3.1.1. Machine learning techniques

Machine learning algorithms are computational models that learn patterns and relationships from data without being explicitly programmed. They use statistical techniques to automatically detect patterns, make predictions, or make decisions based on the input data. Some commonly used machine learning algorithms include Decision Trees, Random Forest (RF), Support Vector Machines (SVM), Naïve Bayes (NB), K-Nearest Neighbours (KNN), Logistic Regression (LR), etc. Researchers frequently employ these supervised, semi-supervised, and unsupervised machine learning techniques for fake review detection.

Supervised ML Algorithms

Supervised ML algorithms like SVM, KNN, and LR are among the most commonly used methods in the field. These ML algorithms have been used by Tufail et al. (Reference Tufail, Ashraf, Alsubhi and Aljahdali2022) to detect fake reviews on the Yelp hotel review dataset. Their model proved to be a robust one, but it only focused on supervised models. Similarly, Kumaran et al. (Reference Kumaran, Chowdhary and Sreekavya2021) have used naïve Bayes, Eristic regression, and SVM for detecting fake reviews in a dictionary based on social media keywords and online reviews. Their model uses a very limited feature set of uni-, bigrams, and length of review. Poonguzhali et al. (Reference Poonguzhali, Sowmiya, Surendar and Vasikaran2022) have implemented SVM for fake review detection by creating an online e-commerce interface. Fake reviews are predicted in this system and are not used in the database while recommending a product on this interface. But their method can’t upload more than one review by one user. So, their prevention method is not feasible in large databases.

Along similar lines, Hussain et al. (Reference Hussain, Mirza, Hussain, Iqbal and Memon2020) have given some important behavioural and linguistic pointers to identify fake reviews. For example, the absence of profile data of the reviewer, posting duplicate reviews, short and often grammatically erroneous reviews, groups of reviews with the same timestamp, and excessive use of positive or negative words. However, they have studied the behavioural model and linguistic model separately. The chances of reviews being classified accurately decrease when only one model is used. Hence, Alsubari et al. (Reference Alsubari, Deshmukh, Alqarni, Alsharif, Aldhyani, Alsaade and Khalaf2022) have extracted features such as sentiment score, four grams, number of verbs, nouns, and strong positive or negative words to identify fake reviews. Then they applied four different supervised classifiers—SVM, Naïve Bayes, Random Forest, and Adaptive Boost and compared their accuracy for fraudulent reviews detection. The limitation of this study was fewer extracted features and, their dataset was limited to the hotel domain. According to the findings of Abdulqader et al. (Reference Abdulqader, Namoun and Alsaawy2022), non-verbal features carry greater significance than verbal features, and their combination can enhance the accuracy of detection. However, they have only given a pure theory-based model, which may or may not apply to datasets other than the one used here. Alawadh et al. (Reference Alawadh, Alabrah, Meraj and Rauf2023a) experimented with a benchmark, balanced hotel review dataset and proved that real-time application of deep learning-based, semantically aware text features on web portals can effectively detect fraudulent reviews. However, they have only used a small dataset which doesn’t fully utilize the neural network advantage. A novel ML framework based on M-SMOTE has been created by Kumar et al. (Reference Kumar, Gopal, Shankar and Tan2022) to address the class imbalance problem. The study’s results confirm that combining reviewer-centric features with review-centric features significantly improves the performance of FRD models. A major limitation of this study is that the features extracted from the datasets used in this study, may not be present in other domains. Theuerkauf and Peters (Reference Theuerkauf and Peters2023) employ labelled reviews sourced from the iOS App Store and combine them with two statistical approaches. The simultaneous utilization of multiple feature sets is demonstrated to enhance the detection of fraudulent reviews, but undetected false positives might affect the evaluation metrics.

Semi-supervised ML Algorithms

The supervised machine algorithms use extensive labelling, which is laborious as well as subjective. Hence, the semi-supervised approach uses a pre-defined set of features to train classifiers. In their research, Jing-Yu et al. (Reference Jing-Yu and Ya-Jun2022) employed a semi-supervised approach called AspamGAN (Generative Adversarial Network), which utilizes an attention mechanism in the classifier to detect fake reviews. They have used the data from the TripAdvisor dataset. But, if the data generated by their model is insufficient, it may result in poor accuracy. The earlier version of this method spamGAN given by Stanton and Irissappane (Reference Stanton and Irissappane2020) was a powerful tool that used a simple classifier. But it has the disadvantage that it may lose important information such as context, the focus of the sentence, etc., and thus won’t be able to detect fake information with high probability. AspamGAN has the advantage that it has better performance with limited label data, than SpamGAN.

The semi-supervised approach relies on a pre-defined set of features to train classifiers. Due to the laborious labelling task of the fake review datasets, researchers such as Ligthart et al. (Reference Ligthart, Catal and Tekinerdogan2021) used four semi-supervised techniques in their study. This consisted of self-training (training on the labelled portion of the data), co-training (utilizing additional perspectives of data), Transductive SVM (variation of traditional SVM used in a semi-supervised setting), and label propagation plus spreading (again used for semi-supervised training). They train on one dataset and perform experiments on two more datasets from Yelp. Although their effort alleviates the labelling task, they have not considered metadata or reviewer-based features. In contrast, Wu et al. (Reference Wu, Liu, Wu and Tan2022) implement a semi-supervised probabilistic ensemble that collectively captures the individual behavioural characteristics of reviewers as well as the reviewer network. They use ten behavioural features such as the number of reviews in a day, burstiness, popularity of product that the user has reviewed, average distance between a user and other users, etc. They assume that the reviewer network presents homophily.

Unsupervised ML Algorithms

As mentioned earlier, detecting fake reviews accurately is highly improbable by humans. Hence, the availability of labelled data is less. Unsupervised algorithms have a good scope here. The work done by Mothukuri et al. (Reference Mothukuri, Aasritha, Maramella, Pokala and Perumalla2022) creates clusters using the extracted features. They perform K-means clustering, GMM Full covariance clustering, and GMM Diagonal covariance unsupervised techniques to detect fake reviews taken from the café dataset of Yelp. They found that K-means shows the highest accuracy among the three. Many other unsupervised algorithms could have been explored in the aforementioned work, and different domains.

There is a major problem of concept drift (Mohawesh et al., Reference Mohawesh, Tran, Ollington and Xu2021; Tommasel & Godoy, Reference Tommasel and Godoy2019) while using machine learning techniques. Concept drift refers to the adaptation of fake reviewers over some time. They start writing in such a way that their writing skills are similar to real reviews and hence, can’t be detected by the detection algorithms in place. For this, Wang et al. (Reference Wang, Fong and Law2022) have incorporated the temporal patterns of reviews. They added a degree of suspicion for FRD by analysing 3D time series, the number of reviews, and the information entropy. Their framework seems comprehensive, but they have only used supervised algorithms. One major drawback of machine learning algorithms is that they require the extraction of features manually and a huge utilization of computational resources (Kaddoura et al., Reference Kaddoura, Chandrasekaran, Popescu and Duraisamy2022).

Many supervised, semi-supervised and unsupervised algorithms have been used to detect fake reviews, to date. These traditional machine learning algorithms perform adequately on small datasets and are highly valued by researchers. In addition to this, they are simpler to implement and computationally cheap but, their performance deteriorates on larger datasets as compared to deep learning models. Also, their solution doesn’t sustain over time and the problem of concept drift arises. In addition to the above shortcomings, they also require manual feature extraction and huge computational resources. Hence, deep learning algorithms were explored by researchers and their studies are discussed next. Table 3 is a summarized representation of the traditional machine learning algorithms used in FRD research work.

Table 3. Machine learning techniques used for FRD

3.1.2. Deep learning techniques

Several deep learning algorithms are suitable for fraudulent review identification. These algorithms leverage the power of deep neural networks to automatically learn features and structural patterns from review data. Some commonly used deep learning algorithms for FRD are Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Transformer-based Models, Deep Belief Networks (DBN) and Siamese Neural Networks. The researchers Alsubari et al. (Reference Alsubari, Deshmukh, Al-Adhaileh, Alsaade and Aldhyani2021) have used four datasets including hotels, restaurants, Yelp, and Amazon to perform convolutional and max-pooling layers of the CNN to reduce dimensions and extract features. They have done cross-domain analysis, but the performance of this model decreases on a single-domain dataset. Jing-Yu and Ya-Jun (Reference Jing-Yu and Ya-Jun2022) utilized a semi-supervised method called AspamGAN (Generative Adversarial Network) that incorporates an attention mechanism to detect fake reviews in their assembled, partially labelled dataset. Stanton and Irissappane (Reference Stanton and Irissappane2020) previously introduced the SpamGAN model, but AspamGAN addresses some of its limitations and performs better with limited labelled data.

The work done by Basyar et al. (Reference Basyar, Adiwijaya and Murdiyansah2020) built a Long-short-term memory (LSTM) model as well as a Gated Recurrent Unit (GRU) model to detect e-mail spam. The former outperformed the latter, but the result was not significant, and they only used a training dataset due to time constraints. The authors Liu et al. (Reference Liu, Wang, Shi and Li2022) give a layered attention network employing two stratums to capture semantic information. The authors then integrate a convolutional structure and Bi-LSTM to extract crucial semantics resulting in superior performance compared to other algorithms. Their limitation is the use of a supervised algorithm and a smaller number of extracted features. The work done by Crawford et al. (Reference Crawford, Khoshgoftaar, Prusa, Richter and Al Najada2021) uses inductive transfer learning to detect hotel review spam. However, they have limited their work to a single domain. As a different approach, Sadiq et al. (Reference Sadiq, Umer, Ullah, Mirjalili, Rupapara and Nappi2021) train deep learning algorithms to predict the star ratings that will match the review. But they don’t predict whether the review is fake or real. Salminen et al. (Reference Salminen, Kandpal, Kamel, Jung and Jansen2022) fine-tuned the RoBERTa model, which is an optimized model of the BERT transformer, and called it fakeRoBERTa. They also create a fake review dataset with the help language generation model GPT-2. They experiment with fakeRoBERTa on their created dataset and conclude that AI algorithms can outperform humans in detecting fake reviews. However, datasets and techniques need to keep evolving to outmanoeuvre the concept drift problem.

Recently, language generation models such as BERT, and GPT-3 have also been used for the classification of fake reviews. Gambetti and Qiwei (Gambetti & Han, Reference Gambetti and Han2023) used the OpenAI GPT-3 model to generate a fake review dataset and then fine-tuned a pre-trained GPT-Neo model for FRD. They have also correlated fake reviews of a restaurant with customer visits. Shukla et al. (Reference Shukla, Agarwal, Mein and Agarwal2023) have used the latest language transformer model GPT-3, on a novel, annotated dataset of physician reviews. Then they compare the results with LR, XG, RF, and SVM and find that GPT-3 is superior to all in terms of accuracy. However, the major disadvantage of using GPT-3 is that it can’t be fine-tuned to train datasets larger than 10 million characters, and the whole dataset can’t be tested at a go.

It is observed that deep learning algorithms outperform traditional machine learning models when applied to large-scale datasets. However deep learning models are computationally expensive. In addition to the above, deep learning models are susceptible to overfitting and thus, they can’t be used for smaller data. Some transformer-based models such as GPT-3 are uninterpretable. They operate as a black box and have higher computational requirements. Table 4 shows a summary of deep learning models used in FRD so far.

Table 4. Summary of Deep Learning models used for FRD

3.1.3. Graph-based techniques

Graph-based techniques in machine learning refer to approaches that leverage graph structures or networks to represent and analyse data. In these techniques, data elements are represented as nodes, while relationships or connections between the elements are represented as edges or links in the graph. Graph Neural Networks (GNNs), Graph Convolutional Networks (GCNs), and Graph Embeddings such as node2vec, GraphSAGE, and DeepWalk are some commonly used graph-based techniques in machine learning. They can capture complex dependencies and make informed predictions.

The authors Ren et al. (Reference Ren, Yuan and Huang2022) use graph neural networks to find associations between reviews, users, and products. This is based on the premise that there is a dependency between the user and product, the reviewer and the time, ratings, etc. Then they introduce the idea of suspicious values based on the TrustRank (Gyongyi et al., Reference Gyongyi, Garcia-Molina and Pederson2004) method. Their model finds more fake reviews, but the accuracy has not increased significantly. Manaskasemsak et al. (Reference Manaskasemsak, Tantisuwankul and Rungsawang2023) have used two novel graph-partitioning algorithms BeGP and BeGPX for FRD. They use the snowball effect to capture all fraudulent users. In the extended version of BeGPX, they also capture the semantic and emotional content of text.

Although graph-based techniques have several advantages, they suffer from scalability and data sparsity issues. As the datasets grow, the size of the graph and in turn, computational complexity grows exponentially. If the labelled data is sparse, as is in the case of most fake review datasets, it can result in weaker signals in graphs. Further, interpreting the results of graph-based techniques is a challenge in itself. Table 5, given below, summarizes the graph-based models used for fake review detection in the recent past.

Table 5. Summary of graph-based techniques used for FR

3.1.4. Swarm intelligence techniques

Swarm techniques for fake review detection draw inspiration from the behaviour of swarms in nature, where collective intelligence emerges from the interactions of simple individuals. These techniques leverage swarm intelligence principles to detect fake reviews by considering the collective behaviour of reviewers or reviews within a dataset. Ant Colony Optimization (ACO), Particle Swarm Optimization (PSO), Bee Algorithm (BA), and Firefly Algorithm (FA) are a few examples of swarm techniques used in fake review detection. The work done by Shringi et al. (Reference Shringi, Sharma and Suthar2022) uses the Fitness-based Grey Wolf Optimization (FGWOK) technique and k-means clustering to classify fake and authentic reviews. The datasets used are synthetic spam review (from the Database and Information System Lab, University of Illinois, TripAdvisor dataset), movie review dataset from IMDB, and Yelp review dataset. After comparing their results with various metaheuristic clustering methods, such as genetic algorithm (GA), particle swarm optimization (PSO), cuckoo search (CS), and artificial bee colony clustering, they found that their algorithm performs better than the current ways. However, they have not considered the feature interactions and could have used better optimizers. Similarly, Jacob and Rajendran (Reference Jacob and Rajendran2022) give a Fuzzy Artificial Bee colony-based CNN-LSTM approach for fake review classification. After data preprocessing, they use the chi-squared technique for feature extraction and selection and CNN-LSTM-FABC is applied, which is found to outperform the earlier approaches. Here also, contextual features are not considered, which may give higher accuracy when combined with the extracted features. Previously, Saini et al. (Reference Saini, Shringi, Sharma and Sharma2021) used k-means artificial bee colony for feature selection after data preprocessing. Then they optimized clusters using ABC with k-means on three datasets namely Synthetic Spam (containing 478 total reviews), Yelp (containing 4952 reviews), and Movie (a subset of IMDB containing 8544 reviews). Earlier Pandey and Rajpoot (Reference Pandey and Rajpoot2019) used a combination of cuckoo search and Fermat spiral to identify spam reviews. It is compared to six metaheuristics clustering methods and was found to be better than these six. But still, the accuracy could have been improved further.

Thus, swarm intelligence approaches give an optimized result for the fake review detection problem, but this area has not been explored fully to date. Many bio-inspired techniques can be utilized and investigated. Table 6 summarizes the few swarm techniques employed for FRD so far.

Table 6. Summary of swarm techniques for FRD

3.1.5. Other techniques

In addition to machine learning techniques, several other approaches and techniques can be used for fake review detection. Some commonly used methods are linguistic analysis, metadata analysis, reviewer behaviour analysis, etc. These identifying features will be discussed in detail in Section 3.2. Hussain et al. (Reference Hussain, Mirza, Hussain, Iqbal and Memon2020) focused on identifying behavioural and linguistic indicators of fake reviews. Filho et al. (Reference Filho, Rafael, Barros and Mesquita2023) experimented on five theoretical studies.

At first, they implemented persuasion knowledge acquisition in which potential Customers can gain insights into the distinguishing features that set apart counterfeit reviews from authentic ones. Alternatively, in the second study, they are just told about which reviews are fake. Their research shows, that persons who know about the linguistic features of fake reviews are better able to detect them. But, due to the structure and vocabulary of fake and real reviews being very similar, their premise seems to fail in real life. The dichotomy of fake and real fails in real-life settings. Hlee et al. (2021) collected 4450 reviews from Yelp and showed that online reviews of new restaurants are manipulated. They elucidated the correlation between extreme ratings and counterfeit reviews. But again, they have just given a theoretical model. Li et al. (Reference Li, Wang, Zhang and Niu2021) use a reviewer grouping method. This is in the context of data mining; the objective is to categorize reviews from reviewers into distinct groups. These groups then contribute to the creation of innovative grouping models that can effectively identify both positive and negative deceptive reviews. In their study, Wang et al. (Reference Wang, Fong and Law2022) investigate the significance of emotional and cognitive cues in detecting Fake Review Deception. Upon experimentation, they found that fake reviews require deliberation at the writer’s end. Hence, emotional, and cognitive cues both play a significant role together. But writing real reviews is stress-free and hence cognitive cues are mostly absent there. However, their experiment is limited to the hospitality domain only and it is yet to be seen whether the results apply to other domains also. Table 7 shows theoretical models and other metadata-based techniques used for deceptive review identification.

Table 7. Summary of techniques used for FRD, other than ML, DL, Graph-based or bio-inspired

It is important to highlight that these techniques have the potential to be employed either individually or in conjunction with machine learning approaches, thereby enhancing the precision and dependability of fake review detection systems. Each technique has its strengths and limitations, and a holistic approach combining multiple methods often yields more robust results. Figure 5 depicts the number of publications year-wise, under various techniques. The analysis of this graph indicates that the trend in the number of publications employing purely traditional machine learning has seen a steep decline as compared to publications employing deep learning. The reason may be the advent of language generator transformers like BERT and GPT-3.

Figure 5. Distribution of various AI techniques for FRD. (The decline in the number for 2023 is because publications are taken only up to May 2023).

3.2.1. Reviewer-centric models

This model focuses on the atypical and suspicious behaviour of the reviewer as well as identifying the connections between a group of reviewers. These are non-linguistic characteristics of the reviews.

Tufail et al. (Reference Tufail, Ashraf, Alsubhi and Aljahdali2022) have extracted the behavioural features of the user. For example, review time, writing style, relationship words, grammatical errors, punctuation, etc. These features may contribute to the classification of the review. This robust model only uses supervised techniques which may not give apt results for big datasets. Alsubari et al. (Reference Alsubari, Deshmukh, Alqarni, Alsharif, Aldhyani, Alsaade and Khalaf2022) ponder upon the fact that 90% of genuine reviewers usually write one review a day, based on the product they bought or the service they used. But 70% of fraudsters may write up to 5 reviews a day. However, their dataset is limited to the hotel domain. Hence, the number of reviews per day is an important characteristic of detecting fake reviews (Heydari et al., Reference Heydari, Tavakoli, Salim and Heydari2015). However, the number of review features may not apply to other domains. According to Sadiq et al. (Reference Sadiq, Umer, Ullah, Mirjalili, Rupapara and Nappi2021), the reviewers tend to give a higher proportion of positive fake reviews compared to the proportion of negative fake reviews. But they also have limited their work to a single domain. The rating of the fake reviewer tends to be different from the ratings of the genuine reviewer. This can also help in identifying fraudsters (Ott et al., Reference Ott, Cardie and Hancock2013). The research conducted by Wang et al. (Reference Wang, Fong and Law2022) encompasses the analysis of the non-verbal behaviour of the reviewer. They say that usually, the fake reviewer attempts to mislead customers by posting reviews as early as possible. They either give maximum rating stars or minimum, depending upon the fake positive or negative review. This may be attributed to the algorithms of e-commerce engines that detect fake reviews after a certain period of days and then only remove them. But the damage has already been done till then. The reviewer’s credibility can also be assessed by determining the proportion of the reviewer’s friends and followers. In addition to these characteristics, users with longer profile timelines would be more authentic than the users who have recently created their profiles. Even with a limited number of non-verbal features employed, the computational cost is less as compared to review-centric feature selection. But theirs is a purely theoretical model and its practical implementation is yet to be seen.

Another view of the authors Sadiq et al. (Reference Sadiq, Umer, Ullah, Mirjalili, Rupapara and Nappi2021) says that the difference between the ratings given in a review and the emotional intensity of the review is also a giveaway. But they don’t predict whether the review is fake or real. Hlee et al. (2021) collected 4450 reviews from Yelp and showed that online reviews of new restaurants are manipulated. They elucidated the correlation between extreme ratings and fake reviews. But again, they have just given a theoretical model. Tang et al. (Reference Tang, Qian and You2020) used a generative adversarial network (GAN) trained model to identify six behaviour features including text, rating, and attribute features.

Thus, some of the important reviewer-centric features can be quantity, user profile, timespan, source credibility, non-immediacy, etc. The fake review detection algorithms show good performance with the incorporation of these non-textual features. This means analysing reviews using the behaviour of the review’s author or creator. But which behavioural features are to be selected for the detection problem, is a big task.

3.2.2. Review-centric model

This kind of fake review detection primarily centres on the textual content of the reviews. Research has indicated that there are substantial linguistic distinctions between authentic and fake reviews, which prove instrumental in their identification. These features may include micro-linguistic content or semantic content such as product characteristics.

Text features such as the count of nouns, verbs, and adjectives, along with the usage of strong positive and negative words in the review, have been recognized as potential indicators of a fake review (Jindal & Liu, Reference Jindal and Liu2007). The writing style of the reviewer can have lexical characters such as the number of characters and their proportion to uppercase letters, numeric characters, to the rate of spaces or tabs, called stylometric features can help in detecting fake reviews. However, they used the common features, and no new features were introduced for fake review detection. Jain et al. (Shan et al., Reference Shan, Zhou and Zhang2021), employed Linguistic Inquiry and Word Count (LIWC), which is considered a deep linguistic feature. Its output such as emotions, self-reference, social words, overall cognitive words, etc. can be incorporated to find the fake review. Moon et al. (Reference Moon, Kim and Lacobucci2021) obtain fake and authentic reviews of hotels via a survey and determine that features such as lack of details, temporal bias, and hyperbole are all part of fraudulent reviews. However, their work is based on surveys and is limited to just the hotel domain. Kumaran et al. (Reference Kumaran, Chowdhary and Sreekavya2021) have used language features like unigrams and their frequency, and bigrams and their frequency and length of reviews but conclude that behavioural features need to be added to accurately identify the reviews. However, their work focuses more on sentiment classification as positive or negative. The work done by Abdulqader et al. (Reference Abdulqader, Namoun and Alsaawy2022) finds that the short length of the online review, review replication, Term frequency-inverse document frequency (TF-IDF), cohesion and coherence measures, and stylometric features are some of the telltale signs of a spam reviewer. Along with these, LIWC such as less usage of personal pronouns (less use of ‘I’,, ‘we’ etc.), less information about time and location, and strong use of positive as well as negative words, all point to fake reviewers. However, their work is based on theory, which may or may not apply to other datasets. Text similarity is widely used as an indication of fake reviews because spammers tend to copy the reviews to save effort (Hussain et al., Reference Hussain, Mirza, Hussain, Iqbal and Memon2020). But their work has studied textual and behavioural models separately which decreases the accuracy of fake review detection. The authors Wang and Kuan (Reference Wang and Kuan2022) went a step further and divided the features into message level (review length and psychological cues), formulation level (readability and linguistic variables), and control level (rating extremity, review age, etc.). Although the dataset they used was Yelp (till 2012), their idea was to understand the process human brains undergo to formulate a message. Their limitation was that they focused on single domain and linguistic summary variables rather than composite variables.

Although review-centric features have been used extensively by researchers, it was found that they alone, or reviewer-centric features alone, are not sufficient to determine fake reviews accurately. Thus, a combination of important features of both should be incorporated to make any fake review detection algorithm outperform others. Table 8 shows some examples of the two types of identifying features used in the existing research. Figure 6 depicts the distribution of publications according to features. From the graph in this figure, it is observed that researchers have predominantly investigated the review-centric features over the reviewer-centric ones. This preference may stem from the fact that textual features are easier to identify as compared to behavioural or network-related ones. Additionally, hybrid (combination of review and reviewer-centric) and other features such as discourse and metadata (depicted in Table 8), are gaining popularity due to their enhanced performance in detecting fake reviews.

Table 8. Summary of identifying features in each publication

Figure 6. Number of publications in FRD according to features.