Source-linked AI summary
Inferring Networks of Substitutable and Complementary Products
Julian McAuley, Rahul Pandey, Jure Leskovec
TL;DR
Recommender systems need to distinguish products that substitute for one another from products purchased together, while product relationships remain difficult to model at scale. Sceptre addresses this with supervised link prediction and topic models over product text and metadata, evaluated on Amazon data, where it accurately predicts both relation types and text is the most useful information source.
Problem
The paper addresses how to infer and explain substitute and complementary relationships between millions of products for context-relevant recommendation.
Method
Sceptre combines topic modeling and supervised link prediction over reviews and descriptions with brand, price, rating, and category features.
Results
Sceptre accurately predicts both substitute and complement links across product categories, with binary-task accuracy between 85.57-96.76%.
Takeaways & Limitations
Product text, including reviews, descriptions, and specifications, is the most useful source for identifying and explaining substitute and complementary relationships.
Takeaways & Limitations
The optimization objective is smooth and has multiple local minima caused by permutations of topic and logistic-parameter order.
Abstract
from arXiv · showhide
In a modern recommender system, it is important to understand how products relate to each other. For example, while a user is looking for mobile phones, it might make sense to recommend other phones, but once they buy a phone, we might instead want to recommend batteries, cases, or chargers. These two types of recommendations are referred to as substitutes and complements: substitutes are products that can be purchased instead of each other, while complements are products that can be purchased in addition to each other. Here we develop a method to infer networks of substitutable and complementary products. We formulate this as a supervised link prediction task, where we learn the semantics of substitutes and complements from data associated with products. The primary source of data we use is the text of product reviews, though our method also makes use of features such as ratings, specifications, prices, and brands. Methodologically, we build topic models that are trained to automatically discover topics from text that are successful at predicting and explaining such relationships. Experimentally, we evaluate our system on the Amazon product catalog, a large dataset consisting of 9 million products, 237 million links, and 144 million reviews.
1 Introduction
The paper frames product-relationship inference as a way to generate context-relevant recommendations, focusing on substitute and complementary links. Sceptre learns these relationships from product text and additional metadata, producing product graphs and fine-grained micro-categories.
- Motivation: Product graphs connect products through relationships such as substitution and complementarity to support navigation, discovery, and product-combination recommendations.Examples include recommending similar t-shirts, complementary clothing items, and related products instead of merely matching user interests.
- Relationship types: Sceptre models substitutes as interchangeable products and complements as products that might be purchased together.The system can identify substitutes and complements for a query item, such as hiking boots, heavy-duty socks, and shoe polish.
- Method: Sceptre combines topic modeling with supervised link prediction to learn predictive topics from product reviews and descriptions.It also incorporates brand, price, rating, and category information and can predict multiple relation types simultaneously.
- Method: Sceptre uses product-category hierarchies to discover fine-grained micro-categories of closely related products.These micro-categories extend the existing hierarchy with more specific groupings.
- Evaluation and applications: The system is trained on 9 million Amazon products and 237 million browsing and co-purchasing connections, with text identified as the most useful information source.The resulting product graphs support ranked substitute and complementary recommendations and candidate generation for context-relevant recommendations.
2 Related work
Related work has largely used recommender systems to estimate user preferences or mine behavioral signals, while this paper focuses on product relationships. It builds on topic models and related approaches for extracting relationships from large, weakly structured data sources.
- Recommender systems: Traditional recommender systems estimate users’ ratings or rankings, whereas this work aims to discover relationships between products.The distinction separates product-link inference from preference estimation.
- Behavioral data: Behavioral co-counting from browsing and co-purchasing logs can produce noisy recommendations for infrequently purchased products and offers limited explanatory power.Examples include “users who bought X also bought Y” and “users who viewed X also viewed Y.”
- Topic models: Topic models have been used in recommender systems mainly to predict user ratings, while this paper uses them to learn product relationships.The approach builds on topic models for networks, including Block-LDA, topic-link LDA, and relational topic models.
- E-commerce data: Related e-commerce work extracts user-relevant attributes or commercial intent from query and browsing data.These studies share the use of large weakly structured data sources to uncover relationships.
3 The Sceptre Model
Sceptre jointly learns topic representations and supervised link predictors to infer product relationships, using sparse hierarchical topics and features for multiple relation types.
- Link Prediction: Products are represented by topic vectors, and logistic regression predicts relationships from pairwise features combining the topics of two products.The adopted pairwise feature vector includes an intercept and elementwise products of corresponding topic memberships.
- High-level Overview: Sceptre combines topic modeling with supervised link prediction, learning topics that are informative for predicting product relationships.The joint objective discovers topics that help explain relationships rather than training topics independently before link prediction.
- Product Graph: The framework learns from product graphs whose construction encodes relationship semantics, such as complements from co-purchasing and substitutes from subsequent purchases after viewing.For each product, the model predicts lists of complementary and substitutable products.
- Directed vs. Undirected Graphs: Sceptre separates relatedness prediction from edge-direction prediction to model asymmetric product relationships and improve interpretability.The relatedness predictor tends to use high-level category topics, while the direction predictor captures qualitative language describing why one product is better than another.
- Sparse Topics via Product Hierarchies: Sceptre uses product hierarchies to impose sparse topic representations, allowing products to share high-level topics while retaining subcategory-specific topics and microcategories.Each hierarchy node has associated topics, restricting which topics a product can activate.
4 Experiments
Sceptre is evaluated on large Amazon product graphs using link-prediction and ranking tasks, with experiments testing data construction, baselines, cold-start text, topic structure, and human judgments. It accurately predicts substitutes and complements across categories, improves substantially over baselines, and produces useful complement recommendations.
- Experimental setting: Sceptre is trained and evaluated on Amazon data, with category-specific splits used mainly to fit models within single-machine memory constraints.The crawl is described as a comprehensive English-language collection; the largest experiment required around 64GB RAM and 2–3 days.
- Evaluation tasks and baselines: Sceptre trains supervised link predictors and rankings against random, LDA, TF-IDF, and collaborative-filtering baselines, using balanced non-edge samples and cross-type links as hard negatives.Substitute links are treated as negative examples for complement prediction and vice versa, in addition to randomly sampled non-edges.
- Link prediction and ranking: 85.57–96.76% accuracy is achieved across the binary substitute and complement tasks, with especially accurate performance for clothing and electronics.The reported range covers the product categories evaluated in Table 3.
- Link prediction and ranking: Sceptre is an order of magnitude more accurate than collaborative filtering, which is itself one to two orders of magnitude above random rankings.LDA and TF-IDF baselines were less accurate than collaborative filtering and are not shown in the ranking comparison.
- Cold-start and topic analysis: Sceptre reaches 91.28–93.67% accuracy using book and electronics descriptions, while its topics discover fine-grained distinctions such as headphone price tiers and running styles.These results support applying the approach when reviews are unavailable and show that the hierarchy yields micro-categories beyond explicit catalog categories.
- User study: In a 200-item clothing user study, Amazon’s browsing links were preferred for substitutes, whereas Sceptre’s complement links were overwhelmingly preferred.Judges separately evaluated the top recommendation and top-five recommendations.
5 Building the product graph
The product graph ranks substitute and complementary links to surface the highest-likelihood recommendations despite the impracticality of exhaustive comparison. Sceptre makes this feasible through search-space restriction, efficient enumeration, cold-start topic estimation, and review-based explanations.
- Ranking recommendations: Sceptre ranks links separately for each graph type and recommends the highest-likelihood substitute or complementary products.The recommendation set excludes the queried product itself.
- Ranking recommendations: Restricting candidates to the hundred-thousand most popular products makes large-scale enumeration feasible within a few milliseconds.Scoring one candidate edge costs O(K), but evaluating all possible recommendations is otherwise too slow.
- Cold-start handling: New products need not trigger retraining because their topic distributions can be estimated under the existing model.The method assumes that adding a small number of products does not fundamentally change the word distribution φ.
- Recommendation explanations: Sceptre explains recommendations by highlighting review sentences whose language most strongly supports the predicted relationship.The interface uses discovered topics to select explanatory sentences.
- Reproducibility: The paper provides the data, code, and Figure 6 interface for reproducing Sceptre.These materials are available on the first author’s webpage.
6 Conclusion
The paper targets contextually relevant product recommendations by learning substitute and complementary relationships from product features, especially review text. It presents Sceptre and applies it to a large Amazon collection, with examples and evaluation interfaces illustrating the resulting recommendations.
- Conclusion: Sceptre learns substitute and complementary relationships from product features, especially product-review text.These relationships distinguish products purchased instead of each other from products purchased in addition to each other.
- Conclusion: Sceptre is presented as a model for predicting and understanding relationships between linked products.The application focuses on identifying substitutable and complementary products.
- Conclusion: Figure 6 shows Sceptre recommendations with substitutes on the left, complements on the right, and a Mechanical Turk preference interface.Turkers compare Amazon’s lists with Sceptre’s lists.
- Conclusion: The evaluation uses 237 million ground-truth relationships based on browsing and co-purchasing logs.The product collection also includes 144 million reviews.