Fake Review Detection Using Machine Learning

Using computational approaches, opinion mining examines and finds views, feelings, and subjective data in enormous amounts of text. It was utilized for testing goods and services, notably in terms of consumer acceptance and perceptions of specific companies and people [1]. Everyone is free to express their views and opinions anonymously and without fear of repercussions. Social media and online posting have made expressing oneself openly and confidently easier. These opinions have pros and cons; while they could help get the right feedback to the right individual, who could help fix the problem, they can also be manipulated. These viewpoints are seen as beneficial. This makes it simple for those with bad intentions to take advantage of the system, provide the appearance of sincerity, and publish comments to promote their goods or disparage those of rivals without disclosing who they are or where they operate. Such folks can write any bogus review. This behavior might be referred to as "opinion spamming”.

Because it has the potential to distort public discourse and galvanize large groups of people behind causes that run counter to legal or ethical norms, opinion spamming on social and political problems can be downright terrifying. As social media opinions become more widely used in real life, it's realistic to assume that opinion spamming will grow more pervasive and sophisticated, making it harder to spot. However, they must be identified to guarantee that social media remains a reliable source of public opinion rather than one rife with propaganda and misinformation.

This study uses the Yelp.com dataset to construct a model for classifying the features derived from text and spam using machine learning techniques to detect spam or non-spam reviews. Several classifiers named Extreme Gradient Boosting, Stochastic Gradient Descent Stochastic, and Support vector classifier are proposed to analyze the fake reviewer “spam and not spam “to suggest a better model for review-centric fake detection. Our model was applied to the balanced and imbalanced datasets as original, with oversamples and undersamples in random techniques.

The remaining part of the paper proceeds as follows: Section 2: Literature Review; Section 3: Tying up the various theoretical backgrounds Section 4 proposes a system framework; and Section 5 discusses the model outline and work. Section Performance Evaluation Section seven illustrates the conclusion and future work.

This section reviews the theoretical concepts used in implementing the proposed system.

3.1 Opinion mining

Consider a consumer-authored review that offers feedback on a product that falls under the classification of reviews. The opinion or opinion created is the review the customer makes to communicate his or her thoughts—mostly positive or negative—about the goods. The review classification seeks to discern whether a person has written a positive or negative review depending on an assessment of the text's point of view.

Opinion mining aims to find the attributes regarding the object on which opinions were offered in each of the reviews r ϵ R and to determine the orientation of comments, i.e., whether the comments are negative or positive if a set of text reviews (R) with opinions on an object is provided.

Assume you are provided with a collection of text reviews (R). When people have ideas about anything, opinion mining aims to locate the aspects discussed in each review (r R) and determine their orientation or whether they are positive or negative. Figure 1 illustrates many interchangeable phrases in opinion mining [9].

1.png

Figure 1. Synonyms of opinion mining [9]

Review mining has emerged in the past few years, which performs the computational evaluation of the users' opinions, sentiments, subjectivity, appraisals, feedback, emotions, etc., expressed in customer reviews. The term "subjectivity" relates to the person's feelings, perspectives, desires, and emotions. Objectivity typically stands in contrast to it. Speech/writing events expressing private states, private states, and expressive subjective elements. Extracting subjective hints, like phrases, terms or expressions, and applying them to determine whether the related sentence (or document) is objective or subjective are the two fundamental goals of subjectivity analysis. Subjectivity classification. It is required to extract any reviews or comments the author has made in order to obtain valuable data [9]. Modified Black Widow Optimization algorithm, which outperforms other bio-inspired approaches in global optimization and convergence speed, but is not as advanced as the best developed ones [10]. Research on the Holy Quran improves speaker identification systems, recognizing Arabic and English speakers using 14 professional reciters' speech signals, with improved LBG-VQ algorithm matching codebook centroids with 96.43% accuracy [11]. The review highlights the effectiveness of deep learning in energy forecasting, highlighting the use of various neural networks, including simple RNN, LSTM, GRU, and bidirectional RNN [12]. The research introduces an efficient odometry method for autonomous path planning, enabling global optimal planning, mapping, and localization in static obstacles, improving computing speed and position accuracy [13].

3.2 Sentiment analysis

Sentiment analysis, which is usually referred to as "opinion mining," is the research subject. Analysis is done on people's attitudes, views, feelings, assessments, and appraisals of goods, services, organizations, people, issues, events, subjects, and traits [14]. Sentiment analysis determines the text's relative polarity. It determines if a text is constructive, destructive, or neutral. It is also known as "opinion mining" since it reveals the speaker's viewpoint or attitude. Users can publish product reviews on social networking and shopping websites, which serve as platforms for classification. A training set is needed to analyze the reviews [15] effectively. In the e-commerce reviews, the sentiment analysis technique is applied. A "tough mountain" is saved for the quantity of reviews.

Sentiment analysis also assists in classifying free-form language as positive, negative, or neutral, summarizing the customer's judgment to comprehend another customer's expression and strengths for products and merchants.

3.3 Fake review detection

Fake review detection can be defined as a field of NLP. It aims to evaluate, identify, and classify product reviews on online e-commerce sites as either fake or authentic. Deceptive product reviews frequently employ false opinions [16]. It is currently a popular academic subject. Fake reviews are often described as spam reviews, deceptive opinions, and spam opinions, and the people who write them could also be called spammers [17]. Review readers are duped by the practice of mind spamming. Users who engage in spamming activities are referred to as "spammers." Spammers' fake reviews aim to give a company a false reputation, either negative or positive. Certain companies hire spammers to download the material in an effort to get new customers or to demote a competent company inside the same company [18]. People publish erroneously good product reviews to advertise products. Sometimes, maliciously unfavourable reviews of other (competition) products are written to harm their reputation. Some are commercials and promos, neither offering reviews or opinions on the goods. Fake reviews have many drawbacks that can influence people's decisions to buy commercial goods, including it might be difficult since a term might be positive in one situation but harmful in another. For example, saying "long" while referring to a laptop's battery life is a favourable attitude, yet using the same word when referring to the startup time is unfavourable. The opinion mining system trained on words from many viewpoints cannot comprehend a word's nature, which has multiple interpretations that depend upon context. Different people have different ways of expressing themselves. Almost all traditional text-processing techniques operate under the presumption that minute textual differences have little impact on meaning. The difficulty in determining reviews that were intentionally edited is tricking people by using various DM methods. People frequently make contradictory statements, which makes it difficult to ascertain their opinions. A negative review can actually have a positive meaning. Opinions of the product may periodically be negative or positive. They also examined the benefits and drawbacks of data mining (DM) techniques to anticipate dishonest and honest ratings.

3.4 XGBoost

The gradient boosting approach has been scaled up and refined, and Extreme Gradient Boosting (XGBoost) was created to increase the model's performance. It is a piece of the Distributed Machine Learning Community's open-source library [19]. XGBoost is a flawless software and hardware solution that quickly and accurately enhances the current boosting approaches. Tianqi Chen developed it. Two tree construction methods are used in it: Scaled Quantile Sketch and Sparsity-aware Split Finding. These are two techniques for finding splits and choosing where to divide. A method for equitably distributing possible splits across the data to take into consideration the significance or weight of a particular data point is called a weighted quantile sketch. Through binning continuous data, a histogram is created; as a result, splits should be performed only once per bin instead of for each function value. A missing value in Split Finding with Handles for Sparsity directs the choice down the left or right path, depending on the feature, by figuring out the default path of the tree node, which shows that the XGBoost disregards rows with missing data. Like every other ML technique, XGBoost optimizes the loss function that defines the model's error. It accomplishes this by applying gradient descent to the loss function.

3.4.1 Stochastic gradient descent stochastic

A linear Gradient Descent classifier increases algorithm performance and lowers the cost function to produce a more accurate model. Using the classifier function SGD Classifier, we put this method into action. Despite the SGD Algorithm, despite its apparent lack of complexity, has emerged as a leading contender among the default standard optimization algorithms for ML classification methods such as the NN and LR. However, the gradient descent of the SGD favours picking a data instance at random. The model is a powerful facilitator because of its simplicity; the SGD Algorithm has been considered one of the effective and default standard optimization algorithms for ML classification, like the NN and LR. On the other hand, the gradient descent of the SGD is biased toward randomly selecting an instance of data [20]. This model is strong facilitation. Data categorization and NLP are two areas where traditional machine learning approaches have struggled, but SGD has proven effective at overcoming these challenges. Efficient and Rapid Computation.

3.4.2 Support vector classifier

The support vector classifier (SVC) can be defined as a subtype of SVM [1]. When dealing with problems involving two groups, a supervised ML model, or SVM, utilizes classification methods [1]. After receiving new data for each category, an SVM model could classify it [1]. An SVM produces a hyperplane or set of hyperplanes in a high- or infinite-dimensional space, which could be utilized for regression, classification, or other tasks like outliers’ detection [4]. Closer to the hyperplane than other data points, the support vectors affect the hyperplane's position and orientation [3]. An SVM algorithm should be able to handle new data that it has never seen before, in addition to classifying objects into two groups [5]. The main step of this approach based on the following parameters;

Parameter C for the regularization process. A more complicated model with less room for error will arise from increasing C. If you set C lower, your model will be easier to understand and have more room for error.

The kernel functions. The data is mapped onto a higher-dimensional space, where the classes are more easily linearly separable, using the kernel function.

Kernel coefficient, or gamma. Each data point's weight in determining the decision boundary is determined by the gamma coefficient.

Degree is the degree of the kernel of the polynomial.

Polynomial kernel constant term coefficient, denoted as coef0.

The suggested model identifies fake reviews using machine learning, including SVC, SGD, and (XGboost). The dataset utilized in the implementation is separated into training and testing data based on its actual characteristics, with 80% for training and 20% for testing. It is arguable if the fake review filters increase the user’s confidence in the review system. The price of filters might go much higher. Such behavior fosters mistrust by increasing the knowledge of bogus reviews. The proposed method looks into and analyzes the comments, classifying them as either real or fake. The Yelp.com dataset's real features, such as spam (1) and not spam (0), as well as features (stop words removed), are used in this model's feature extraction and sentiment analysis. Figure 3 demonstrates our ML technique. Most crucially, we process three ML algorithms' approaches concurrently.

3.png

Figure 3. Machine learning approach

In this paper, we applied three machine learning approaches to detect fake reviewer, and the results were somewhat similar. The approach of the current work can accurately identify the false reviews on the Yelp dataset in terms of balanced and unbalanced.

Our methodology shows effective results with an imbalanced dataset for binary classes. In contrast to the prior study, the suggested model in the present study could accurately detect fake reviews in the Yelp data set.