A Systematic Review of Machine Learning Methods for Movie Genre Classification and Age-Appropriateness Prediction

A systematic literature represents an attempt to provide a clear understanding of a specific topic in response to research questions by exploring the previous literature published within a specific timeframe. It discusses the results of these selected studies to deliver new knowledge to researchers, thereby saving them the effort and time required to browse each of these studies individually. This current review was conducted according to the principles of the PRISMA statement [Preferred reporting items for systematic reviews and meta-analyses] [12]. The PRISMA statement consists of a checklist of items for researchers to include when reporting systematic reviews and meta-analyses. This protocol sets out the optimal steps for preparing a study, including a search strategy, determining the sources of information, identifying the literature, screening and selecting the studies, extracting the data, determining the eligibility criteria, synthesizing the results, and assessing the quality of the studies against the risk of bias. To ensure clarity in their reporting, researchers should carefully follow the PRISMA guidelines to present a complete and standardized format, thereby providing an understanding and evaluation of the results.

2.1 Search strategy

Two inter-related topics were explored in this study: movie genre classification and the prediction of movie age-appropriateness. Five electronic databases were consequently searched: ScienceDirect, IEEE, Springer, and arXiv, which belongs to the University of Cornell and Multidisciplinary Digital Publishing Institute (MDPI). In addition, Google Scholar was used as a search engine to retrieve relevant materials. Using the combination operator ‘OR’, different search terms were entered into the advanced search feature of these databases – the same search terms were used for the databases and search engine. Table 1 presents the number of search results and their relevance to the inclusion criteria. The search was conducted over the period of one week, ending 24 April 2024, using the three-term keywords namely, ‘movie genre classification’, ‘film type prediction’, and ‘movie MPAA age rating’.

Table 1. Data sources and search terms used

2.2 Eligibility criteria

Studies related to movie genre classification and the prediction of movie age-appropriate rating were selected based on four inclusion criteria:

1. Research studies written in English.
2. Peer-reviewed journal and conference papers.
3. Fully accessible papers.
4. Papers published between 1 January 2010 and 31 December 2023, inclusive.

Meanwhile, research studies were excluded according to the following criteria:

5. Articles published before 1 January 2010 or after 31 December 2023.
6. Papers without empirical results.
7. Articles not downloaded in their full text.
8. Studies written in a language other than English.
9. Studies with no known publisher and not indexed in Science Direct or Clarivate.
10. Studies focused on specific movie genres, such as action movies only.

A total of 798 studies were sourced, of which 139 were found in the identified databases, while 659 were located by the Google Scholar search engine. The steps of the study selection procedure are illustrated in Figure 1, which depicts a PRISMA flow diagram for this review.

1.png

Figure 1. Study selection in a PRISMA statement

In the identification step, it was found that 136 studies were duplicated, out of a total of 798. As such, only 662 papers were left for further investigation. The screening step involved reviewing the title and abstract for each paper, whereupon the Rayyan.ai website [13] was accessed to remove the duplicate studies. As a result, 476 studies were excluded because they did not meet the inclusion criteria. The content of the remaining 186 studies was then carefully checked by two authors to identify the most relevant papers after applying the specified criteria. Out of a total of 186 studies, 133 were considered eligible for further content reading, out of which 53 articles were found to be inaccessible in their full texts. Moreover, 55 studies were deemed to be irrelevant to this study, according to the inclusion and exclusion metrics. Finally, just 78 studies were included after checking their titles, abstracts, and entire content.

Results

The included studies were analyzed according to five categories: 1) Identifying the study topic: movie genre classification or the prediction of movie age rating, 2) Determining the prediction or classification method, 3) Exploring the datasets used, 4) Identifying the data type used, and 5) Determining the evaluation metrics. Based on these themes, this systematic analysis highlights the key aspects of the research topic, identifies the robustness of the previous literature, and sheds light on any unaddressed research gaps that could open the door to further research.

2.3 Description of the included studies

Table 2 shows the total number of studies that resulted from the search query, combining the results of the search engine records with those of the corresponding database source.

Table 2. Sources of data used in the studies searched

The search results indicated that the Springer database contained the largest number of publications from 2010 to 2023, comprising 17% of all the papers retrieved. From the IEEE database, a 15% publication rate was found, followed by arXiv at 6%, and Science Direct at 5%. Finally, the MDPI database yielded 2% of the publications sourced. Figure 2 presents the publication rate of all the studies retrieved by the Google Scholar search engine and the above-mentioned databases, demonstrating that around 55% of the papers appeared in the search engine and met the eligibility criteria.

2.png

Figure 2. Rate of publication of the studies sourced

Meanwhile, Figure 3 illustrates the number of published studies after removing duplicate papers and implementing the eligibility criteria. It indicates that studies on movie genre classification exceed the literature on age-appropriate movie rating. Moreover, Figure 4 depicts the number of publications per year in the five databases, indicating that the publication rate has increased since 2018, whereby the Springer and IEEE databases were found to have a higher publication rate than the MDPI database. Furthermore, Figure 5 demonstrates the publication rate for movie classification and age rating, revealing that the rate of publication increased from 2014 onwards.

3.png

Figure 3. Rate of publication for both topics

4.png

Figure 4. Studies per year in the databases

5.png

Figure 5. Publications retrieved from Google Scholar

Movie genre classification and the prediction of movie age-appropriateness can be determined through content analysis or collaborative filtering methods. Figure 6 presents a taxonomy for the available methods used in this area. In content analysis, different data types may be employed for this task, consisting of text, image, video, and audio content formats, as shown in Figure 7. Textual data can take the form of a plot summary, script, subtitle, or user review [14-34]. Conversely, visual data can be a movie poster [35-51] or movie trailer clip [52-65]. Meanwhile, audio files, such as movie soundtracks, have been used in other studies [66-68]. However, some researchers have combined different data types to enhance their models. These studies may be described as multimodal [69-90]. Collaborative filtering [15] methods offer another approach by leveraging users’ viewing habits and preferences or exploring movie ratings to infer genre preferences. Table 3 summarizes the findings of the reviewed literature, identifying the research objectives, evaluation criteria, feature description, datasets, and data types.

Figure 6. A proposed taxonomy of approaches in movie genre classification and movie age rating

Figure 7. Ratio of data types used in the included studies

Table 3 provides an organized overview of the key characteristics of earlier research. It identifies each work along with its primary goals, either for genre classification (G) or age rating prediction (A). It also determines the evaluation measures used to assess models' performance as well as the computational methods used across studies. Moreover, the source and scale of the utilized datasets and the type of the used data are highlighted. This could include text (T), poster (P), audio (Au) and/or video (V). Hence, Table 3 offers a comparative view of research trends, highlighting dominant modalities and algorithms in this field. From the results summarized in this Table 3, it is clear that various types of data were used in the reviewed studies, such as movie posters, plot summaries, and trailers. These were deployed individually or in combination, for example, including a poster with audio data. This is known as a multimodal approach. Table 4 lists the studies that adopted a multimodal classification approach, which integrates multimedia information and harnesses the features of each type of media used. Such algorithms and multimodels were used to enhance classification performance and make trustworthy predictions. It is clear that visual data could include posters, trailers or videos, while textual data could encompass multiple sources, such as scripts, plot summaries, synopses, subtitles, metadata, and descriptions or overviews. Moreover, the auditory features extracted from movies or trailers were incorporated with other data types.

Table 3. Characteristics of the reviewed studies

Table 3. Characteristics of the reviewed studies (Continued)

Table 3. Characteristics of the reviewed studies (Continued)

Table 4. Studies on multiple data types

2.4 Datasets

A movie genre dataset typically contains information about various films and their associated genre. Such a dataset is essential for various applications in film studies, data analysis, machine learning, and recommendation systems. The earlier literature included numerous datasets, some containing textual data such as plot summaries, script, subtitles, synopses, and metadata. Examples of these datasets comprise the Internet Movie Database (IMDb) [52, 78], Movie Database (TMDB) [48, 81], Multi- language Movie Review Dataset (MLMRD) [34], and the Movie Summary Corpus (CMU) [22, 30].

Other datasets contain visual data, divided into poster, trailer, or video files, with examples including open movie datasets (OMDb) [31], MovieLens [15, 31, 39, 55], the Labelled Movie Trailer Dataset (LMTD) [24, 39, 86], and MediaEval [64, 68, 77]. In contrast, other datasets contain multiple data types such as posters, plot summaries, audio content, and movie trailers, for example, trailers in Moviescope [73, 80, 89] and multimodal movie trailer datasets (MM-Trailer) [91]. Conversely, some researchers have created their own datasets with a variety of data types to address data quality and reliability in relation to the problem of imbalance labels [52, 66, 68, 84, 90].

Figure 8 illustrates the employed datasets and the number of studies that used each one, along with the using ratio across the included studies. The IMDb dataset [18, 19, 26, 27, 29, 33, 35, 37, 40-42, 76, 85, 92] is considered to be the most significant and is widely accessed for information on films, TV shows, and the entertainment industry in general. It is especially useful for movie data analysis, machine learning models, and developing recommendation systems, since it contains basic information about films such as the title, genre, year of release, cast, director, and screenwriter. This dataset has garnered a great deal of attention, having been sourced in 14 published papers (15%), as illustrated in Figure 8. Meanwhile, the MovieLens dataset [15, 20, 24, 31, 36, 43, 45, 46, 54, 55, 58, 60, 63, 70, 74, 83, 86] has been effectively deployed for movie genre classification, with 17 research studies using it for this purpose (18%). Meanwhile, the Labeled Movie Trailer Data (LMTD) dataset [24, 39, 46, 58, 60, 63, 70, 72, 74, 86] contains around 500 trailer files and was included in 10 studies (10%). The Movie Database (TMDb) [16, 28, 32, 48, 50, 51, 81] represents another important dataset, frequently utilized in the development and testing of various machine learning models in movie research. It contains comprehensive information about films and TV shows and was referenced in seven of the selected studies (7%). Alternatively, the MediaEval dataset [68] is a compilation of multimedia datasets, de-signed for benchmarking and evaluating diverse multi-media processing and analysis tasks. The above-mentioned dataset, which encompasses a variety of media types such as video, audio, text, and images, serves as a valuable resource for multimodal analysis and was found to have been employed in three of the research papers reviewed (3%).

Figure 8. Datasets used in previous studies

In contrast, the Movie Summary Corpus (CMU) was sourced in just two of the included papers (2%). The CMU contains detailed plot summaries for a large number of movies, along with metadata such as genre information [22, 30]. Meanwhile, the Moviescope dataset contains rich metadata and annotated content, for example, title and overview, with a movie summary included in three of the studies (3%) [73, 80, 89]. Finally, other datasets were only mentioned once in the literature, such as the Kaggle Movie Dataset [36], SUBTIEL Corpus [14], Opensubtitles [21], MovieNet [82], Mul-ti-modal IMDb (MM-IMDb) [76], MovieLens Mul-ti-relational Dataset (MLMRD) [34], Internet Movie Script Database (IMSDb) [25], Open Movie Database (OMDb) [31], and Scene Understanding (SUN) [65]. Moreover, 31 datasets were created in several different studies to gather the distribution of different data types (visual and textual) in data records of varying size (32%).

2.5 Machine learning models applied

Different methods are utilized for predicting the label or class of trained datasets. The selection of a model is based on various metrics such as the type of data, complexity of the problem, effort required in terms of time and resource consumption, and level of accuracy. Traditional machine learning and artificial neural networks (ANNs) represent two different approaches in the machine learning domain. Traditional machine learning techniques involve a set of algorithms that learn patterns from data to make predictions or decisions. These methods often require careful feature engineering and domain knowledge to improve their performance.

The most common techniques in traditional machine learning include logistic regression (LR), decision trees (DT), random forest (RF), SVM, K-nearest neighbors (k-NN), Naïve Bayes (NB), and gradient boosting machines (GBM). These models have some advantages since they are typically considered faster to train and evaluate with small datasets. However, they can struggle with very large datasets or high-dimensional data, and their performance depends heavily on the quality and relevance of the features. A brief description of these techniques is presented in this subsection.

Artificial neural networks are computational models that are inspired by the human brain. Hence, they consist of layers of interconnected neurons that can learn complex patterns. This category includes several different techniques such as feedforward neural networks, recurrent neural networks (RNNs), convolutional neural networks (CNNs), and Transformers. Feedforward neural networks are especially powerful in deep learning architecture, whereas CNNs are specialized neural networks, designed to process grid-like data such as images. Convolutional neural networks are widely used in image classification tasks and many models are constructed upon them, for example, ResNet, AlexNet, MobileNet, Inception, and VGG networks. In contrast, RNNs are designed for sequential data, such as time series or natural language processing, wherein they capture temporal dependencies. Numerous models are designed around their structure, including Long Short-term Memory (LSTM), Bidirectional Long Short-term Memory (BiLSTM), and the Gated Recurrent Unit (GRU).

Alternatively, Transformers constitute an advanced neural network architecture that is designed to deal with sequential data based on an Attention Network Model. These models have become state-of-the-art in the field of natural language processing since they can handle large datasets, complex high-dimensional data, and learning features, directly from raw data and by reducing the need for manual feature engineering. They can be applied to a wide range of tasks, including image, text, audio, and time-series analysis.

However, they require significant computational resources and training time to achieve high performance, especially in the case of deep networks with large amounts of labeled data. Nevertheless, to enhance models' performance, multiple models may be combined to produce reliable final output predictions, potentially outperforming the results of individual models.

These are known as ensemble methods, two popular examples being Random k-Labelsets (RAkEL), and Gated Multimodal Units (GMU).

Some models, such as Binary Relevance (BR) can only handle binary classification. Thus, in order to address multi-class or multi-label classification, techniques such as one-vs-rest (OvR) or one-vs-one (OvO) are required, using multiple binary classification. Moreover, other models, like Multinomial Logistic Regression or Multi-label K-nearest Neighbors (ML-kNN) have been extended to handle multi-class and multi-label classification directly using sigmoid activation or optimizing a softmax function.

Figure 9 illustrates the distribution of methods and models used across the reviewed studies. Each model type is shown with the frequency used and the corresponding percentage of adoption in the included research. The researchers deployed SVM [3, 16-18, 20-22, 25, 28, 33, 34, 39, 43, 66-69, 72, 78, 81, 87, 88, 90, 91] in 24 studies (13%); 15 studies (8%) included Naïve Bayes [16, 18, 21, 27, 29, 30, 42-45, 50, 59, 79, 84] and CNN [3, 28, 34, 38, 40, 41, 46, 57, 70, 72, 73, 83, 85, 86, 91]; 11 studies (6%) contained LSTM [3, 28, 31, 33, 34, 56, 72, 73, 85, 88, 91]; nine studies (5%) employed kNN [26, 28, 36, 42, 48, 67, 69, 88, 93]; VGG [37, 41, 42, 73, 83-85, 89] was used in seven studies (4%); eight studies (4%) utilized ResNet [36-39, 41, 42, 71, 83] and ANN [22, 26, 32, 33, 39, 82, 88, 90]; seven publications (4%) featured DT [16,32,66,77,84,88] and RF [21, 32, 33, 43, 67, 73, 92]; five studies (3%) employed the RAkEL ensemble approach [44, 48, 50, 51, 84]; five studies (3%) utilized LR [18, 22, 27, 31, 53] and RNN [3, 28, 29, 31, 91]; Inception [49, 71, 83] and ML-kNN [36, 43, 48, 84] were applied in four studies (2%); BiLSTM [14, 31, 33], Attention Neural Network [13, 33, 76], MobileNet [37, 47, 71], and AlexNet [41, 42, 45] were used in three studies (2%), and GRU [72, 83] and GMU [80, 91] were deployed in just two studies (1%).

Figure 9. Machine learning models used in the reviewed studies

Finally, some methods were used in just one study, with a combined 18% rate of inclusion. These methods included CTT-MMC [86], Movie-CLIP [82], the Vector Space Model (VSM) [81], the Hidden Markov Model (HMM) [59], content-based filtering (CBF) [55], distance ranking (DR) [48], YOLO v3 [45] SEMPD [43], GloVe 42B [89], PART [66], Lenet, canonical correlation analysis (CCA) [67], HAN architecture [14], EfficientNetB7 [83], XGBoost [29], AdaBoost [16], ConvNeXt [37], the Universal Sentence Encoder [71], DenseNet [37], SAS-MC-v2 [24], CRNN [73], the Residual Dense Transformer (RDT) [35], BERT [91], Topological Data Analysis (TDA) [19], Linear Discriminant Analysis (LDA)[69], Knowledge Graph Feature Formation [76], the Content-based Recommender System (CB) [54], ILDNet [58], TFAnet [60], the Feature Fusion Network (FFN) [63], Dual Image and Video Transformer Architecture (DIViTA) [75], and hierarchical clustering relevance feedback (HCRF) [77].

2.6 Feature selection techniques

Machine learning models require feature extraction techniques during the important step of converting raw data into a more suitable form for analysis and processing. The choice of feature extractor will depend on the type of data and task. Various types of feature extractor are used with different data types such as textual, visual, and/or audio-data. The most frequently used feature extractors in textual data include Bag of Words (BoW), Term Frequency-inverse Document Frequency (TF-IDF), and N-gram. Bag of Words converts text into fixed-length vectors by counting the occurrence of each word in a document. Meanwhile, TF-IDF weighs the importance of words by considering their frequency in a document and their rarity across all documents. On the other hand, N-gram is used to transform text into numerical features based on contiguous sequences of ‘N’ items. Other techniques involve word-embedding and are exemplified by Word2Vec, GloVe, fasttext, and Doc2Vec [94]. These techniques map words to dense vectors in a continuous vector space where semantically similar words are close together. Moreover, contextual embedding, as performed by BERT [95], provides dynamic word representations based on the context in which words appear. Conversely, the LSTM model is based on RNN architecture and can be used to capture long-term dependencies in textual data.

There are a large number of feature descriptors specified for visual data. These are distributed via traditional methods such as SIFT, YOLO, and others, which collect the key points, edges, and textures from images or CNNs that deploy convolutional layers, in order to learn hierarchical feature representations from raw pixel data via automated means [96]. Some models are pre-trained; for instance, VGG, ResNet, and Inception use pre-trained models on large datasets (like ImageNet) to extract features from images. Furthermore, MPEG-7 refers to Multimedia Content Description Interface, which is used to describe multimedia content data such as audio and image data, and GIST is a descriptor that can capture the spatial structure in image data. Meanwhile, for audio-data, some feature extractors like Mel-Frequency Cepstral Coefficients (MFCCs) can capture the power spectrum of audio-signals and represent them in a way that closely resembles human auditory perception or spectrograms. This consists of visualizing the spectrum of frequencies in a signal as it fluctuates over time.

Figure 10 depicts the variety of feature extractors employed in the literature. Each one is presented alongside its usage frequency. The TF-IDF [17-21, 23, 26, 31, 32, 44, 82, 84, 88, 90, 91] appearing in 14 studies (22%), representing a relatively high rate. Meanwhile, BOW [23, 27, 29, 31, 81, 88], MFCC [66, 72, 79, 83, 86, 97], Glove [3, 24, 32, 84, 89], Word2Vec [14, 29, 32, 84], and Doc2Vec [22, 25, 28] were involved in six (9%), seven (11%), five (8%), four (6%), and three (5%) of the studies, respectively. Conversely, FastText [32, 73], CNN [83, 88], GIST [50, 52], and N-gram [30, 88] were used in 3% of the studies, while other feature descriptors were used only once, such as DNN [55], Local Binary Pattern (LBP) [88], Wavelet features [90], VGG16 [84], LSTM [88], Inception v3 [88], resNet50 [84], fastVideo [73], Wang2vec [32], Discrete Fourier Transform (DFT) [83], and Discrete Cosine Transform (DCT) [83].

Figure 10. Feature extractors used in the reviewed literature

2.7 Assessment criteria

Evaluation metrics are deployed to measure the performance of classification models. According to the metric used, different insights may be gained into the level of a model’s performance. There are three types of classification: binary, multi-class, and multi-label. Accordingly, different metrics are designed for this purpose. In binary classification, accuracy metrics [98] are noted as the most commonly used evaluation measure for the proportion of correctly predicted instances out of a total number of instances. Precision [98] is considered as another important metric, if measuring the proportion of true positive predictions out of all positive predictions. Recall (sensitivity, true positive, or hit rate) measures [98] the proportion of true positive predictions out of all actual positives (number of relevant items selected). The F1-Score measures [98] the harmonic mean of Precision and Recall, providing a balance between the two. Specificity (true negative rate) measures [98] the proportion of true negative predictions out of all actual negatives. The ROC-AUC (Receiver Operating Characteristic - Area Under Curve) [99] is used to measure the trade-off between true and false positive rates across different thresholds, whereby a higher AUC indicates better model performance. Meanwhile, the Loss metric measures the performance of a classification model if the prediction input is a probability value of between 0 and 1, with lower log loss indicating a better model.

The Jaccard Index [100] and the Matthews Correlation Coefficient (MCC) [99] are metrics used to evaluate the performance of binary classifiers. In multi-class classification, the same accuracy metric is used but considers all classes, whereby Precision, Recall, and F1-score [98] can be calculated for each class separately. Alternatively, if in macro averaged format, the metric can be averaged across all classes. However, when micro-averaged [98], the metric is calculated globally across all instances and classes. Sample Average Precision (sAP) [98] is an evaluation metric that is primarily used in the context of object detection and image classification tasks, where it measures the average precision of the detection results over a set of samples. Unlike traditional average precision (AP) which considers Precision-Recall curves across an entire dataset, sAP calculates Precision for each sample and then averages the Precision. The calculation of sAP [98] involves two steps: first, Precision is calculated for each sample. Average Precision is calculated from the Precision values across all the samples.

The Area Under the Precision-Recall Curve AU (PRC) or AUC-PR measure [101] is an evaluation metric used to assess the performance of binary classification models, especially when dealing with imbalanced datasets where one class is significantly more frequent than the other. It focuses on the trade-off between Precision and Recall, which are often more informative than the ROC curve in such scenarios.

Another metric is the Confusion Matrix, this being a table that presents the number of correct or incorrect predictions for each class, thereby helping to clarify where the model is making errors. A further evaluation metric, known as Ranking Loss, indicates how frequently false prediction labels are given a higher ranking than true prediction labels.

In multi-label classification, the Hamming Loss measure [98] is considered to be one of the most important metrics to measure the percentage of wrongly predicted labels across all samples in a dataset.

Some studies included more than one metric for evaluation, for instance [63, 74, 79, 87], whereas others [52, 68, 81] included just one metric. Figure 11 provides an overview of the evaluation metrics utilized in the included studies. It identifies not only the total occurrences of each metric, but also their relative distributions within the literature. This provides insight into measures that are most frequently used for evaluating models' performance. The F1-score [69, 72, 84] was included in 36 of the studies (21%); Recall measurement [33, 45, 83] was employed in 35 studies (20%); the Accuracy metric [14, 19, 34, 37] occurred in 33 (19%) of the studies, and Precision [21, 22, 24, 25, 34] was mentioned in 27 of the reviewed publications (16%).

Figure 11. Metrics used in the studies

The other metrics, mentioned in Table 3, had a lower rate of inclusion, starting with Hamming Loss [21, 41], in six publications (4%); mAP [73-75, 80] in six publications (4%); AU(PRC) [24, 61, 63] in six publications (4%); sAP [73-75, 80] in four publications (2%), and uAP [73-75, 80] in three publications (2%). However, some metrics were used in only two (1%) studies, namely, Loss [37], Jaccard Index [29, 43], Ranking Loss [74], AUC [37, 72] and CD [69]. Meanwhile, other metrics appeared just once (1%), these being Specificity [65], EM [21], MCC [53], RMSE [54], and MAE [54].

This study represents a systematic review of the literature on movie classification, both in relation to movie genre and the prediction of age-appropriate rating. These areas of movie classification are considered important in research on information retrieval and multimedia use, due to their significant applications in recommendation systems, content organization, and digital libraries. Various data types were used in the studies sourced for this review, classifiable into textual data (for example, plot summaries, synopses, and scripts) or visual data (for example, posters and movie trailers). However, even though audio data was considered to be an important source, this type of data has received less attention in earlier studies. Thus, a combination of diverse data sources could have attained greater accuracy in the results.

In the reviewed studies, various data types were fed into different classifier models that were distributed across traditional machine learning and deep neural networks. The authors of the previous studies explored and created different datasets, as well as using an array of evaluation metrics to evaluate the performance of their models, handle multi-label classification problems, and deal with data imbalance. Therefore, due to the heterogeneity of the studies, a direct comparison between their outcomes is difficult.

Although this review makes significant contributions in this direction, it is not without its limitations. Therefore, this review invites further systematic reviews to consider the following recommendations for future research. First, creating benchmark datasets and standard evaluation metrics could help establish a foundation for comparing different algorithms and proving their performance. Second, most studies have limited their attention to English language movies, without considering movies in other languages. Therefore, it is important to address the challenge of genre classification across different languages and cultures. Third, advanced deep learning techniques based on Transformer models, such as BERT, should be explored to enhance the capture of complex relationships in the data. Forth, conducting cross-cultural studies is essential for testing the generalizability of models across different linguistic, cultural, and contextual settings, which remains a major gap in current research. Moreover, adding user data such as reviews could enrich the datasets used in genre classification and thereby contribute to a robust outcome.

Additionally, the search for pertinent studies was executed on just one occasion, which could have led to the omission of relevant papers published after the specified period. Second, the exclusion of studies that were not fully accessible and not written in the English language could have affected the findings and generalizability of this research. Moreover, this systematic review only included papers sourced from a few selected databases. Therefore, studies in other databases were overlooked, although an attempt was made to overcome this limitation by using the Google Scholar search engine. Finally, the use of predefined keywords may have restricted the inclusion of other studies. Hence, the use of different keywords could expand the search results and enrich the study with a higher volume of publications.