Predicting Age and Gender in Author Profiling: A Multi-Feature Exploration

1  Introduction

Author Profiling (AP) evaluates author’s personal information based on various linguistic features, e.g., stylistic, semantic, and syntactic [1]. AP is considered an important task in digital security forensics. Also, it plays a key role in anti-harassment, targeted advertisement, analysis of text for personality identification, medicine, and marketing [2,3]. AP has multiple applications for educational cognitive purposes. Such as, researchers can find out the level of knowledge of exceptionally talented students by evaluating their writing skills. Also, AP can be used to identify various author’s attributes from a given text [4]. Scammers and hackers employ false names, gender, age, and location on social media to conceal one’s identity [5]. Law enforcement organizations can use AP to track individuals using forged identities. Likewise, people exhibit vandalism, harassment, and Cyberbullying using alternative or fake identities to ensure that they never get caught [6,7].

AP could help to identify such cases using different natural language processing (NLP) algorithms [8]. From a marketing point of view, organizations will get to know each other by analyzing online websites and items. What kind of people comment on their items, and accordingly, they will promote their efforts towards a specific gender or age limit. It can also determine which age group, gender, occupation, etc., use certain products or comment on certain posts. This information can be used for targeted advertising and increasing revenue for different companies. Research into the AP domain shows that the language features of Facebook, status updates, tweets, messages, and blog posts allow us to evaluate the age and gender of authors accurately [9]. Deep Learning and Machine Learning significantly contribute to modeling various factors to optimize AP systems. Also, few traditional classification methods classify authors based on demographic factors like age, gender, and personality traits [1,10]. Additionally, researchers have explored linguistic features such as stylometric analysis, semantic and syntactic, and, more recently, deep learning techniques to improve classification accuracy. However, while existing studies have made significant strides in this field, there remains a research gap in effectively combining ensemble models like Random Forest and Adaboost for potentially enhanced performance in author profiling tasks, especially regarding age and gender classification.

Despite the existing research efforts in author profiling using various machine learning techniques and stylistic features, a significant research gap emerges in the lack of exploration and comparison of ensemble models combining Random Forest (RF) and AdaboostM1 (AB). While individual models have been examined extensively, more comprehensive investigations must be conducted into the potential performance enhancements and synergies that can be achieved by integrating these two powerful techniques [10]. Furthermore, most prior studies have focused on specific languages or datasets, neglecting the potential generalizability of ensemble approaches across diverse linguistic and textual domains [11]. This research gap warrants a thorough examination of the effectiveness of ensemble techniques in author profiling tasks and the potential improvements they can offer over traditional single-model approaches while also considering broader applicability across linguistic and demographic variations [12]. This research contributes to the advancement of author profiling methodologies by offering a novel ensemble-based approach that outperforms traditional single-model methods, thereby expanding the toolset available for researchers and practitioners. The key contributions of the paper are:

1.    Investigate and improve the accuracy of AP for Roman Urdu text messages while evaluating authors’ attributes (such as age and gender) based on linguistic features.

2.    Analyze the FIRE’18-MAPonSMS dataset for AP purposes and propose an AdaBoostM1 and Random Forest (AMBRF) ensemble model that leverages multiple linguistic features.

3.    Examine the effectiveness of this ensemble model compared to existing models. Perform a comprehensive comparative analysis of the proposed AMBRF model against well-known AP models from the literature, including J48 Decision Tree (J48), Naive Bayes (NB), K Nearest Neighbor (KNN), and Composite Hypercube on Random Projection (CHIRP), NB-Updatable, RF, and AdaboostM1.

4.    Examine the advantages of ensemble models over single models regarding predictive accuracy and performance. Focus on Boosting as an ensemble technique for constructing a robust classifier from weak classifiers, aiming to enhance overall predictive performance.

The rest of the paper is organized: Section 2 presents the brief literature, and Section 3 discusses the proposed model. Section 4 presents the experimental setup, while Section 5 illustrates the results achieved and the discussion. Finally, Section 6 concludes the study.

2  State of the Art

The FIRE’18-MAPonSMS dataset is used in [13] for AP. They used a multilingual (English and Roman Urdu) SMS-based document and conducted different experimental analyses. They concluded that RF achieved the best accuracy of 73.714% for gender while using all 14 language-independent features together and an accuracy of 58.571% for the age group by using all 29 features together. For the best result the authors concluded that they obtained the best accuracy of 58.571% for the age. The overall result was compared with the baseline technique.

In [14], the authors focused on the ML techniques such as KNN, Support Vector Machine (SVM), Logistic Regression (LR), Radial basis Support Vector Machine function (RBFSVM), SVM Linear (SVML), (Convolution Neural Network (CNN). Seventeen stylometric features were used to train the model. The best accuracy for both detections was 92.45%, achieved from the English dataset, and 90.36 for gender classification. For the Spanish dataset, 89.68% and 88.88% for bot detection. The authors in [15] used various ML algorithms, which are SVM, LR, deep pyramid Convolution Neural Network (DPCNN), NB, Gaussian Naive Bays (GNB), NB complement (NBC), LR, and RF (Random Forest), Region Based Convolution Neural Network (RCNNs) for AP. They used stylistic and word-level features. The models do not predict unfrequented demographics, i.e., non, binary gender, and work that is not single-topic (e.g., manager, professional, and science). Predicting the date of birth year works best between 1980 and 2000 for that age is 20–40, but it is not good for older personalities.

The author in [16] proposed a technique for predicting Age and Gender where they used Multilingual (English and Spanish) corpus (PAN-2018) datasets and applied RF to classify age and gender. They used different features, i.e., Lexical, Grammatical category, Close words, Suffixes, and Signs. Precision, recall, and F1 scores are used as evaluation measures. The results obtained using only the training set show a better classification in gender than in age; however, in neither case, it exceeded 55% of the F1 measure. Besides, this measure lowers when the two classifiers are joined, reaching an F1 value between 40% and 44%. In [17] authors have used a combination of Semantic, Syntactic, and NLP as a feature, and then all these combinations are fed as an ensemble model that classifies age and gender. They employed a supervised Random Forest ensemble classifier for the AP using the PAN2014 dataset.

The only working language indicates readability criteria, function words, and structural features, which play a vital role in identifying the age and gender of the writer. For predicting an author’s profile, researchers [18] proposed a technique for Twitter bots that can only categorize human gender as female and male. Both users were there in 11 Twitter bots, and from their profiles, only a hundred tweets were selected overall, and a hundred tweets were selected randomly. They focused on the Semantic feature category, which is present in the tweets. They joined those semantic features with other stylistic features and POS tags. They used different machine learning methods, having an ensemble model; they said that Adaboost is fruitful, and the F1 score is 0.99. For the English language, the results gain an accuracy of 89.17%. RF technique was used to predict the profile of an author. In [19] describes an approach to working with an author’s profile PAN 2013 Challenge. This work is based on the idea of a linguistic method successfully used in other classification tasks, such as document writing. They considered three different features: Syntactic, stylistic, and semantic. Each represents a different aspect of the text. They extracted similarity relationships between attribute vectors in test files and center-specific modality clusters for each method. The authors [20] presented a clear feature in the form of a group; each group is then together with appropriate pre-processing steps for each group. Structural trigrams, counts of Twitter’s most important characteristics, and stylometric grouping were used.

The authors clarified that age and gender prediction as a classification job and character prediction are regression problems employing SVM and Support Vasomotor Rhinitis (SVMR), respectively. Researchers [10] focused on age and gender prediction and carried out experiments using deep learning (DL) methods, i.e., LSTM, Gated recurrent units (GRU), BI-LSTM, and CNN. BAT-AP-19, RUEN-AP-17 corpus, and SMS-AP-18 corpus datasets were used for training and testing. This research focused on POS features. The best accuracy was achieved when the Bi-LSTM classifier was used. Best accuracy was achieved accuracy = 0.882, F1 score = 0.839 and accuracy = 0.735, F1 score = 0.739 for age and gender, respectively, using LSTM. To predict the age, gender, education, language, country, and emotions of AP [21], researchers used SMO, RF, Bagging, and Instance bases learning with parameter k (KNN) and Lib SVM. Character (case word length), lexical (function word correlate), structural (document category Hyper Text Markup Language (HTML)), and POS features are extracted from the English Email dataset in this research. The best result was achieved using RF for the language classification, which is 84.22%. In this research [22], the author proposed a technique for age, gender, and country prediction. They used a linear, SVM, and MLP classifier as an ML technique. Consequently, we proposed the following AdaboostM1 with Random Forest (ABMRF) model to address the highlighted issues.

3  Proposed Methodology

All authors’ text messages are fed into the system in .txt format. The proposed system architecture, depicted in Fig. 1, comprises six components. Upon inputting the text documents into the system, preprocessing steps are applied to each document. After the preprocessing stage, in phase 3, a comparative analysis was undertaken using two distinct strategies (feature selection) to predict the author’s age and gender. This prediction was accomplished by examining the author’s writing behavior through eight learning models: RF, AdaboostM1, CHIRP, J48, NB, NBupdatable, and ABMRF. The comparison was conducted based on the chosen strategy.

Figure 1: Proposed methodology

During the fifth phase, the model selects a file, typically the top one, and makes predictions for a specific class. Additionally, the class assigned to the test files is determined by the majority votes, with the highest count dictating the chosen class. The system’s output is the predicted gender and age class for the test files.

In real-world applications, multi-class problems are prevalent. Therefore, understanding how RF operates in the context of Multi-class AdaBoostM1 becomes crucial. For instance, AdaBoostM1 can execute RF in multi-class scenarios, provided the application of AdaBoostM1 achieved accuracy exceeding 50%. Both algorithms should be used correctly for optimal results [23].

Ensemble models are preferred over single models for two primary interconnected reasons. An ensemble model can yield more accurate predictions and outperform individual contributing models. It can also reduce the variance in predictions and enhance overall model performance. Consequently, our focus has been on Boosting, an ensemble technique aiming to construct a robust classifier from a set of weak classifiers. This approach can improve a model’s robustness and reliability. The proposed research methodology is illustrated in Fig. 1.

It is an ensemble learning method that combines multiple weak classifiers to create a strong classifier. It operates by iteratively training a sequence of weak classifiers, assigning higher weights to misclassified instances in each iteration. The weak classifiers are typically simple classifiers, such as decision stumps. During training, ABM1 adjusts the weights of the training instances based on their classification accuracy. The subsequent weak classifiers focus more on the previously misclassified instances, allowing them to correct the mistakes made by previous weak classifiers. The final prediction is determined by combining the predictions of all weak classifiers, weighted by their accuracy. The process is discussed in Algorithm 1.

In the aforementioned equations, xi represents the feature vector of instance i, and yi represents its true label. ht (xi) represents the prediction of weak classifier ht, for instance, i. The sign function returns +1 for positive predictions and −1 for negative predictions. The proposed ensemble model ABMRF works flow as described in Algorithm 2.

The proposed model relies on AdaboostM1 (ABM1) and Random Forest (RF) to enhance analysis compared to other employed models. ABM1, a widely-used machine learning model, is specifically designed for classification tasks. The quasi-code for ABMRF is shown Algorithm 3.

Why did we select an ensemble model rather than a single model?

1.   AdaBoostM1 assigns higher weights to misclassified samples in each iteration, forcing the model to focus more on challenging instances. In contrast, Random Forest builds trees on random subsets of the data without explicit emphasis on misclassifications. Combining these two approaches allows the ensemble to give extra attention to difficult-to-learn patterns, contributing to a more robust model.

2.   Random Forest inherently helps reduce overfitting by building multiple trees and averaging their predictions. Combining it with AdaBoostM1, which sequentially fits models to the residuals, can further enhance the ability to generalize well to unseen data.

3.   AdaBoostM1 effectively handles imbalanced datasets by assigning higher weights to misclassified samples. Combined with Random Forest, which inherently has mechanisms for handling imbalanced data, the ensemble model can perform tasks with imbalanced class distributions better.

4.   Random Forest provides a measure of feature importance, which can be beneficial in understanding the relevance of different features in the NLP (Natural Language Processing) task. The combination with AdaBoostM1 may further refine the importance rankings and contribute to a more accurate feature selection process.

5.   AdaBoostM1 and Random Forest are built on different principles, and combining them introduces diversity to the ensemble. Diversity is often crucial in improving the model’s overall performance, as different models may capture different patterns or nuances in the data.

4  Experimental Setup

The app runs on Google Colaboratory (Colab), which provides free CPU cloud services via TensorFlow. The Google Colab has a 33 GB hard drive, 13 GB of RAM, and a 2-core Xeon 2.2 GHz processor. Python 3. x (3.6) runs and tests the system. The output produced using stylometric features is given as input to ML algorithms. For the experimental setup, our model uses 10-fold cross-validation to evaluate the performance of the model. We conducted our experiments using eight different ML algorithms: NB, J48, KNN, CHIRP, AdaBoostM1, NB Updatable, RF, and ABMRF.

4.1 Dataset Description

This study focuses on FIRE’18-MAPonSMS, a Roman Urdu dataset available at https://lahore.comsats.edu.pk/cs/MAPonSMS/de.html. The age and gender categories with their counts are presented in Table 1. FIRE’18 MAP on the SMS dataset consists of testing and training instances. Training consists of 350 instances and 1 truth file having mentioned gender as defined as male or female, while age is sub-categorized as the following groups: 15–19, 20–24, 25–XX. In training 350 instances, there are 210 male 18 files and 140 female files, while age-wise from 15–19, there are 108 records; from 20–24, there are 176 records 25–XX, there are 66 records. Testing consists of 150 instances, which will be used to test the proposed research model.

4.2 Stylistic Features

The words or sentences are arranged in such a way that it gives us understanding. Narrative perspectives, the structure of stanzas, and position are stylistic features. The (14) Stylistic features are shown in Table 2.

4.3 Word-Based Features

The word-based features depend upon the length of the word. We can define these features based on the dataset we are using word-based features are based on the words used in a document [4]. The Word-Based features used in this study are presented in Table 3.

4.4 Pre-Processing Steps

Pre-processing comprises three phases, and traditionally, text pre-processing has held significant importance in NLP. It involves simplifying language to enhance the effectiveness of machine learning algorithms. The dataset underwent the following three preprocessing steps:

4.4.1 Remove Whitespaces

Almost all textual data in the world contains white space, the removal of which provides readability and ease of understanding [14]. To remove the starting and ending spaces of a text line, we have used the strip () function in Python. This is illustrated in Fig. 2. Part A shows white space before the word “Jaldi.” This space is called the starting space of a sentence. Part B shows the preprocessed text; when the text is preprocessed, the white space is removed, so there is no space before the word “Jaldi”.

Figure 2: Remove whitespaces

4.4.2 Stop Words Removal

Stop words have no significant semantic relation to the context in which they exist [15]. In this phase, stop words are removed using Python code, e.g., in the example shown in Fig. 3, Part A has “Mai” and “hai,” both considered stop words. In Part B, the stop words are removed when the text is preprocessed.

Figure 3: Stop word removal

4.4.3 Tokenization

It is the method by which text is broken down into smaller pieces known as tokens [18]. Words, numbers, and punctuation marks are supposed to be tokens. In this preprocessing phase, the sentences are tokenized using the split () function in Python, which separates the tokens inside the file by identifying the space between the tokens inside the sentence. In Part A of Fig. 4, the text is neither preprocessed nor tokenized. In Part B, each word is tokenized.

Figure 4: Tokenization

4.5 Training and Evaluation

Model training and testing are the core phases of any ML-based analysis. To this end, the study focuses on 10-fold cross-validation [19,23], which is a process for assessment that splits the complete data into ten subgroups of equal sizes; one subgroup is used for testing, whereas the rest of the subgroups are used for training, continuing until each subgroup has been used for testing [24–29]. The performance of the proposed and other ML models is evaluated using standard assessment measures, including accuracy, precision, recall, F-measure, and MCC (Matthews Correlation Coefficient) [20–22]. These can be calculated as shown in Eqs. (1)–(5).

Accuracy=TP+TNTP+TN+FP+FN(1)

Precision=TPTP+FP(2)

Recall=TPTP+FN(3)

F−measure=2×[(Precision×Recall)(Precision+Recall)].(4)

MCC=TP∗TN−FP∗FN(TP+FP)(TP+FN)(TN+FP)(TN+FN)(5)

5  Experimental Results and Analysis

This section provides a detailed discussion of experimental results. The term “Classifier” denotes the machine learning algorithm, including NB, NB Updatable, J48, KNN, RF, CHIRP, AdaBoostM1, and ABMRF are employed to generate numeric scores. All the evaluations, including precision, recall, F-measure accuracy, and MCC, are derived from the Confusion Matrix (CM).

The “Classifier” indicates the ML algorithm we applied to produce NB, NB Updatable, J48, KNN, RF, CHIRP, AdaBoostM1, and ABMRF numeric scores. All the evaluations, including precision, recall, F-measure accuracy, and MCC, are derived from the CM. The analysis determines the superior performance of the proposed model compared to other models. Table 4 shows each technique’s Correctly Classified Instances (CCI) and Incorrectly Classified Instances (ICI).

The outcomes achieved via precision, F-measure, and recall for age are shown in Table 5. Some table rows, e.g., Table 5, contain a (?) sign. Here, the question mark sign is used instead of the message “DIV/0!” due to its value “0”. In the confusion matrix, according to different equations, the division cannot be performed if some values need to be divided and that value becomes “0”. Nevertheless, in the tables, we used “?” instead of “DIV/0!”

Here, we have just shown the average values of each attribute. When precision, recall, and F-measurement are evaluated using stylistic features, the suggested ABMRF surpasses existing approaches, attaining 0.723, 0.543 for precision and recall, while for F-measure RF produced 0.46, and, respectively. Conversely, CHIRP and AdaboostM1 produces the poorest results in precision, recall, and F-measure, which are 0.409, 0.354, and 0.232, respectively.

The outcomes achieved via precision, F-measure, and recall for gender are shown in Table 6. Here, we have just shown the average values for each attribute. When precision, recall, and F-measure are evaluated using stylistic features, the suggested ABMRF surpasses existing approaches, attaining 0.723, 0.711, and 0.687 for precision, recall, and F-measure, respectively. Conversely, CHIRP produces the poorest results in precision, recall, and F-measure, which are 0.309, 0.395, and 0.232, respectively.

Moreover, Table 7 shows the outcomes of each technique using MCC for age and gender. These analyses show the better performance of ABMRF, with an accuracy of 54.29% and 71.14% for age and gender, respectively. However, using stylistic features, CHIRP shows the weakest performance.

Comparing ABMRF with other applied techniques, the best result was achieved by ABMRF, i.e., 54.29% for age. On the other hand, the worst performance of NB and NB Updatable is noted with an accuracy of 35.4%, while ABMRF achieved the best result of 71.14%, whereas the worst performance of NB is noted with an accuracy of 35% for age while CHIRP is noted with an accuracy of 39.54% for gender. The accuracy of stylistic features (age and gender) is illustrated in Table 8.

5.1 Result Analysis Using Word-Based Features

The best result achieved by ABMRF for age was 177, and KNN misclassified 173 instances. For gender, AdaBoostM1 gained the best result, i.e., 211, while AdaBoostM1, i.e., 139, achieved the worst result among the ten classifiers. The outcomes achieved via CCI and ICI are illustrated in Table 9.

The outcomes achieved via precision, F-measure, and recall for age are shown in Table 10. Here, we have just shown the average values of each attribute. When precision, recall, and F-measure are evaluated using word-based features, the suggested NB and NB updatable surpass existing approaches, attaining 0.506 for recall, 0.489, and 0.484 for precision and f-measure, respectively. CHIRP and KNN, on the other hand, produce the poorest results in precision, recall, and F-measure, which are 0.417 and 0.362 using CHIRP and 0.489 using KNN, respectively.

The outcomes achieved via precision, F-measure, and recall for gender are shown in Table 11. Here, we have just shown the average values of each attribute. When precision, recall, and F-measure are evaluated using word-based features, the suggested CHIRP, AdaBoostM1, and KNN surpass existing approaches, attaining 0.700 for recall, 0.603, and 0.588 for precision and F-measure, respectively. CHIRP, J48, and KNN, on the other hand, produce the poorest results in precision, recall, and F-measure, which are 0.571, and 0.441 using CHIRP and 0.319 using KNN, respectively.

For age, the best outcomes gained by NB and NB Updatable, i.e., 0.153 while the worse result gained by CHIRP was 0.028. For gender the best result achieved by NB Updatable, i.e., 56.8571% while the worse result achieved by J48, i.e., 0.094. The illustration of MCC are shown in Table 12.

Comparing all applied eight techniques, the best result was achieved by ABMRF, i.e., 50.5714%, while the worst performance of KNN is noted with an accuracy of 41.71% for age. For gender, AdaBoostM1 gained 60.29%, and the worst performance of CHIRP is noted with an accuracy of 44.1261%. The accuracy of word-based features is illustrated in Table 13.

Fig. 5 shows the percentage difference (PD) for the age of the better technique compared with the rest of the other employed techniques. PD is calculated as shown in Eq. (6).

Figure 5: PD for age

PD=n1−n2n1+n22∗100(6)

where n1 represents the value of ABMRF and n2 represents the value of other techniques. For Stylistic features, the illustration shows the minimal difference between ABMRF and RF is 1.07% and the highest difference is achieved using ABMRF with NB updatable is 42.04%. While for word based features the minimal difference between ABMRF and NB updatable is 0.57% and the highest difference between ABMRF and IBK is 19.21%.

Fig. 6 shows the PD for gender using Stylistic features. The illustration shows the minimal difference between ABMRF and RF is 1.21%, and the highest difference between ABMRF with CHIRP is 57.10% using stylistic features. While for word based features the minimal difference between ABMRF and RF is 0% and the highest difference between ABMRF with CHIRP is 30.49%.

Figure 6: PD for gender

There are two primary reasons to prefer an ensemble model over one interconnected. An ensemble model can outperform any contributing model in terms of performance and accuracy. It may also stop the prevalence or dispersion of predictions and model performance. As a result, we focused on Boosting, an ensemble strategy for attempting to build a strong classifier from a set of weak classifiers. It may also improve the overall robustness or reliability of a model. With AdaBoostM1, we have ensemble RF. In an ML project, these are significant considerations, and we may sometimes choose one or both qualities from a model. The following are some of the reasons why we chose ensemble ABMRF. It gives variable weight, which helps identify the variable with a beneficial influence. ML models are frequently overfitted, but RF classifiers are not. We have different amounts of text in each file in this situation. Furthermore, when a class is rarer than other classes in the data, as we have, it may automatically balance datasets. Ensemble classifiers also outperform nonlinear classifiers on a wide range of tasks.

As mentioned in the analysis, ABMRF outperforms in contrast with other working classifiers to improve accuracy. This study focuses on eight diverse ML techniques for AP detection. The techniques are evaluated using multiple assessment measures on a Fire’18-MAPonSMS dataset.

ABMRF has been due to its better performance than other models employed in this study on the Fire’18-MAPonSMS dataset. We compared CHIRP, J48, RF, NB, IBK, AdaboostM1, NB Updatable, and ABMRF techniques and found that the Ensemble ABMRF is the most optimal age and gender classification for AP. Tables 8 and 13present the accuracy analysis of age and gender, which shows the better performance of ABMRF, which achieved an accuracy of 54.2857% for age prediction and 71.1429% for gender prediction calculated on stylistic features. Regarding word-based features, age and gender were considered in 50.5714% and 60%, respectively. On the other hand, KNN and CHIRP show the weakest performance using all the linguistic features for age and gender prediction.

6  Conclusion

The focus of this research is to improve the accuracy of AP. The first and foremost approaches discussed in this research are ML approaches to achieve this goal. The dataset was taken from the web address “FIRE 18-MAPonSMS” in Roman Urdu, and accuracy, precision, recall, F-measure, and MCC are used as evaluation metrics. The accuracy results from the proposed model are compared with those of J48, NB, NB Updatable, CHIRP, KNN, RF, and AdaboostM1. The results show that ABMRF performs best compared to the other used models. This study focuses on proposing an ensemble model based on RF and AdaboostM1. The proposed model outperforms other employed models; however, there are also some limitations. The proposed model is less interpretable, and any wrong selection can lead to lower predictive accuracy than an individual. Features tuning is very hard while mapping with both ensemble models. Moreover, it is a bit expensive in terms of space and time-consuming.

However, it is important to acknowledge certain limitations of the proposed model. Firstly, the model’s interpretability is reduced, and incorrect selections may lead to decreased predictive accuracy compared to individual models. Additionally, feature tuning proves challenging when integrating RF and AdaboostM1 models in an ensemble. Furthermore, the model requires considerable resources in terms of space and time, resulting in higher computational costs. Future research directions in AP may involve broadening the focus to incorporate more demographic characteristics, incorporating deep learning strategies, addressing privacy issues, investigating cross-lingual profiling, and adjusting to new digital communication platforms, allowing a more thorough and flexible approach to understanding people through their online content. Furthermore, further investigations should involve larger datasets to expand the scope of analysis. Furthermore, using deep learning (DL) techniques can lead to better results. By doing so, a more comprehensive understanding of author profiling in Roman Urdu can be achieved, ultimately improving accuracy and effectiveness.

7  Declarations

I declare that this research paper, titled “Predicting Age and Gender in Author Profiling: AMulti-Feature Exploration” is an original work conducted by Aiman et al. The study focuses on detecting authors’ age and gender in Roman Urdu text messages through an ensemble model (AMBRF) based on AdaBoostM1 and Random Forest. The research utilizes the Fire18 Map on SMS dataset, proposing a model contrasted with established approaches such as ABMRF, J48, NB, KNN, CHIRP, NBUpdatable, RF, and AdaboostM1.

Acknowledgement: The authors thank Prince Sultan University and the EIAS: Data Science and Blockchain Laboratory for their valuable support.

Funding Statement: The authors would like to acknowledge the support of Prince Sultan University for the Article Processing Charges (APC) of this publication.

Author Contributions: The authors confirm their contribution to the paper as follows: Study conception and design: Aiman, Bilal Khan, Muhammad Arshad; data collection: Aiman, Muhammad Arshad; analysis and interpretation of results: Aiman, Bilal Khan; draft manuscript preparation. All authors reviewed the results and approved the final version of the manuscript.

Availability of Data and Materials: The data used in this study is online available at https://lahore.comsats.edu.pk/cs/MAPonSMS/de.html. The code for this research is available on GitHub at https://github.com/aiman-syed/Features-Extraction.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.

References

1. I. Ameer, G. Sidorov, and R. M. A. Nawab, “Author profiling for age and gender using combinations of features of various types,” J. Intell. Fuzzy Syst., vol. 36, no. 5, pp. 4833–4843, 2019. doi: 10.3233/JIFS-179031. [Google Scholar] [CrossRef]

2. F. C. Hsieh, R. F. S. Dias, and I. Paraboni, “Author profiling from Facebook corpora,” in Proc. Int. Conf. Lang. Resour. Eval., Miyazaki, Japan, 2018, pp. 2566–2570. [Google Scholar]

3. G. Farnadi, J. Tang, M. de Cock, and M. F. Moens, “User profiling through deep multimodal fusion experimental results: User profiling,” in Proc. Int. Conf. Web Search Data Min., Marina Del Rey, CA, USA, 2018, pp. 171–179. [Google Scholar]

4. D. Radha and P. C. Sekhar, “Author profiling using stylistic and N-gram features,” Int. J. Eng. Adv. Technol., vol. 9, no. 1, pp. 3044–3049, 2019. doi: 10.35940/ijeat.A1621.109119. [Google Scholar] [CrossRef]

5. C. Peersman, W. Daelemans, and L. V. Vaerenbergh, “Predicting age and gender in online social networks,” in Proc. 3rd Int. Workshop Search Min. User-Gen. Contents, Glasgow, Scotland, UK, 2011, pp. 37–44. [Google Scholar]

6. A. Sboev, T. Litvinova, D. Gudovskikh, R. Rybka, and I. Moloshnikov, “Machine learning models of text categorization by author gender using topic-independent features,” Procedia Comput. Sci., vol. 101, pp. 135–142, 2016. doi: 10.1016/j.procs.2016.11.017. [Google Scholar] [CrossRef]

7. T. R. Reddy, B. V. Vardhan, and P. V. Reddy, “N-gram approach for gender prediction,” in Proc. IEEE 7th Int. Adv. Comput. Conf. (IACC), Hyderabad, India, 2017, pp. 860–865. [Google Scholar]

8. E. R. D. Weren et al., “Examining multiple features for author profiling,” J. Inf. Data Manage., vol. 5, no. 3, pp. 266, 2014. [Google Scholar]

9. T. K. Koch, P. Romero, and C. Stachl, “Age and gender in language, emoji, and emoticon usage in instant messages,” Comput. Hum. Behav., vol. 126, pp. 106990, 2022. doi: 10.1016/j.chb.2021.106990. [Google Scholar] [CrossRef]

10. M. A. Ashraf, R. M. A. Nawab, and F. Nie, “Author profiling on bi-lingual tweets,” J. Intell. Fuzzy Syst., vol. 39, no. 2, pp. 2379–2389, 2020. doi: 10.3233/JIFS-179898. [Google Scholar] [CrossRef]

11. A. Alajmi, E. M. Saad, and R. R. Darwish, “Toward an Arabic stop-words list generation,” Int. J. Comput. Appl., vol. 46, no. 8, pp. 8–13, 2018. [Google Scholar]

12. J. Soler-Company, and L. Wanner, “On the relevance of syntactic and discourse features for author profiling and identification,” in Proc. 15th Conf. Eur. Chapter Assoc. Comput. Linguist., Valencia, Spain, 2017, pp. 681–687. [Google Scholar]

13. A. Sittar and I. Ameer, “Multi-lingual author profiling using stylistic features,” in Proc. Working Notes FIRE, Gandhinagar, India, 2018, pp. 240–246. [Google Scholar]

14. S. Ouni, F. Fkih, and M. N. Omri, “Toward a new approach to author profiling based on the extraction of statistical features,” Soc. Netw. Anal. Min., vol. 11, no. 1, pp. 1–16, 2021. doi: 10.1007/s13278-021-00768-6. [Google Scholar] [CrossRef]

15. M. Wiegmann, B. Stein, and M. Potthast, “Overview of the celebrity profiling task at PAN 2019,” in Proc. Working Notes Conf. Labs Eval. Forum, Lugano, Switzerland, vol. 2380, 2019. [Google Scholar]

16. J. Silva et al., “A method for detecting the profile of an author,” Procedia Comput. Sci., vol. 170, no. 2020, pp. 959–964, 2020. doi: 10.1016/j.procs.2020.03.101. [Google Scholar] [CrossRef]

17. A. Nemati, “Gender and age prediction multilingual author profiles based on comments,” in Proc. Working Notes FIRE 2018-Forum Inf. Retr. Eval., Gandhinagar, India, 2018. [Google Scholar]

18. G. Kovács et al., “Author profiling using semantic and syntactic features: Notebook for PAN at CLEF 2019,” in Proc. Working Notes Conf. Labs Eval. Forum, Lugano, Switzerland, vol. 2380, 2019. [Google Scholar]

19. U. Sapkota, T. Solorio, M. Montes-Y-Gómez, and G. Ramírez-de-la-Rosa, “Author profiling for English and Spanish text: Notebook for PAN at CLEF 2013,” in Proc. Working Notes Conf. Labs Eval. Forum, Valencia, Spain, vol. 1179, 2013. [Google Scholar]

20. A. Grivas, A. Krithara, and G. Giannakopoulos, “Author profiling using stylometric and structural feature groupings notebook for PAN at CLEF 2015,” in Proc. Working Notes Conf. Labs Eval. Forum, Toulouse, France, 2015. [Google Scholar]

21. D. Estival, T. Gaustad, S. B. Pham, W. Radford, and B. Hutchinson, “Author profiling for English emails,” in Proc. Int. Conf. Pacific Assoc. Comput. Linguist., Melbourne, Australia, 2007, pp. 263–272. [Google Scholar]

22. H. A. Nayel, “NAYEL @ APDA: Machine learning approach for author profiling and deception detection in Arabic texts,” in Proc. Working Notes FIRE 2019-Forum Inf. Retr. Eval., Kolkata, India, 2019, pp. 92–99. [Google Scholar]

23. Z. Zhang and X. Xie, “Research on AdaBoost.M1 with random forest,” in Int. Conf. Adv. Emerg. Trends Comput. Technol., Chennai, India, 2010. [Google Scholar]

24. S. Mamgain, R. C. Balabantaray, and A. K. Das, “Author profiling: Prediction of gender and language variety from document,” in 2019 Int. Ceronf. Inf. Technol. (ICIT), 2019, pp. 473–477. [Google Scholar]

25. B. Khan, M. Arshad, and S. S. Khan, “Comparative analysis of machine learning models for PDF malware detection: Evaluating different training and testing criteria,” J. Cybersecur., vol. 5, pp. 1–11, 2023. [Google Scholar]

26. R. Asif et al., “Hyper-tuned convolutional neural networks for authorship verification in digital forensic investigations,” Comput., Mat. Continu., vol. 76, no. 2, pp. 1947–1975, 2023. [Google Scholar]

27. M. Nawaz et al., “Single and multiple regions duplication detections in digital images with applications in image forensic,” J. Intell. Fuzzy Syst., vol. 40, no. 6, pp. 10351–10371, 2021. [Google Scholar]

28. J. Abdul, I. Sajid, I. T. Manzoor, R. Amjad, A. B. Saeed and S. Tanzila, “Analytical analysis of text stemming methodologies in information retrieval and natural language processing systems,” in Inst. Electr. Electron. Eng. Inc., Nov. 2023. doi: 10.1109/ACCESS.2023.3332710. [Google Scholar] [CrossRef]

29. L. S. Zoya, L. Rabia, M. Hammad, and S. M. J. Nor, “Assessing urdu language processing tools via statistical and outlier detection methods on urdu tweets,” ACM Trans. Asian Low-Resour. Lang. Inf. Process., vol. 22, no. 10, pp. 1–31, 2023. [Google Scholar]