Leveraging Product Adopter Information from Online Reviews for Product Recommendation

Jinpeng Wang, Wayne Xin Zhao, Yulan He, Xiaoming Li

Abstract: The availability of the sheer volume of online product reviews makes it possible to derive implicit demographic information of product adopters from review documents. This paper proposes a novel approach to the extraction of product adopter mentions from online reviews. The extracted product adopters are then categorise into a number of different demographic user groups. The aggregated demographic information of many product adopters can be used to characterise both products and users, which can be incorporated into a recommendation method using weighted regularised matrix factorisation. Our experimental results on over 15 million reviews crawled from JINGDONG, the largest B2C e-commerce website in China, show the feasibility and effectiveness of our proposed framework for product recommendation.

Introduction

Standard recommendation engines typically operate on a simple assumption: the person purchasing a product is the one intended to use it. However, in e-commerce, a significant portion of transactions involves "purchasing for others"—buying gifts for children, spouses, or parents. This creates a demographic mismatch. For instance, if a middle-aged man buys a toy for his daughter, traditional collaborative filtering might incorrectly assume he has developed a personal interest in toys, leading to irrelevant future recommendations.

This research addresses this "mismatch" problem by mining "product adopter" mentions from online reviews (e.g., "I bought this for my son") to infer the actual user's demographics. By aligning a user's purchasing habits with a product's target audience, recommendation accuracy can be significantly improved.

Methodology

The proposed framework consists of three main stages:

  1. Unsupervised Extraction: To avoid massive manual labeling, a bootstrapping algorithm was employed. Starting with seed patterns (e.g., "buy [someone] [something]"), the system iteratively scanned 139 million reviews to identify new linguistic patterns and adopter mentions, using Jaccard similarity to maintain quality.
  2. Demographic Categorization: Extracted mentions (like "grandpa" or "little niece") were mapped into six broad categories: Children, Young female, Old female, Young male, Old male, and Colleagues.
  3. Preference Biased Matrix Factorization (PBMF): The core algorithm modifies standard Matrix Factorization by adding a "confidence" weight to the loss function. This weight is derived from the probability that a user would buy a product based on their historical preference for specific adopter categories.

Data Analysis and Stability

The study utilized a massive dataset from JINGDONG (JD.com), containing 139 million reviews. Analysis revealed that roughly 10.8% of reviews contain explicit adopter mentions, representing a vast untapped data source.

A key finding is the stability of these distributions. While user preferences (who they buy for) can be dynamic, the product adopter distributions (who a product is for) are remarkably stable. As shown in the KL Divergence analysis, both distributions converge quickly, providing a reliable basis for recommendation modeling.

Stability of product adopter and user preference distributions over time.

Experimental Results

The proposed PBMF method was compared against Most Popular (MP), User-based Collaborative Filtering (UCF), and standard Matrix Factorization (MF).

Across various training-to-test data ratios, PBMF consistently outperformed all baselines. The performance gap was particularly notable when the number of latent factors was small, suggesting that incorporating adopter information allows the model to learn effective representations more efficiently than standard methods.

AUC performance comparison between UCF, MF, and PBMF with a 2:1 training/test ratio.

When the training data was at least double the size of the test set, PBMF showed a 12-18% improvement in Precision and Recall over standard MF. This demonstrates that demographic context extracted from text is a powerful signal for resolving the buyer-user mismatch.

Limitations and Future Work

Currently, the model treats all adopter mentions equally, regardless of sentiment. Future iterations will integrate sentiment analysis to distinguish between successful gifts and those the recipient disliked. Additionally, the researchers aim to move beyond the six broad categories toward more fine-grained demographic segmentation and test the framework on English-language datasets like Amazon.

Summary updated

Abstract

The availability of the sheer volume of online product reviews makes it possible to derive implicit demographic information of product adopters from review documents. This paper proposes a novel approach to the extraction of product adopter mentions from online reviews. The extracted product adopters are then categorise into a number of different demographic user groups. The aggregated demographic information of many product adopters can be used to characterise both products and users, which can be incorporated into a recommendation method using weighted regularised matrix factorisation. Our experimental results on over 15 million reviews crawled from JINGDONG, the largest B2C e-commerce website in China, show the feasibility and effectiveness of our proposed framework for product recommendation.
BibTeX
@inproceedings{wang2021leveraging,
  author = {Jinpeng Wang and Wayne Xin Zhao and Yulan He and Xiaoming Li},
  title = {{Leveraging Product Adopter Information from Online Reviews for Product Recommendation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v9i1.14585},
}
Figure