A systematic literature review on spam content detection and classification

Introduction

The word spam generally means some unwanted text sent or received through social media sites such as Facebook, Twitter, YouTube, e-mail, etc. It is generated by spammers to divert the attention of the users of social media for the purpose of marketing and spreading some malware etc. The e-mail spam messages are sent in bulk to various users, with the intention of tricking them into clicking on fake advertisements and spreading malware on their devices. The spam messages provide a good source of income for the spammers (Bauer, 2018) and, hence, they continue to spread them rapidly. To combat spam in e-mail, a lot of techniques have been involved, but the spam content continues to increase (Statista, 2017). These spam messages cause financial loss to business e-mail consumers and also to the general users of e-mail (Okunade, 2017).

Spam is common on social media sites like YouTube, and it mainly consists of comments and links to pornographic websites, as well as irrelevant videos. These comments are sometimes created automatically by bots. Although the definition of spam on online video game sharing services is debatable, instances of message flooding, requests to join a specific group, violations of copyrights, and so on are occasionally referred to as spam. Spam in blogs, often known as splog, refers to comments that have nothing to do with the topic of discussion. Frequently, these comments are accompanied by links to commercial websites. Some splogs are devoid of unique content and contain stuff plagiarized from other websites (Rouse, 2015).

Spam is also included in written reviews of products that are available on social networking sites. According to Liu & Pang (2018), about 30–35% of online reviews are deemed spam. These spam reviews are intended to influence people’s purchasing decisions and to affect product ratings (Saini, Saumya & Singh, 2017; Ho-Dac, Carson & Moore, 2013). As a result, detecting bogus reviews appears to be a major worry, and online review systems may become utterly useless unless this vital issue is addressed (Jin et al., 2011; Govtnaukries, http://www.govtnaukries.com/you-wont-ever-use-head-and-shoulder-shampoo-after-watching-this-video-facebook-spam/). Fake/spam profiles abound on social networking platforms like Facebook and Twitter, and users are bombarded with SMS messages from these identities. To analyze the spam content many researchers Song, Lee & Kim (2011) have employed the attributes from Facebook including community, URL, videos and Images. By identifying and filtering the spam and non-spam accounts Stringhini, Kruegel & Vigna (2010) could identify and characterize the spam using statistical techniques. Mateen et al. (2017) have used honey-profiles to record the activity of the spammers and applied this technique to social media content for spam detection using a novel tool. The graph models were also popular to detect spam based on the different features of the map and they could find the relationships that exist among the social media users (Benevenuto et al., 2010). In recent times, the machine learning algorithms are getting popular and they are used in spam detection (Rathore, Loia & Park, 2018; Liu et al., 2016; Zheng et al., 2016; Serrano-Guerrero et al., 2015).

The steps in detecting spam on social media are often as follows. Obtaining the spam text collection (dataset) is the initial step. Because these datasets frequently have unstructured text and may contain noisy data, preprocessing is almost always necessary. The following step is to select a feature extraction method, such as Word2Vec, n-grams, TF-IDF, and so on. Finally, a variety of spam detection technologies, such as machine learning, deep learning, and Lexicon-based algorithms, are utilized to decide whether texts are spam.

The rationale of our work is to bring out a detailed survey of several spam detection and categorization algorithms. We are aware that many previous surveys on spam detection may not have acquired the information that we obtained from various popular academic data sources. Some previous efforts on spam identification from social media have constrained themselves to only a few limited academic sources. Some earlier studies failed to highlight the benefits and drawbacks of various spam detection and classification systems. The novelty of our work is that we used data from a variety of reputable academic sources to achieve our goal of identifying spam content on social media. We have also highlighted certain significant strategies, along with their benefits and drawbacks when applied to various spam datasets. We also covered deep learning and other crucial Artificial Intelligence (AI)-based spam detection approaches that have previously only been found in restricted investigations.

This extensive survey will assist academics who are interested in spotting social media spam using AI techniques, as well as addressing the issues associated with it. Using the proposed survey, researchers will be able to select optimal detection and control mechanisms for spam eradication. Our work will let academics compare the many existing spam detection works in terms of their merits, limits, approaches, and datasets employed. This study will also assist researchers in addressing current research possibilities, concerns, and challenges connected to spam text feature extraction and classification, as well as specifics on various data sets used by other researchers for spam text detection.

We compare the accuracy of existing spam text detection systems in order to determine which ones are the most effective. “Survey Methodology” describes the survey methodology used to conduct our comprehensive review. “Steps for Detecting Spam in Social Media Text” uses a block diagram to explain the multiple steps involved in spam detection. “Collection of Social Media Textual Data (Dataset Collection)” provides a summary of the datasets available for social media spam text. The following section, “Pre-processing of Textual Data”, goes over the various spam text pre-processing procedures. “Feature-Extraction Techniques” and “Spam Text Classification Techniques” investigate several feature extraction methodologies and spam categorization algorithms. Deep learning techniques for spam classification are discussed in “Deep Learning (DL) Approaches for Spam Classification”. “Challenges in Spam Detection/classification from Social Media Content” discusses the difficulties encountered in spam detection, and “Open Issues and Future Directions” concludes with a list of references.

Survey methodology

The goal of this survey is to undertake a thorough literature evaluation on approaches for detecting and classifying spam content in social media. There are several sources of textual data on social media platforms such as Facebook, Twitter, E-mail, and YouTube. A variety of ways have been used to detect and regulate spam text. Our efforts are primarily motivated by a desire to learn more about different spam text detection and categorization algorithms. This section discusses the survey methodology that we used to conduct our detailed spam detection review.

Database selection

We extracted research papers from a few academic digital sources to conduct the literature review. Expert advice was sought regarding source selection, and databases such as Web of Science (WoS), Scopus, Springer, IEEE Xplore, and ACM digital library were used to collect research papers for our study. We used search query terms such as “social media spam,” “twitter spam,” “review spam,” and “spam text,” among others. The academic data sources with their links that are used in our work is listed in the Table 1 below.

Table 1:

Description about academic databases and their links.

Academic Data sources	Search string	Links
WoS	Social spam	https://apps.webofknowledge.com/
Scopus	Spam AND Twitter	https://www.scopus.com/
Springer	Spam AND Artificial Intelligence	https://link.springer.com/
IEEE Xplore	Social spam AND Artificial Intelligence	https://ieeexplore.ieee.org/
ACM Digital Library	Online spam AND Review Spam	http://dl.acm.org/
Science Direct	Social media AND Spam	http://www.sciencedirect.com/

DOI: 10.7717/peerj-cs.830/table-1

In this review, the title of each paper was scanned and identified for possible relevance to this review. Any paper that does not refer to social media spam was eliminated from further investigation. The abstract and keywords of the publications were scanned for a deeper review and a better understanding of the papers. The Fig. 1 below displays the distribution of articles depending on publishing types such as journals, conference proceedings, books, and other reference materials that were referred for our extensive spam detection survey.

We may conclude from the article distribution pie-chart that for our work, the majority of the articles referred to were from journals and conference proceedings, and that some technical reports were also used to obtain material for our systematic literature review.

Steps for detecting spam in social media text

The task of spam detection and classification requires several processes, as depicted in Fig. 2. Data is collected in the first stage from social networking sites such as Twitter, Facebook, e-mail, and online review sites. Following data collecting, the pre-processing activity begins, which employs several Natural Language Processing (NLP) approaches to remove the unwanted/redundant data. The third phase entails extracting features from the text data using approaches such as Term Frequency-Inverse Document Frequency (TF-IDF), N-grams, and Word embedding. These feature extraction/encoding approaches convert words/text into a numerical vector that can be used for classification.

The last step is the spam detection phase, which employs several Machine Learning (ML) and Deep Learning techniques to classify the text into categories like spam and non-spam (ham).

Collection of social media textual data (dataset collection)

The first phase in spam identification is the collecting of textual data, comprising spam and non-spam (ham) material, from social media sites such as Twitter, Facebook, online reviews, hotel evaluations, and e-mails. They are extracted with the help of an appropriate API, such as the Facebook API or the Twitter API, which are both free and allow users to search and collect data from several accounts. They also enable the capture of data using a “hashtag” or “keyword,” as well as the collecting of data posted over time. Based on the text content, we can identify data as spam or ham, and official social networking sites may flag some accounts or postings as spam. The following Table 2 presents some of the datasets regarding E-mail spam and Twitter spams. It also displays a description of the dataset as well as some of the reference studies performed on those datasets.

Twitter, a prominent microblogging network, has attracted people from all around the world looking to express themselves through multimedia content. Spammers transmit uninvited information, including malware URLs and popular hashtags. Twitter suspends accounts that send a high volume of friend requests to people they don’t know, as well as accounts with a high number of followers but few followers. Table 3 below includes descriptions and references for some of the Twitter spam datasets.

Sites such as TripAdvisor, Amazon, and Yelp, among others, have online reviews of a product, hotel, or movie. These reviews include input from previous customers who have purchased a product or stayed at a hotel. Spammers blend spam content with these reviews to convey a negative impression about a product or service, causing the firm financial harm. Table 4 below covers a few datasets linked to online reviews, as well as several reference studies on detecting spam in reviews.

Table 5 below contains some of the most prevalent spam words seen in e-mail, Twitter, and Facebook posts. If your e-mail contains any of these words, it’s quite likely that it'll end up in the spam bin.

Pre-processing of textual data

Text-preprocessing is a significant technique for cleaning the raw data in a dataset, and it is the first and most important stage in removing extraneous text (Albalawi, Buckley & Nikolov, 2021; HaCohen-Kerner, Miller & Yigal, 2020). Before extracting features from text, it is necessary to eliminate any undesired data from the dataset. Unwanted data in the text dataset include punctuation, http links, special characters, and stop words.

As illustrated in the Fig. 3, there are numerous text-preprocessing techniques available that can be used to remove superfluous information from incoming text input.

Feature-extraction techniques

Because many machine learning algorithms rely on numerical data rather than text, it is required to convert the text input into numerical vectors. This method’s goal is to extract meaningful information from a text that describes essential aspects of it.

Spam text classification techniques

Text classifiers can organize and categorize practically any sort of material, including documents and internet text. Text classification is an important stage in natural language processing, with applications ranging from sentiment analysis to subject labelling and spam detection. Text classification can be done manually or automatically, however in the manual approach, a human annotator assesses the text’s content and categorizes it correctly. Machine learning techniques and other Artificial Intelligence (AI) technologies are used to automatically classify text in a faster and more accurate manner utilizing automatic text classification models. As shown in the Fig. 4 below, there are three techniques of classifying the text.

Deep learning (dl) approaches for spam classification

Deep learning models are gaining popularity among NLP researchers due to their ability to solve challenging problems (Kłosowski, 2018; Torfi et al., 2020). Deep learning is based on the idea of building a very large neural network inspired by brain activities and training it using a massive amount of data. They can cope with the scalability issue and extract the features from the data automatically. The most popular deep learning models among NLP researchers are Convolutional Neural Networks (CNN) and Long Short Tern Memory (LSTM) networks. Convolutional Neural Networks (CNN), one of the most important and extensively used Deep Learning approaches, has received a lot of attention in recent times for performing NLP tasks. It has been used successfully for sentiment analysis (Kim & Jeong, 2019), image (Sharma, Jain & Mishra, 2018) and text categorization (Song, Geng & Li, 2019), pattern recognition (Mo et al., 2019), and other tasks. For text categorization, Lai et al. (2015) used a recurrent structure to capture contextual information from textual data. Their technique was able to capture semantic information from text and outperformed CNN in classifying text texts. Tai, Socher & Manning (2015) employed the Long Short Term Memory Network (LSTM) to capture sequential information in textual data, and they built a tree LSTM model that could perform well for NLP applications. Basyar, Adiwijaya & Murdiansyah (2020) built a Long Short Term Memory (LSTM) network and a Gated Recurrent Unit (GRU) model to detect spam in the Enron e-mail spam dataset, which contained 34,519 records. The LSTM model outperformed the GRU model in spam detection, achieving an accuracy of 98.39%. Alauthman (2020) employed the Gated Recurrent Unit-Recurrent Neural Network (GRU-RNN) to recognize Botnet spam E-mails. On the SPAMBASE dataset, which included 4,601 spam and 2,788 non-spam e-mails, they achieved an accuracy of 98.7%. They evaluated the performance of GRU with several machine learning algorithms, but the GRU-based strategy produced the best results for spam detection. Hossain, Uddin & Halder (2021) used feature selection techniques including Heatmap, Recursive Feature Elimination, and Chi-Square feature selection techniques, along with Deep Learning models such as RNN, to select the most effective features for spam e-mail detection. On spam text information obtained from the UCI machine learning repository, they achieved a 99% accuracy. Tong et al. (2021) used a deep learning model based on LSTM and BERT to overcome issues such as unfair representation, inadequate detection effect, and poor practicality in Chinese spam detection. They created this model to capture complex text features using a long-short attention mechanism. In their work to detect spam reviews related to hotels, Liu et al. (2022) used a combination of Convolution structure and Bi-LSTM to extract important and comprehensive semantics in a document. They could be able to outperform current methods in terms of classification performance by achieving an F1-Score of around 92.8. There are many other research works (Crawford & Khoshgoftaar, 2021; Bathla & Kumar, 2021) employing Deep Learning (DL) techniques for spam detection that could capture contextual information of text for spam identification.

Based on the prior work on spam classification with Deep Learning approaches presented in Table 14. These Deep Learning techniques definitely helps in improving the performance of the spam detection model and also helps in reducing the effects of over-fitting that is seen in Machine Learning models. Unlike ML techniques, deep learning methods do not necessitate a manual feature extraction process or a large amount of computational resources. It can adapt to a wide range of spam content found in social media text and will be very effective at extracting spam data from the text. Based on previous research, we can deduce that combining word-embedding techniques with Deep Learning methods improves spam classification performance. However, with less training data, it is more difficult to avoid over-fitting, and the presence of unlabeled text in the input corpus will lower performance. The deep learning method is used to classify text that saves a lot of manpower and resources while also improving text classification accuracy.

Challenges in spam detection/classification from social media content

Spam content on social media continues to rise as people’s use of social media grows dramatically. The technology underlying spam spread is amazing, and some social media sites were unable to correctly identify spam contents/spammers. Some legitimate social media users manufacture duplicates in order to communicate with a group of recognized pals. It is tough to distinguish between a spammer and a legitimate user with a duplicate profile. Spammers also employ many fake identities to distribute dangerous and fraudulent material, making it harder to track them down. A spammer may also employ social bots to automatically post messages based on the user’s interests. Many businesses use “crowdsourcing” to enhance production, in which some people are paid to offer false reviews about a product that is not good. The machine learning method for spam detection suffers from over-fitting and sometimes suffers from a lack of training samples. They may also encounter difficulties if the spammer is intelligent and quick enough to adapt. When the input dataset is quite large, ML approaches suffer from temporal complexity, and memory requirements are also an issue. If there are undesirable features in the dataset, the classifier’s performance suffers, and an efficient feature selection algorithm is required.

Unsupervised learning suffers from a storage shortage, as well as a scarcity of efficient spam detection methods. As a result, there is a strong need to pursue a method that is flexible and efficient, such as Deep Learning, in order to tackle the challenges encountered by traditional Machine Learning methodologies. Spammers also employ Deep Learning algorithms to manipulate social media material in order to generate spam. These bogus contents developed using Deep Learning algorithms are difficult to detect, necessitating more effort to resist them. If there is a shortage of properly annotated data available, the notion of transfer-learning might be used as an alternative to Machine Learning.

Open issues and Future Directions

Some of the issues in spam detection are the presence of sarcastic text, multilingual data, and improper labelling of the datasets. Many researchers use APIs to gather data related to a given language and geographical area, there is a bias in the data collected through social media. Some studies employ raw data without much pre-processing, which results in duplicated features and lower classification performance. Some datasets exhibit a class imbalance, for example, the ‘spam’ class has a large number of samples whereas the ‘ham’ class has a small number of samples.

There are a limited number of labelled datasets available for spam text, as well as a limited number of attributes available in these text datasets, which is a problem. For efficient research, a dataset with correct labelling is required, as is large computational power in the case of a large dataset. Only a few studies have used deep learning techniques and semantic approaches to detect spam. Exploring the use of multimodal content (text and images) from social media for social media would be a significant future challenge.

Conclusion

We have described numerous strategies for spam text identification in depth in our systematic literature review on spam content detection and categorization. Our research also looked into the various techniques for pre-processing, feature extraction, and spam text classification. This survey will assist researchers in conducting research in the field of social media spam detection as it highlights some of the best works done in this field. We’ve also provided details on a number of databases that can be used for spam detection studies. The various previous works on spam text pre-processing, feature extraction, and classification will aid researchers in determining the most appropriate strategies for their research in this area. In future development, we’d like to include some other spam detection approaches, as well as their benefits and drawbacks.