💻 computer science

Fake Review Detection on E-Commerce Platforms Using a Hybrid Anomaly Detection Approach: Combining Autoencoder and Isolation Forest

This paper proposes a hybrid unsupervised anomaly detection framework combining Autoencoders and Isolation Forests to effectively identify fake e-commerce reviews using TF-IDF and BERT features, offering a scalable and cost-effective alternative to supervised methods by eliminating the need for labeled training data.

Original authors: Steven Laychi, Samuel Paul Arthur Karuntu, Daniel Hamonangan Sihombing, Rhio Sutoyo, Gabriel Asael Tarigan

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Steven Laychi, Samuel Paul Arthur Karuntu, Daniel Hamonangan Sihombing, Rhio Sutoyo, Gabriel Asael Tarigan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a massive online store, like a digital mall. You want to buy a new coffee maker, so you look at the reviews. But here's the problem: some people are paid to write fake, glowing reviews to trick you, while others use robots (AI) to write thousands of them instantly. This is "opinion spam," and it makes the whole store feel untrustworthy.

This paper is about building a smart security guard for these online stores. The goal is to spot the fake reviews without needing a list of "known bad guys" to start with.

Here is how the authors built their solution, explained simply:

The Problem with Current Guards

Most current security guards are like bouncers with a "Wanted" poster. They only know who to kick out if they have seen that specific person before or have a photo of them. In the world of data, this means they need a huge list of reviews that humans have already labeled as "fake."

The Issue: Getting these lists is expensive, slow, and only works for one specific store. If a new type of fake review appears, the bouncer doesn't know what to do.

The New Approach: The "Hybrid" Security Team

The authors created a new system that doesn't need a "Wanted" poster. Instead, it learns what a "normal" review looks like and flags anything that feels "off." They used two different types of guards working together:

1. The Copycat (Autoencoder)

Imagine a student trying to copy a teacher's handwriting perfectly.

How it works: This system is trained only on real, honest reviews. It tries to "reconstruct" or rewrite every review it sees.
The Trick: If it sees a real review, it can copy it easily. If it sees a fake review (even a smart AI one), it struggles to copy it because the pattern is slightly different.
The Score: The harder it is for the system to copy the review, the higher the "suspicion score." It's like saying, "I couldn't copy this handwriting, so it must be a forgery."

2. The Crowd Spotter (Isolation Forest)

Imagine a crowded party where everyone is wearing a red shirt.

How it works: This system looks at the whole room. It knows that most people (real reviews) are wearing red. It randomly picks groups of people to separate.
The Trick: If someone is wearing a neon green shirt (a fake review), they are easy to spot and separate from the crowd immediately. They stand out because they are in a "sparse" area.
The Score: If a review is easy to isolate from the crowd, it gets a high suspicion score.

Putting Them Together (The Hybrid)

The authors realized that sometimes the Copycat misses a fake review that looks very similar to a real one, and sometimes the Crowd Spotter gets confused by a weirdly written real review.

The Solution: They combined the two. They took the "suspicion score" from the Copycat and the "suspicion score" from the Crowd Spotter and mixed them together.
The Result: If either guard thinks something is fishy, the system flags it. This makes the security much tighter than using just one guard.

The Tools They Used

To make these guards smarter, they gave them two different "languages" to understand the reviews:

TF-IDF (The Word Counter): This is like counting how many times words like "great" or "amazing" appear. It's simple but misses the meaning behind the words.
BERT (The Context Reader): This is a more advanced AI that understands the context. It knows that "This coffee maker is a bomb" could mean it's great (slang) or dangerous, depending on the other words.

Finding: The "Context Reader" (BERT) was much better at spotting fakes because it understood the story of the review, not just the word count.

The Results

The authors tested this system on a dataset of 40,000 reviews (half real, half fake AI-generated ones).

The Winner: The Hybrid Team (Copycat + Crowd Spotter) using the Context Reader (BERT) was the best unsupervised method. It caught fake reviews with an accuracy (F1-Score) of 0.84.
The Comparison:
- It beat the "Copycat" alone and the "Crowd Spotter" alone.
- It was slightly worse than a "Supervised" guard (one with a "Wanted" poster), which scored 0.89.
- The Big Win: The Hybrid Team achieved nearly the same accuracy as the expensive "Wanted Poster" guard, but it did it without needing any labeled data. It learned on its own.

What They Didn't Claim

The paper is very specific about what they did and didn't do:

They did not test this on real-time live traffic from stores like Tokopedia or Shopee yet; they used a specific dataset of Amazon reviews.
They did not claim this works for languages other than English (like Bahasa Indonesia) yet, because their "Context Reader" was trained on English.
They did not test it against the newest, most advanced AI (like GPT-4), only against GPT-2 generated reviews.

The Bottom Line

The paper proves that you can build a very effective "fake review detector" without spending money on labeling thousands of reviews. By combining a system that learns to copy normal text with a system that spots outliers, you get a strong, cheap, and adaptable security guard for e-commerce.

Technical Summary: Hybrid Anomaly Detection for Fake Review Detection

Problem Statement
E-commerce platforms face a significant challenge from "opinion spam" or fake reviews, which distort product ratings and erode consumer trust. While high-performance detection methods exist, they predominantly rely on supervised learning, necessitating large, labeled datasets that are expensive to collect, platform-specific, and difficult to maintain as fraud tactics evolve. Unsupervised approaches offer a cost-effective alternative by eliminating the need for labeled training data, yet existing unsupervised methods often lack the robustness of supervised baselines. Furthermore, hybrid approaches combining reconstruction-based and isolation-based anomaly detection have been successful in network security but remain underexplored in the context of textual fake review detection.

Methodology
The authors propose a hybrid unsupervised anomaly detection framework that integrates Autoencoders (AE) and Isolation Forests (IF). The system operates through the following pipeline:

Dataset & Training Strategy: The study utilizes the Amazon Fake Reviews Dataset, comprising 20,000 real human reviews and 20,000 fake reviews generated by GPT-2. Crucially, the model is trained unsupervised, using only the real reviews to learn the distribution of genuine data. Labels are reserved exclusively for the evaluation phase.
Feature Extraction: Two distinct feature representations are evaluated to determine the impact of semantic understanding:
- TF-IDF: Bag-of-words vectors (unigrams and bigrams) limited to 10,000 terms.
- BERT Embeddings: Contextual sentence embeddings (768-dimensional) derived from the bert-base-uncased model via mean-pooling, without fine-tuning.
Hybrid Model Architecture:
- Autoencoder: A neural network trained to reconstruct real review embeddings. It calculates an anomaly score ( $s_{AE}$ ) based on the Mean Squared Error (MSE) of the reconstruction. High error indicates a deviation from the "normal" manifold of real reviews.
- Isolation Forest: An ensemble of trees that isolates data points by random partitioning. It calculates an anomaly score ( $s_{IF}$ ) based on the path length required to isolate a review; shorter paths indicate anomalies located in sparse regions of the feature space.
- Score Fusion: The final decision is made using a weighted linear combination: $s_{hybrid} = \alpha \cdot s_{AE} + (1 - \alpha) \cdot s_{IF}$ . The weights ( $\alpha$ ) and classification threshold are optimized on a validation set to maximize the F1-Score.
Baselines: The hybrid approach is compared against standalone Autoencoders, standalone Isolation Forests, and a supervised Support Vector Machine (SVM) trained on labeled data.

Key Results
Experiments were conducted on a local workstation using Python, PyTorch, and Scikit-learn. The results indicate:

Hybrid Superiority: The hybrid framework consistently outperformed both standalone unsupervised models across all metrics (Precision, Recall, F1-Score, AUC-ROC).
Feature Impact: BERT embeddings significantly outperformed TF-IDF across all model configurations, yielding a performance gain of approximately 5–8 percentage points. This suggests that semantic context is critical for distinguishing GPT-2 generated text from human writing.
Specific Performance (BERT Features):
- Hybrid AE+IF: Achieved an F1-Score of 0.84 and an AUC-ROC of 0.89.
- Standalone AE: F1-Score of 0.74.
- Standalone IF: F1-Score of 0.75.
- Supervised SVM (Upper Bound): Achieved an F1-Score of 0.89 and AUC-ROC of 0.93.
Comparison: While the supervised SVM achieved the highest performance (as expected due to access to labels), the hybrid unsupervised method achieved competitive results (0.84 F1) without requiring any labeled training data.

Significance and Claims
The paper claims that the proposed hybrid framework offers a scalable and cost-effective substitute for traditional supervised detection systems in real-world e-commerce environments. Its primary significance lies in its ability to deliver competitive detection performance (approaching supervised baselines) while completely removing the dependency on expensive, platform-specific labeled datasets.

The authors position this work as a practical solution for platforms where collecting labeled fake review data is prohibitive. They acknowledge that while the method does not yet surpass supervised learning, it bridges the gap between purely unsupervised methods and supervised performance. The study also highlights the importance of semantic feature representation (BERT) over traditional frequency-based methods (TF-IDF) for detecting AI-generated spam.

Limitations and Future Directions
The authors modestly note several limitations and future directions:

Dataset Scope: The evaluation relies on GPT-2 generated fakes; performance against more advanced models (e.g., GPT-4) or manually written fakes remains to be tested.
Language Constraints: The current implementation is English-centric. The authors suggest extending the framework to support Southeast Asian languages (e.g., Bahasa Indonesia) for platforms like Tokopedia and Shopee.
Feature Expansion: Future work could integrate reviewer behavioral features and graph-based methods to model reviewer-product relationships, which are currently excluded due to dataset limitations.
Hyperparameter Sensitivity: The contamination parameter in the Isolation Forest and the bottleneck dimensionality in the Autoencoder require careful tuning and may not generalize perfectly across datasets with different fake-to-real ratios.