric priors to automatically determine the number of latent features and methods to better integrate expert human knowledge through variable "clamping."
Summary updated