PaperlayerView exact revision on arXiv

Source-linked AI summary

ScoreGAN: A Fraud Review Detector based on Multi Task Learning of Regulated GAN with Data Augmentation

Saeedreza Shehnepoor, Roberto Togneri, Wei Liu, Mohammed Bennamoun

arXiv:2006.06561v2cs.LGcs.CRstat.ML

TL;DR

Fraud-review detectors face limited trusted labels, while existing GAN approaches do not incorporate review metadata and may miss bot-generated reviews. ScoreGAN conditions generated reviews on rating scores and uses text–score representations in detection. It reports improved detection over state-of-the-art systems on Yelp and TripAdvisor.

  • Problem

    Fraud-review detection is constrained by scarce trusted labels, and prior GAN approaches omit metadata and overlook possible bot-generated reviews.

    Source: §I · pp. 1–2

  • Method

    ScoreGAN incorporates rating scores through Information Gain Maximization to generate score-correlated reviews and train detection using joint text–score representations.

    Source: §I · p. 2, B. Information Gain Maximization GAN… · p. 4

  • Results

    7% on Yelp and 5% on TripAdvisor: score-customized reviews improved fraud-review detection over state-of-the-art systems.

    Source: §I · p. 2

  • Takeaways & Limitations

    Score-customized synthetic reviews offer an approach to address the lack of high-quality data for fraud detection.

    Source: §I · p. 2

Abstract

from arXiv · show

The promising performance of Deep Neural Networks (DNNs) in text classification, has attracted researchers to use them for fraud review detection. However, the lack of trusted labeled data has limited the performance of the current solutions in detecting fraud reviews. The Generative Adversarial Network (GAN) as a semi-supervised method has demonstrated to be effective for data augmentation purposes. The state-of-the-art solutions utilize GANs to overcome the data scarcity problem. However, they fail to incorporate the behavioral clues in fraud generation. Additionally, state-of-the-art approaches overlook the possible bot-generated reviews in the dataset. Finally, they also suffer from a common limitation in scalability and stability of the GAN, slowing down the training procedure. In this work, we propose ScoreGAN for fraud review detection that makes use of both review text and review rating scores in the generation and detection process. Scores are incorporated through Information Gain Maximization (IGM) into the loss function for three reasons. One is to generate score-correlated reviews based on the scores given to the generator. Second, the generated reviews are employed to train the discriminator, so the discriminator can correctly label the possible bot-generated reviews through joint representations learned from the concatenation of GLobal Vector for Word representation (GLoVe) extracted from the text and the score. Finally, it can be used to improve the stability and scalability of the GAN. Results show that the proposed framework outperformed the existing state-of-the-art framework, namely FakeGAN, in terms of AP by 7\%, and 5\% on the Yelp and TripAdvisor datasets, respectively.

I. INTRODUCTION

ScoreGAN addresses gaps in fraud-review detection by conditioning generated reviews on rating scores and using text–score representations to detect fraud, including bot-generated reviews. The paper also examines whether regularization improves GAN convergence and whether augmentation helps with data scarcity.

I. INTRODUCTION

  • Prior GAN-based fraud-review generation did not associate generated text with metadata, while bot reviews can evade metadata-based behavioral cues.The paper identifies rating scores as a useful metadata feature for joint representations.Source: §I · pp. 1–2
  • ScoreGAN uses Information Gain Maximization to generate reviews correlated with supplied scores and uses text–score representations in detection.The score-probability auxiliary function is included in the GAN objective to encourage score-correlated generation.Source: §I · p. 2
  • 7% on Yelp and 5% on TripAdvisor: score-customized generated reviews improved fraud-review detection over state-of-the-art systems.The comparison is attributed to customization of generated reviews based on score.Source: §I · p. 2
  • The paper reports that regularization reduces the epochs required for ScoreGAN convergence and that generated labeled data helps address data scarcity.Its scalability experiments examine whether smaller training subsets can reach full-dataset performance on Yelp and TripAdvisor.Source: §I · p. 2

II. RELATED WORKS

The related work covers classical fraud-detection features, deep-learning classifiers, and GAN-based review generation. It describes their feature choices and the use of GANs to augment review data.

A. Fraud Detection

  • Fraud detection approaches are divided into classical and deep-learning methods.Classical methods use hand-crafted features, while deep-learning methods map detection to text classification.Source: A. Fraud Detection · p. 2

1) Classical Approaches:

  • Classical approaches combine text cues, such as n-grams and capitalization, with behavioral signals such as review frequency and score patterns.NetSpam combines text and behavioral features to rank reviews by fraud probability.Source: 1) Classical Approaches · pp. 2–3

2) Deep Learning:

  • Deep-learning work represents reviews with neural models or engineered topic features, then classifies them using models such as recurrent networks, SVMs, or MLPs.Reported examples include a 3% fraud-classification improvement on TripAdvisor and 81.3% accuracy for an MLP on Yelp.Source: 2) Deep Learning · p. 3

B. Generative Adversarial Networks

  • GAN-based approaches generate review samples for augmentation; FakeGAN uses two discriminators, including one distinguishing human-written from generated fraud reviews.FakeGAN uses its dual discriminators to address mode collapse and was evaluated on 800 TripAdvisor reviews.Source: B. Generative Adversarial Networks · p. 3

III. PROPOSED METHODOLOGY

ScoreGAN generates fraud bot reviews correlated with scores and trains a discriminator to distinguish fraud from genuine reviews. Its regularization maximizes mutual information between a generation constraint and generated text.

A. Problem Definition

  • Given genuine and human-written fraud reviews with scores, ScoreGAN aims to generate score-correlated fraud bot reviews and train discriminators for fraud-type and fraud-versus-genuine classification.The system first trains Df to distinguish human-written from generated fraud and estimate score probability, then uses the generated reviews to train Dg.Source: A. Problem Definition · p. 3

B. Information Gain Maximization GAN Regularization

  • The regularizer maximizes mutual information between constraint c and generated review Gθ(z,c), using a variational lower bound and auxiliary distribution Q(c|x).A fully connected layer on Df estimates score probability Qs, and the objective preserves the constraint's contribution during generation.Source: B. Information Gain Maximization GAN… · p. 4

1) Generator (Module 1):

The generator conditions review generation on a rating score, while two discriminators distinguish fraud from genuine reviews and support score-correlated generation. The discriminator representations combine text-derived features with the review score.

1) Generator (Module 1):

  • The generator takes random noise z and score s as constraints and generates fake bot reviews Xfb.It uses an LSTM with embedding size and hidden dimension 32, trained with batch size 50.Source: 1) Generator (Module 1) · p. 4

2) Discriminators (Module 2, 3):

  • Dg classifies reviews as fraud or genuine, while Df distinguishes real from generated fraud reviews and predicts review scores.Df uses softmax outputs P and Q; Q predicts scores from 1 to 5 as the auxiliary function.Source: 2) Discriminators (Module 2, 3) · p. 5, Figure 3 · p. 5
  • The review’s kernel-derived representation is concatenated with score s to form vector m; max-pooling combines kernel outputs into f, and CNN feature representations are fed to a softmax layer for final classification.Source: 2) Discriminators (Module 2, 3) · p. 5

D. Training

ScoreGAN training has a pre-training step followed by training. The training step uses Monte-Carlo search to address discrete word generation.

D. Training

  • Training has two steps: pre-training generates initial reviews for subsequent Df training, followed by training.The initial reviews are used with Xfh and Xfb.Source: D. Training · p. 5
  • During training, Monte-Carlo search addresses the generator’s discrete token-generation problem.Source: D. Training · p. 5
  • Monte-Carlo search selects words by evaluating complete sequences through repeated policy-based rollouts and their rewards.The algorithm plays to the end a fixed number of times from each state and selects the best move.Source: D. Training · p. 5

B. Experimental Setup

The section specifies ScoreGAN’s training configuration and evaluates Dg with AP, AUC, and accuracy across several experimental factors.

B. Experimental Setup

  • The setup uses 50-dimensional GLoVe, batch size 64, CNN filters of widths 1–20, and discriminator/generator learning rates of 10^-4/10^-20.Adversarial training uses 120 iterations, with five generator inner-loop iterations and three discriminator training epochs.Source: B. Experimental Setup · pp. 6–7
  • Table I is designated to report dataset details.Source: Table I · p. 7

C. Main Results

  • Evaluation uses AP, AUC, and accuracy to examine score, behavioral-feature, and regularization effects, as well as robustness to data scarcity.Source: C. Main Results · p. 7

1) Performance:

ScoreGAN outperforms other systems on both datasets, while feature and score ablations show that results depend on the feature combination and dataset. Regularization also shortens convergence relative to the compared GAN approaches.

1) Performance:

  • ScoreGAN outperforms other systems on both datasets across AP, AUC, and accuracy; generated reviews improve Dg performance over human-fraud-only training.The cited table is scoped to Yelp-Main with a 70/30 train/test split.Source: 1) Performance · p. 7, Table II · p. 8, Figure 4 · p. 8

2) Behavioral Feature Performance on ScoreGAN:

3) Ablative Study:

  • On Yelp, score-conditioned generation improves more than discriminator scoring; on TripAdvisor, effects vary by metric, with binary labels offered as a possible constraint.Source: 3) Ablative Study · pp. 7–8, Figure 7 · p. 9
  • With regularization, ScoreGAN converges to final performance in 100 adversarial iterations, versus FakeGAN’s 140; unregularized ScoreGAN had not reached its best after 200.Source: 3) Ablative Study · p. 9

4) Robustness with Data Scarcity:

ScoreGAN is reported to retain performance as supervision and training-set size vary, with cross-dataset results showing smaller performance changes than for other frameworks. The section also compares how supervised and semi-supervised approaches respond to reduced training data.

4) Robustness with Data Scarcity:

  • ScoreGAN performs best across the tested supervision levels, while the two supervised approaches decline sharply with less training data and semi-supervised approaches are less affected.The tested training proportions are 0.7, 0.5, 0.3, and 0.1; ScoreGAN’s Yelp performance converges to a constant value at 30% supervision.Source: 4) Robustness with Data Scarcity · p. 9, Figure 8 · p. 10, Figure 9 · p. 10
  • On TripAdvisor, ScoreGAN’s best AP occurs at 0.7 supervision, while its AP at 0.5 remains the same; AP varies more than AUC or accuracy.The authors attribute AP’s greater variation to its relatively low values, where small changes in training data produce larger performance changes.Source: 4) Robustness with Data Scarcity · p. 9
  • With fewer training reviews, robustness declines across frameworks, but ScoreGAN exhibits imperceptible performance changes compared with the others.In the cross-dataset experiment, the frameworks are compared using the same number of samples from each dataset.Source: 4) Robustness with Data Scarcity · p. 10
  • With more TripAdvisor samples, FakeGAN degrades on AP, AUC, and accuracy while ScoreGAN improves on all three; ScoreGAN changes little across Yelp sample sizes.On Yelp, ScoreGAN’s AP decreases slightly at 1,000 samples, while its AUC and accuracy improve; Yelp has substantially more genuine than fraud reviews.Source: 4) Robustness with Data Scarcity · pp. 10–11, Figure 10 · p. 11

V. CONCLUSION

ScoreGAN uses information gain maximization between review scores and generated score-correlated text in a regulated GAN with one generator and dual discriminators. The conclusion reports Yelp AP of 65.16% and AUC of 88.78%, and identifies combining text with other performance-related features as future work.

V. CONCLUSION

  • ScoreGAN is a regulated GAN with one generator and dual discriminators that uses review text and metadata such as scores.Its loss uses information gain maximization between scores and generated score-correlated reviews.Source: §V · p. 11
  • Yelp results were 65.16% AP and 88.78% AUC, described as significant improvements over FakeGAN, NetSpam, Li et al., and Shaalan et al..Source: §V · p. 11
  • Future work will consider combining text features with other features that affect performance, potentially supporting joint representations of text and metadata.Source: §V · p. 11

Related papers

Loading 2006.06561v2…