Detection of Review Abuse via Semi-Supervised Binary Multi-Target Tensor Decomposition

Anil R. Yelundur, Vineet Chaoji, Bamdev Mishra

Abstract

Product reviews and ratings on e-commerce websites provide customers with detailed insights about various aspects of the product such as quality, usefulness, etc. Since they influence customers' buying decisions, product reviews have become a fertile ground for abuse by sellers (colluding with reviewers) to promote their own products or to tarnish the reputation of competitor's products. In this paper, our focus is on detecting such abusive entities (both sellers and reviewers) by applying tensor decomposition on the product reviews data. While tensor decomposition is mostly unsupervised, we formulate our problem as a semi-supervised binary multi-target tensor decomposition, to take advantage of currently known abusive entities. We empirically show that our multi-target semi-supervised model achieves higher precision and recall in detecting abusive entities as compared to unsupervised techniques. Finally, we show that our proposed stochastic partial natural gradient inference for our model empirically achieves faster convergence than stochastic gradient and Online-EM with sufficient statistics.

The Problem of Review Abuse

Product reviews heavily influence customer purchasing decisions, making them a prime target for manipulation. Unscrupulous sellers often collude with paid reviewers or hijack inactive accounts to post fake positive reviews for their own products or fake negative reviews for competitors. This recruitment often happens through intermediaries on social media platforms, where "agents" solicit individuals to perform quick review tasks for small payments.

An example of an intermediary recruiting fake reviewers on social media to manipulate product ratings.

To make a noticeable impact on a product's rating, a seller must recruit a large number of fake reviewers in a short period. In a data graph, this suspicious behavior manifests as "dense bipartite cores"—tightly knit clusters of connections between specific sellers, products, and reviewers, all sharing similar ratings within a narrow time window. The challenge lies in efficiently finding these multidimensional dense blocks in massive, highly sparse e-commerce datasets.

The recruitment of fake reviewers leads to the formation of dense bipartite cores, where a group of reviewers targets a specific set of sellers and products.

Methodology: The SENTINEL Approach

The authors model the e-commerce ecosystem as a multidimensional array, or tensor, where the dimensions (modes) are reviewer, product, seller, rating, and time. To identify dense cores, they utilize Logistic CP Tensor Decomposition, which breaks the binary tensor down into latent factors. The paper introduces two major innovations:

  1. Semi-Supervised Multi-Target Learning (SENTINEL): Standard tensor decomposition is typically unsupervised. SENTINEL incorporates partial supervision by using a small set of known abusive entities identified via manual audits. Furthermore, it treats different types of abuse (e.g., paid reviews vs. compromised accounts) as distinct but correlated targets, learning them jointly to improve predictive power.
  2. Stochastic Partial Natural Gradient Inference: Standard stochastic gradient descent (SGD) can be slow for large-scale tensor problems. The authors developed a scalable inference method that exploits the mathematical structure of their model (Pólya-Gamma data augmentation) to compute a partial Fisher matrix. This allows them to use natural gradients, which offer faster convergence, in a scalable stochastic setting.

Key Results and Real-World Impact

The researchers evaluated SENTINEL on both proprietary and public Amazon datasets, yielding several key insights:

  • Superior Detection Performance: SENTINEL achieved an AUC of 0.88 for seller abuse and 0.85 for reviewer abuse, significantly outperforming unsupervised baselines like BNBCP (which scored 0.74 and 0.61, respectively).
  • The Power of Supervision: Seeding the tensor decomposition with known bad actors forces the model to find new entities with similar behavioral signatures, drastically improving precision over purely unsupervised methods.
  • Multi-Target Boost: Jointly learning separate models for different abuse types yielded a 3% gain in AUC compared to learning a single model for all abuse types combined.
  • Early Detection: In a retrospective analysis, SENTINEL flagged 40% of abusive reviewers at "Week 0" that existing production systems took up to 5 weeks to catch. Banning these users earlier prevents a massive number of fake review impressions.

Conclusion and Future Work

SENTINEL demonstrates that combining partial supervision with advanced tensor decomposition can effectively unmask sophisticated review abuse. While the current model is highly effective, the authors note that tensor decomposition is inherently non-convex. Future work will explore the use of Graph Convolutional Networks (GCNs) and hierarchical Tucker decomposition (Convolutional Rectifier Networks) to further refine detection capabilities in the ever-evolving landscape of e-commerce fraud.

Summary updated

Abstract

Product reviews and ratings on e-commerce websites provide customers with detailed insights about various aspects of the product such as quality, usefulness, etc. Since they influence customers' buying decisions, product reviews have become a fertile ground for abuse by sellers (colluding with reviewers) to promote their own products or to tarnish the reputation of competitor's products. In this paper, our focus is on detecting such abusive entities (both sellers and reviewers) by applying tensor decomposition on the product reviews data. While tensor decomposition is mostly unsupervised, we formulate our problem as a semi-supervised binary multi-target tensor decomposition, to take advantage of currently known abusive entities. We empirically show that our multi-target semi-supervised model achieves higher precision and recall in detecting abusive entities as compared to unsupervised techniques. Finally, we show that our proposed stochastic partial natural gradient inference for our model empirically achieves faster convergence than stochastic gradient and Online-EM with sufficient statistics.
BibTeX
@inproceedings{yelundur2019detection,
  author = {Anil R. Yelundur and Vineet Chaoji and Bamdev Mishra},
  title = {{Detection of Review Abuse via Semi-Supervised Binary Multi-Target Tensor Decomposition}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2019},
  url = {https://openreview.net/forum?id=0FzswO-Bqj},
}
Figure