Detecting malicious “professional” online reviews using machine learning

[*]

A new research collaboration between China and the United States offers a way to detect malicious e-commerce reviews designed to undermine competitors or facilitate blackmail, by leveraging the signature behavior of such reviews.

The system, called malicious user detection model (MMD), uses metric learning, a technique commonly used in computer vision and recommendation systems, along with a recurrent neural network (RNN), to identify and label the output of these reviewers, which the article appointed Malicious business users (PMU).

Awesome! 1 star

Most online e-commerce reviews provide two forms of user feedback: a star rating (or a rating out of 10) and a text review, and in a typical case these will logically match (i.e. say that a bad review will be accompanied by a low rating).

PMUs, however, usually reverse this logic, either leaving a bad text review with a high rating, or a bad rating accompanied by a good review.

This allows the user’s review to damage their reputation without triggering the relatively simple filters deployed by e-commerce sites to identify and process negative malicious review results. If a filter based on natural language processing (NLP) identifies the invective in the text of a review, this “flag” is effectively overridden by the high (or decimal) rating that the PMU has also assigned, thus making the “neutral” malicious content. , from a statistical point of view.

An example of how a malicious review can be mixed, statistically, with genuine reviews, from the perspective of a collaborative filtering system attempting to identify such behavior. Source: https://arxiv.org/pdf/2205.09673.pdf

The new document notes that the intent of a PMU is often to extort money from online retailers in exchange for changing negative reviews and/or promising to stop posting negative reviews. In some cases, the actors are ad hoc individuals looking for discounts, although the PMU is frequently used by the victim’s competitors.

Hiding negative reviews

The current generation of automated detectors for such reviews use collaborative filtering or a content-based model, and look for clear and unambiguous “outliers” – reviews that are uniformly negative in both feedback methods, and that diverge notably from the general tendency of the revision. sentiment and rating.

The other classic signature that these filters rely on is high posting frequency, whereas a PMU will post strategically and only occasionally (since each review can represent either an individual commission or a step in a larger strategy). long designed to obscure the “frequency” metric).

Therefore, the researchers of the new paper have integrated the strange polarity of professional malicious reviews into a dedicated system, resulting in an algorithm that is almost comparable to a human reviewer’s ability to “smell a rat” to the disparity between rating and review. text content.

The conceptual architecture of MMD, composed of two core modules: Malicious User Profiling (MUP) and Attention Metric Learning (MLC, in grey).

The conceptual architecture of MMD, composed of two core modules: Malicious User Profiling (MUP) and Attention Metric Learning (MLC, in grey).

Comparison with previous approaches

Since MMD is, according to the authors, the first system to attempt to identify UGPs based on their schizophrenic posting style, there is no direct prior work to compare it to. Therefore, the researchers pitted their system against a number of component algorithms that traditional automated filters frequently depend on, including K-means++ clustering; the venerable Statistic Outlier Detection (SOD); Hysades; Semi-sad; CNN-sad; and Slanderous User Detection Recommendation System (SDRS).

Tested against labeled datasets from Amazon and Yelp, MMD is able to identify professional online detractors with the highest accuracy rate, the authors claim.  The bold represents MMD, while the asterisk

indicates the best performance. In the above case, MMD was only beaten in two tasks, by an autonomous technology (MUP) that is already built into it, but not tooled by default for the task at hand.

Tested against labeled datasets from Amazon and Yelp, MMD is able to identify professional online detractors with the highest accuracy rate, the authors claim.  The bold represents MMD, while the asterisk

indicates the best performance. In the above case, MMD was only beaten in two tasks, by an autonomous technology (MUP) that is already built into it, but not tooled by default for the task at hand.

In this case, MMD was pitted against unlabeled datasets from Taobao and Jindong, making it an unsupervised learning task. Again, MMD is only enhanced by one of its own constituent technologies, highly suited to the task for testing purposes.

In this case, MMD was pitted against unlabeled datasets from Taobao and Jindong, making it an unsupervised learning task. Again, MMD is only enhanced by one of its own constituent technologies, highly suited to the task for testing purposes.[On] The researchers observe:

all four datasets, our proposed model MMD (MLC+MUP) outperforms all baselines in terms of F-score. Note that MMD is a combination of MLC and MUP, which ensures its superiority over supervised and unsupervised models in general.

The article also suggests that MMD could serve as a useful pre-processing method for traditional automated filtering systems and provides experimental results on a number of datasets including User-Based Collaborative Filtering (UBCF), Item-Based Collaborative Filtering (IBCF), Matrix Factorization (MF-eALS), Bayesian Custom Ranking (MF-BPR) and Neural Collaborative Filtering (NCF).

In terms of Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG) in the results of these tested increases, the authors state: [feedback]“Among the four datasets, MMD significantly improves the recommendation models in terms of HR and NDCG. Specifically, MMD can improve HR performance by an average of 28.7% and HDCG by an average of 17.3%. [intuitive]“By removing malicious business users, MMD can improve the quality of datasets. Without those fake malicious business users

the dataset becomes more .’The paper is titled

Detect malicious business users with metric learning in the recommender system

s, and comes from researchers from the Department of Computer Science and Technology of Jilin University; the Intelligent Information Processing Key Laboratory of the Chinese Academy of Sciences in Beijing; and the Rutgers School of Business in New Jersey.

Data and approach

The detection of UGPs is a multimodal challenge, since two non-equivalent parameters (a numeric-valued star/decimal notation and a textual review) must be taken into account.  The authors of the new paper say that no previous work has addressed this challenge.

MMD uses a Hierarchical Dual Attention Recurrent Neural Network (HDAN) to assimilate review content into a sentiment score.

Projecting an opinion into a sentiment score with HDAN, which helps with word integration and sentence integration to get a sentiment score. Projecting an opinion into a sentiment score with HDAN, which helps with word integration and sentence integration to get a sentiment score. HDAN uses attention mechanisms to assign weights to each word and sentence. In the image above, the authors state, the word

poorer

should clearly be given more weight than competing words in the review.

For the project, HDAN took the product ratings on four datasets as ground truth. The datasets were Amazon.com; Yelp for RecSys (2013); and two “real-world” (rather than experimental) datasets, from Taobao and Jindong.

MMD leverages Metric Learning, which attempts to estimate an accurate distance between features to characterize the overall group of relationships in the data. MMD starts with hot encoding to select the user and item, through a latent factor model (LFM), which gets a base evaluation score. In the meantime, HDAN projects the content of the review into the sentiment score as supporting data. The results are then processed into a malicious user profiling (MUP) model, which produces the

feeling gap vector

– the difference between the rating and the estimated sentiment score of the textual content of the review. Thus, for the first time, UGPs can be categorized and labeled.

Attention-based metric learning for clustering.

Attention-based metric learning for clustering.

Metric Learning for Clustering (MLC) uses these output labels to establish a metric against which the likelihood of a user review being malicious is calculated. Human tests In addition to the quantitative results detailed above, the researchers conducted a user study that tasked 20 college students with identifying malicious reviews, based solely on content and star ratings. Participants were asked to rate the comments as 0 (for “normal” reviewers) or

1

(for a malicious professional user). [*] On a 50/50 split between normal and malicious reviews, students tagged an average of 24 true positives and 24 true negatives.  In comparison, MMD was able to label an average of 23 true positive and 24 true negative users, operating almost at the level of human discernment and exceeding the baselines for the task.

Students against MMD. Asterisk [*] indicates best results and bold indicates MMD results.

Students against MMD. Asterisk

indicates best results and bold indicates MMD results.

The authors conclude:

“Essentially, MMD is a generic solution, which can not only detect the professional malicious users that are explored in this document, but also serve as a general basis for malicious user detections. With more data, such as image, video or sound, MMD’s idea can be instructive to detect the sentiment gap between their title and content, which has a bright future to counter different masking strategies in different applications.

First published May 20, 2022.

Sherry J. Basler