A Hybrid CNN-LSTM and XGBoost Approach for Crime Detection in Tweets Using an Intelligent Dictionary

Social media has evolved into a powerful communication tool for sharing news and expressing feelings on a variety of themes, particularly in Arabic, a difficult language with fewer resources than English and Chinese in addition, Arabic language instruction is popular, local information is easily accessible, and cultural idiosyncrasies make it a good tool for spotting illegal activities on Twitter. Dialects, colloquialisms, false positives, the amount of data, and ethics are some of the challenges. For responsible and accurate Arabic Twitter Crime Detection, linguistic competency, natural language processing, and cultural awareness are required. Instead of focusing on a single crime, this study looked at multiple crimes [10]. The study examines Twitter crimes in Arabic using DL and ML methodologies, comparing English outcomes. Machine learning (ML) and deep learning (DL) models are well-suited for the analysis of social media material owing to their capacity to process substantial volumes of data and discern intricate patterns. Deep learning algorithms can gradually acquire high-level features, enhance accuracy through iterative processes, and obtain advanced characteristics through the practice of feature engineering. Convolutional neural networks (CNNs) have demonstrated superior performance compared to conventional machine learning methods across several domains, including but not limited to cybersecurity, natural language processing, bioinformatics, robotics, and medical information processing. Deep learning algorithms possess the capability to acquire knowledge from extensive datasets and exhibit expedited testing durations, rendering them well-suited for real-time applications [11]. Kaddoura et al. [12] offer a strategy for identifying Arabic tweets using machine learning and deep learning techniques, using feature extraction and N-gram models, and computing performance metrics. With an F1-Score of 99.73%, the neural network method outscored the other algorithms, but GloVe outperformed rapid Text by 0.5%. Support vector machines, neural networks, logistics regression, and naive Bayes are the classical machine-learning techniques used. Deep learning approaches make use of global vector (GloVe) and fast Text learning models. The study [13] used DL models to identify offensive remarks on Arab YouTube channels, including Bi-LSTM with attention mechanism, CNN, Bi-LSTM, and CNN-LSTM. Results showed accuracy of 85.7%, 86.4%, 87.2%, and 87.8%, with CNN outperforming competitors. CNN, and CNN-LSTM models, were employed by the authors in this paper [14], in their "fast and simple" methods. On a dataset of 8K Arabic tweets, they tested each model. According to the findings, ML models cannot compete with neural learning models. Using the hybrid model CNN-LSTM, the best F1-Score was 73%. This paper's authors [15], The text analysis algorithm categorizes tweets as cyberbullying using TF-IDF and Word2Vec, using seven machine learning classifiers. The model achieves a 96.4% F1-Score rate and outperforms other entity recognition methods [16]. used ML algorithms like NB, DT, RF, and SVM to classify religious hate speech in Arabic tweets. They trained DL models using Fast Text and word2vec embedding, achieving a high F-Measure and accuracy of 71%. Finally, Aljarah et al. [17], used machine learning and natural language processing to identify cyber hate speech on Twitter in Arab environments, with the RF method achieving 91.3% accuracy.

The main smart tools used to develop the proposed approach are described in this section.

4.1 Aho-Corasick algorithm

Developed in 1975 by Aho and Corasick, the Aho-Corasick algorithm compares input text patterns using a tree-like structure, using a finite state machine and common prefixes [19]. The Aho-Corasick algorithm is an efficient string-matching method using a tree structure and failure transitions. It is widely used in applications like intrusion detection, virus scanning, text mining, and lexical due to its linear time complexity. The preprocessing step builds the automaton once, allowing multiple patterns to be compared against input texts. The utilization of the Aho-Corasick algorithm in constructing a dictionary can be characterized as advantageous in situations when there is a requirement to efficiently and effectively search a substantial collection of terms within texts. After identifying the keywords inside the text, one can categorize them or execute a certain action by the identified term. The construction of the dictionary is designed to align with the proposed approach and employ outcomes to identify criminal activities associated with Twitter posts.

4.2 Natural language processing (NLP) technique

This section demonstrates how to efficiently operate machine learning algorithms by representing text documents as numerical vectors and using the TF-IDF technique to extract features in NLP tasks. Because TF-IDF can recognize significant words in a document or corpus and give them weights depending on their significance, it is useful for feature extraction. It can handle big datasets: TF-IDF is scalable and capable of handling big datasets, which makes it appropriate for processing massive volumes of text data.

4.3 Basic machine learning

Computer scientists have focused on machine learning since the 1950s [20]. ML applications include translation and sentiment analysis [21]. ML algorithms are categorized into supervised, semi-supervised, and unsupervised types based on application goals [22, 23]. Supervised algorithms train using labeled data for tasks like crime detection and sentiment classification, using techniques like SVM, Naive Bayes, Decision Trees, Random Forests, KNN, and Logistic Regression [24, 25]. Semi-supervised algorithms use labeled and unlabeled data during training, which is useful when limited or expensive labeled data is scarce. Unsupervised algorithms, like K-Means and DBSCAN, focus on discovering patterns, clusters, or relationships in unknown, unlabeled data [26]. ML algorithms revolutionize tasks like translation and sentiment analysis, advancing techniques and applications [21]. Understanding the history of machine learning can help us appreciate XGBoost, a popular ensemble learning algorithm family [27]. It's beneficial for crime detection as it transitions from rule-based AI to data-driven methods. XGBoost excels in processing and learning from big datasets, making it a valuable tool in crime detection-based intelligent Dictionaries for a method proposed.

4.4 Deep learning (DL)

Deep learning extracts intricate features from dimensional data, creating models connecting inputs to outputs using multi-layered networks and layered neural networks for abstract computation [28]. DL techniques improve accuracy and reduce training time in complex problems, enabling breakthroughs in science, engineering, and engineering tasks like data analysis, pattern recognition, and prediction [29]. DL algorithms excel in autonomous vehicles, robotics, intelligent systems, community-related applications, and social network analysis, uncovering patterns and trends for valuable insights into user behavior and preferences [30]. DL models offer potential in health, imaging, disease diagnosis, drug discovery, and personalized medicine. DL algorithms are supervised and unsupervised, requiring labeled training data for predictions [31-33]. Deep learning revolutionizes data analysis, pattern recognition, decision-making, and driving innovation. These deep learning models consist of two parts, A and B:

A. CNN (Convolutional neural network)

CNNs are used for image analysis and text processing in cybercrime detection. They use convolutional filters to capture local patterns and features, identifying keywords, linguistic patterns, and structural characteristics. For additional processing or categorization, the output is passed into connected layers [34]. Furthermore, CNN has a specific architecture that consists of numerous hidden layers and is a Multi-Layered Perceptron network (MLP) [35, 36].

B. LSTM (Long short-term memory)

Long short-term memory (LSTM) networks, a subtype of recurrent neural networks, are capable of storing enormous volumes of data in a short period [37]. LSTMs recognize long-term dependencies, sequential patterns, and temporal correlations in data, making them ideal for handling text, modeling context, and detecting cybercrime [38]. Because CNN and LSTM are both potent deep learning models that work well with sequence prediction problems including spatial inputs, such as text data, they were selected for this model. While the LSTM is employed for sequence modeling and prediction, the CNN is utilized for feature extraction, which is crucial for text classification tasks. Because the LSTM can selectively ignore irrelevant information and remember previous inputs, it is especially helpful for sequence prediction issues. These two models work together to forecast material linked to crime in social media posts and to extract significant features from the text data. Furthermore, it has been demonstrated that CNN and LSTM perform better than conventional machine learning algorithms in a variety of fields, such as sequence prediction and text classification. To identify crime tweets on Twitter, a hybrid model combining these two algorithms was employed.

The experiments and findings for each of the models put forward in the preceding part are summarized in this section. On two datasets, one of which included a dataset of Arabic spam and Ham tweets [41], which is globally measured, and the other on a dataset that was built and compiled from Twitter based on the Aho-Corasick algorithm with the scrape Python library for data scraping, which includes Several Arabic tweets crime editorials were used to detect crimes of Tweeter tweets to study the effect of features on model learning and the resulting accuracy.

6.1 Environmental settings

The research is conducted on an Intel Core i7 computer with Python 3.11.4, Google Collab, the Keras library for deep learning models, and Sklearn for machine learning techniques for comparisons.

6.2 Experiment 1 compares results with other ML and DL models

The proposed techniques are compared to various ML algorithms and deep learning on a 20% dataset and created data set. The technique assesses classifier performance using accuracy, precision, recall, and f1-measure variables. The best classifier is picked based on higher measurement values. These parameters are Precision (Eq. (3)), Recall (Eq. (4)), Accuracy (Eq. (5)), and F1-Score (Eq. (6)) [42].

Precision $=\frac{\mathrm{TP}}{T P+F P}$           (3)

Recall $=\frac{\mathrm{TP}}{T P+T N}$           (4)

Accuracy $=\frac{\mathrm{TP}+\mathrm{TN}}{T P+T N+F P+F N}$             (5)

F1score $=\frac{2 \times \text { Precision } \times \text { Recall }}{\text { Precision }+ \text { Recall }}$            (6)

The study focuses on the efficiency and performance of classification methods using evaluation measurements. The dataset includes keywords for crimes, user names, tweets, and other features. The proposed method categorizes the dataset using ML classifiers and DL models to select the best classifier for better-assessed metrics. The results were compared to a dataset of (13240) tweets [41], and compared with a paper [12] model using the same data set. The outcomes of the DL and ML were as follows. The hybrid CNN-LSTM model achieved the highest accuracy of 99.84, and the results for RF using f1-measure is 98.76, an accuracy of 100. The results also show that the machine learning of the model outperforms most classification techniques in terms of accuracy and precision. Table 4 illustrates measurements of the ML and DL models' performance. The table shows that the proposed approach produced better outcomes than the proposed method in the paper [12]. While the paper [12] utilizing ML, and DL techniques (Fast Text+LSTM) and (GloVe+LSTM) yielded the following results for the Text+LSTM, the classifier obtained with no computation of overall accuracy, the total classifier precision, recall, and f1-measure are 99.1 92.4 2, and 95.1, respectively. The classifier obtains for the GloVe+ LSTM The total classifier precision, recall, and f1-measure are 98.88, 95.92, and 97, respectively, with no measurement of overall accuracy, but machine learning outperformed most classification algorithms in terms of precision.The results show high-performance classifiers, effectively processing text data through pre-processing steps like tokenization, stop word removal, derivation, and text normalization. These models provide accurate and valid classification based on pre-processed data. The second group (the constructed data set, which totaled 18493), and the findings are shown in Table 5: The proposed method for classifying Twitter Crime Detection performances.

4.png

Figure 4. Confusion matrix for XGBoost model based on the dataset [41]

When compared to F1, the findings showed that the Hybrid CNN-LSTM model performed the best. Figure 4 depicts for XGBoost Model machine learning classifier.

When comparing the performance of ML and DL algorithms, the test results from the data set that it built are efficient and have high performance in all machine learning categories of crime and deep learning, As the results of the XGBoost model, the classifier achieves an accuracy of 100, while the overall classifier precision, recall, and f1-measure are 99, 99.63, and 99.36, respectively. Table 6 summarizes the Twitter Crime Detection classifiers with the highest accuracy, recall, and F1 measurement findings, all of which are higher than the paper [12] using Dataset [41]. XGBoost measurement scores are greater in terms of accuracy, precision, recall, and f1. The Hybrid CNN-LSTM classifier has an overall accuracy of 99.75. The total precision, recall, and f1-measure of the classifier are 98.76, 98.75, and 99.86, respectively. According to the results, the Hybrid CNN-LSTM has a higher accuracy of 99.75 than the F1-Score. Figure 5 depicts the Machine Learning Confusion Matrix.

5.png

Figure 5. Confusion Matrix for XGBoost Model based on data set built

6a.png

(a)

6b.png

(b)

Figure 6. Performance of proposed model (a) accuracy curve (b) loss curve

7.png

Figure 7. Confusion matrix of hybrid CNN-LSTM model of crimes tweets

Compare the proposed model's performance to additional measures such as Mean Average Precision (MAP) and Average Precision for each class, as given in Table 6: Deep Learning Performance Model Comparison structure.

The Hybrid CNN-LSTM model achieved a 99.84% accuracy rate in detecting tweet crime for 100 epochs and batch size = 64, as shown in Figure 6, and the loss and accuracy model and confusion matrix are illustrated in Figure 7.

6.3 Experiment 2 compares proposed algorithms with the literature

This section illustrates compare in Table 7 displays the best results of proposed algorithms compared to the literature.

Table 4. Measurements of the ML and DL models’ performance

Table 5. The proposed method for classifying Twitter Crime Detection performances

Table 6. Deep learning performance model comparison structure

Table 7. Comparison of proposed paper with related publications