ABSTRACT
Sex determination has been found interesting in forensic handwriting examinations and has been researched by scientists. The inclusion of the sex parameter as a supporting element in the examination of forensic handwriting while deciding belonging will increase the reliability of the results. In addition, it will help reduce the number of people to be investigated for a large group of suspects, both men and women. In this study, it was aimed to investigate the contribution of the ascender and descender parts of the letters to sex prediction by measuring them.
In line with this purpose, handwriting samples were collected from 50 female and 50 male participants by having them write 11 sentences containing the letters “b, d, f, g, h, k, t, y, p” at initial, medial, and end positions. The ascender and descender parts of these letters were measured in millimeters. Logistics, k-nearest neighbor (KNN), support vector machine (SVM) and artificial neural network (ANN) were selected and applied to these data.
The ascender and descender parts of these letters were measured in millimeters and statistically significant differences were found between male and female participants. The ascender parts of the “b, d, h, k, t” were determined to be statistically significantly longer in males. Accuracy rates are 0.65, 0.60, 0.71 and 0.82 for Logistics, KNN, SVM and ANN, respectively.
In our opinion, this result is promising. If the studies on this subject are increased, higher success rates can be achieved, and more contributions can be made to forensic handwriting examination.
Keywords: Sex prediction, forensic handwriting examination, artificial neural networks, data mining
INTRODUCTION
Handwriting is affected by many factors such as age, social habits, and biological factors, which is the starting point of studies aimed at classifying handwriting into demographic classes (1). Characteristic elements found in people’s handwriting have been investigated for sex determination (2).
In forensic handwriting examinations, belonging is determined by comparing the distinguishing features of handwriting. Huber and Headrick (2) reported that handwriting has 21 distinctive features (3).
In forensic handwriting examinations, the ability to classify the handwriting into demographic data such as age, sex, hand dominance, and nationality and then performing eliminations is of great benefit in practice (1). This classification can help forensic document examiners to focus on a particular category of suspects (4). Further, the classification of demographic information is regarded to be objective because it can be experimentally verified through quantitative results (5). Moreover, by processing all these demographic data separately, improved results can be produced for the determination and verification of the person who wrote the examined handwriting (4).
Although it is a two-class problem, sex determination through handwriting is difficult owing to a large number of variations. The fact that some men have feminine handwriting and vice versa cause significant differences in studies in this area (5).
Determination of certain characteristics through handwriting and the studies that classify them into sex information have generally attracted the interest of psychologists.
The oldest study on this issue was conducted by Goodenough in which 10 female and 10 male graduate students classified the handwritings of 115 high-school students into sex, and approximately 2/3 of the writings were correctly classified into sex (6).
The classification of handwriting into demographic data is conducted in two steps: Feature extraction and classification. The performance of the system has been reported to depend on the feature extraction step because extracted features are used to distinguish individuals (4).
In a study conducted by Marzinotto et al. (7) with two-layer clustering analysis on the writings created online, sex determination was not as good as in the classification of demographic information such as age; the vertical and connected text was observed together in males’ writings whereas in females’ writings only one or the other form was observed.
In the study conducted by Hamid and Loewenthal (8), handwriting samples were collected from 30 subjects (16 females and 14 males) in both English and Urdu languages, and 25 examiners were asked to classify them into sex; the accuracy was found to be approximately 68%. Multiple analysis of variance was used in this study, and it was reported that language is not an important source of variance in terms of sex. In other words, approximately the same results were obtained in both languages.
In a study conducted by Binet (9), writing samples belonging to a total of 180 participants, 91 male and 89 female, 2 graphologists, and 15 people who were disinterested in such business were asked to classify them into sex. Results indicated that one of the experts classified 78.3% correctly, and 10 disinterested examiners classified correctly at a percentage ranging between 65.9% and 72% (10).
Many features such as the size of the letters, formation features, line, and word spacing, dotting of the letter “i”, inclination, and slant of the writings were determined by Kumar et al. (3) using the Z test. They investigated whether a statistical difference existed between male and female writings and reported that a significant difference existed in terms of sex.
Referring to the study conducted by Briggs with 100 people, Tomai et al. (11) reported that the former had concluded that distinguishing male and female writings was not possible. However, one of the frequently studied topics regarding demographic features in the literature is sex determination, and these studies have concluded that a strong relationship exists between sex and certain handwriting features (6,8).
In the literature, there are studies aimed at determining certain features that are thought to be important in sex classification and investigating their relation to sex (4,10). Moreover, studies on automatically classifying sex through artificial neural networks (ANNs) and various image processing techniques (1,5,11-16) using software systems exist.
In the literature research, it was determined that the classification rates of the studies on sex prediction were in the range of 61.93%-85.7% (1,4,5,8,13,15-24). Some of these studies are given in Table 1.
In the studies of Kumar et al. (3), they showed that the handwriting of men and women can be distinguished from each other with the features obtained from the sentences of 200 people. In the same way, Hamid and Loewenthal (8), in their study with writing samples from 30 people, made a sex estimation using 25 experts and showed that this is possible. In addition, in the study of Al Maadeed and Hassaine (4), they stated that they could use the automatic sex handwriting classification system with a certain sentence written to the students. However, Liwicki et al. (15) also used a mathematical model to classify sex with an accuracy of 67.5% in the data they collected. In the study of Riza et al. (19), on the other hand, they estimated the sex with 76% accuracy by using 49 variables with effects such as height, pressure, margin etc. of certain words obtained from 75 people.
These studies show that by looking at people’s handwriting, sex can be determined and an automatic system can be established.
In the articles made using the QUWI, MHSH, IAM, KHATT, ICDAR2013 and CEDAR databases, it is aimed to perform sex classification with different structuring in the pattern recognition method (1,3-5,13,15-18,20,23,24).
Success achieved varies between 68.90% and 82%. Using combinations different patterns (LBP, HOG, GLCM, SFTA) ANN, S, DT, KNN and RF analyzes were implemented in best success. Apart from this, any success rate varies between 68.90% and 77%. You must be a good computer scientist for these analyses. It is difficult for these non-expert scientists to perform these analyzes and expert systems are very demanding. In this study, it is aimed to develop an easier, faster and higher classification success method with simpler data.
While determining the belonging in forensic handwriting comparisons, conducting a research on the sex of the writer will both facilitate the comparison and increase the reliability of the conclusion reached. Therefore, developing methods to be used in sex prediction and increasing their accuracy will greatly advantage the forensic document examination society.
In the present study, parts of the same letters in the same words in identical texts written by 50 female and 50 male participants were measured aiming at revealing the differences between them as well as to investigate the success rate in sex determination using data mining.
MATERIALS AND METHODS
In line with the aim of the study, 11 sentences in Turkish containing the letters “b, d, f, g, h, k, p, t, y” at initial, medial, and end positions were written by 50 female and 50 male individuals who were higher education students and/or graduates. Cursive samples were not included in this study. Handwriting samples from people were collected on A4 paper. These A4 papers were scanned as a whole at 300 dpi and saved in jpeg format. Then opened in A4 size in Photoshop. Parts of letters were measured at x300 magnification (Figure 1).
The body and extension parts of these letters were measured in millimeters in Adobe Photoshop CS6 by three different researchers, as shown in Figures 1, 2. Whether a statistical difference existed between the extensions of the letters written by male and female participants was evaluated by an Independent sample t-test.
The most commonly used methods for classifying variables in data mining methods are support vector machine (SVM), artificial neural network, logistic regression (LR) analysis, and k-nearest neighbor (k-NN). In different studies, these methods have been reported to be superior to one another according to the type of data used. Twenty-seven different measurements of the letters were obtained at initial, medial, and end positions in the present study. As the number of subjects was 100, variables that showed significant differences by sex were included in the selection of variables. Accordingly, in this study, sex determination was performed based on the stroke lengths of the letters “b, d, h, k, t,” which were determined as significant in the independent sample t-test and of the letter “p,” which was significant at the medial position, using the aforementioned four methods.
Support Vector Machine
SVM is a machine learning method developed in the late 1960s by Vladimir Vapnik and Alexey Chervonenkis and is primarily based on statistical learning theory. The SVM method has been used frequently in recent years, especially in data mining, for classification problems in datasets where patterns between variables are unknown (25). This method is basically intended as a linear classifier in solving two-class problems and then generalized to the solution of linearly inseparable or multiclass classification problems, and it has been widely used in the solution of these problems. When applied to linearly separable data, SVM aims to select the line that will make the margin the highest among an infinite number of lines that can separate the data. In the case of linearly inseparable data, SVM transforms the original data into a higher dimensional space with a mapping method and tries to find the linear separating hyperplane that can be optimized to classify the data (26). Models use kernel functions for this purpose. The kernel function of choice affects the performance of the system, and different results can be obtained with different kernel functions.
Below is an illustration of how the SVM method works in a two-dimensional space (Figure 3).
Artificial Neural Networks
ANNs are computer systems inspired by the characteristics of biological nervous systems (information generation, description, estimation, etc.) (29). As in the biological nervous systems, ANNs are formed by a combination of cells. Generally, ANN architecture is defined in three layers: input, intermediate or hidden, and output (27).
More than one intermediate hidden layer can be present in a network. To date, it has not been determined how many hidden layers should be used in an ANN and how many nerve cells should be in each hidden layer. The solution to this situation, which varies according to the problem, has been through trial and error (28,29). A network with several hidden neurons cannot distinguish complex patterns because it can only make linear predictions. Moreover, a large number of hidden neurons prevents the network from generalizing (28,30). Because additional layers exist between the input and output layers in solving non-linear problems, the network architecture becomes multilayered, as shown below (Figure 4).
The backpropagation algorithm is widely used as the learning algorithm of ANN in multilayer feedforward networks.
In backpropagation networks, data is processed from the input layer to the hidden layer and then to the output layer. The purpose in of obtaining values close to the targets as output is to find the optimal weights.
In the present study, Levenberg-Marquardt learning algorithm is used to adjust weights in multilayer feedforward networks. In the hidden and output layers, logistic sigmoid nonlinear function (logsig) and linear transfer function (purelin), respectively, have been used as activation functions.
Logistic Regression
LR analysis is a regression method that helps classification and assignment. In most biological, health, and socio-economic studies conducted to reveal cause and effect relationships, some of the variables examined comprise two-level data such as positive-negative, successful-unsuccessful, or yes-no. In this way, if the dependent variable comprises two-level or multilevel categorical data, LR analysis has an important place in examining the cause-effect relationship between the dependent and independent variables (31). In the LR analysis, one of the purposes of which is classification and the other is to investigate the relationships between dependent and independent variables, the dependent variable takes categorical values. Independent variables can be continuous or categorical variables. Besides being applicable when the dependent variable is a two-level variable, such as 0 or 1, or a discrete variable with more than two levels, its mathematical flexibility and easy interpretability increase the interest in this method (32).
K-nearest Neighbor (k-NN) Algorithm
K-NN algorithm, T. M. Cover ve P. E. K-NN algorithm, proposed by Hart, is a classification method by which the nearest neighbor of the class in which the sample data point is present is determined according to the k value (33). This algorithm is one of the best-known, old, simple, and effective pattern classification methods and is popularly used among machine learning algorithms (34). Classification of objects is an important research area and is applied in a wide variety of fields such as pattern recognition, data mining, artificial intelligence, statistics, cognitive psychology, medicine, and bioinformatics (35). The k-NN algorithm is especially preferred in classification applications owing to its advantages such as easy applicability and resistance to noisy training data. Despite these advantages, it also brings some disadvantages such as processing load increases with the number of datasets and variables’ performance being affected by parameters or features such as the number of neighbors, distance criteria, and the number of variables (33). k-NN calculates the probability of data that is considered to belong to the class of its neighbors according to the status of its closest neighbor (25). The following figure shows to which class the data indicated with an asterisk will belong in cases k=3 and 6 (Figure 5).
Classification Criteria
The confusion matrix evaluates the performance of classification models and tells us how good our classification model is while making predictions on test data. The rows of the matrix contain actual values whereas the columns contain predicted values. Predicted values are values calculated by the model, and actual values are true values for the given observations. With the help of the confusion matrix, different parameters such as accuracy and precision can be calculated for the model. These values indicate how effective the used methods are. True positive (TP), true negative (TN), false positive (FP), and false-negative (FN) values in the confusion matrix are used to calculate the following values. Because our aim in this study was to correctly predict females, the correct prediction of females in the confusion matrix was TP, and the correct prediction of males was FP; moreover, the incorrect prediction of females was TN, and the incorrect prediction of males was FN.
In this study, accuracy (ACC), error rate (ERR), precision (PREC), sensitivity (SENS), specificity (SPEC), F-measure (FM), Youden’s index (YI), kappa (k) statistics, true negative rate (TPR), false positive rate (FPR), and receiver operating characteristic (ROC) area values were used.
RESULTS
Statistical Analyses
The data collected from 100 different people were measured by three different people, and measurement error was examined using the Friedman analysis. Consequently, no statistically significant difference was observed among the measurements performed by the three different researchers. Therefore, the analysis was carried on with the data measured by one person. Extensions of the letters “b, d, h, k, t” were found to be statistically significantly longer in males. No such difference was found in the letters “p,” “f,” “y,” and “g” (Table 2).
Here, 9 letters have a total of 27 measurements, including the initially middle and end. This means 27 different variables. In prediction models; If the number of variables is high and the number of samples is few, “overfitting” occurs. To avoid this, one of the variable selection methods such as random forest, chaid analysis or variable reduction methods such as principal component analysis can be used. Here, the independent sample t-test was preferred because of the grouping. When the sample size of the groups was 50 and the alpha value was 0.05, the effect size was found to be 0.80 when examined. This effect size is sufficient. For this, the G power software was used. Accordingly, it is seen that the number of samples used in the study is sufficient. A prediction model has been established with the data we have. Thanks to this model, it can be determined with 82% accuracy whether a person writing the same words is male or female.
Data Mining Analyses
With the inclusion of letter “p,” which was determined as significant at the medial position, to the letters “b, d, f, h, k, t”; ANN, k-NN, SVM, and LR were applied, and the most successful result was obtained with ANN (Tables 3 and 4). The analyses were conducted with five cross-validations. The confusion matrix obtained as a result of the analyses is shown in Table 2.
With ANN, 43 out of 50 females were correctly predicted as females, whereas 39 of the males were correctly predicted as males (Table 3). The positive predictive value for ANN was 0.86, and the negative predictive value was 0.78. The analysis with the least error was ANN with an error ratio of 18%. ACC value for ANN was 0.82. The highest Kappa value was obtained with 0.64 in the ANN analysis. The LR (+) value obtained as a result of the analysis was best in the ANN analysis with 5.23. Considering the ROC test, the best explanation was approached with the ANN analysis. Consequently, there is an 88% probability of determining a person as a male or a female when the measurement values of that person are known (Table 4).
Correct Classification Rate (Accuracy): The closer the correct classification rate to 1, the higher the performance of the test. When this value is below 0.50, the classification made by the test performed can be said to be by chance. Herein, the highest correct classification rate was in ANN (82%), and the lowest was in the KNN analysis (60%).
Kappa Coefficient: It is a coefficient that provides information about reliability by correcting chance agreement that occurs solely by chance. This coefficient takes a value between 0 and 1. A value of 0-0.39 implies poor agreement, 0.40-0.75 good agreement, and 0.76-1.00 perfect agreement. Herein, the Kappa value varied between 0.20 and 0.64. The analysis with the lowest agreement was KNN, and the analysis with the highest was ANN. Accordingly, a good level of agreement existed in the assignments performed using the ANN and SVM analyses. Conversely, a weak agreement existed in the assignments performed using KNN and LR.
Likelihood Ratio: Two likelihood ratios, positive and negative, exist. LR (+) indicates the number of correct positives the model gives versus an FP. The higher this ratio, the better it distinguishes the positive state. LR (−) defines the number of false negatives a test gives for each TN. The smaller this ratio (closer to zero), the better the negativity success of the test. The fact that it is equal to 1 in both the likelihood ratios indicates the situation where the test is the most unsuccessful. The LR (+) value obtained as a result of the analysis was best in the ANN analysis with 5.23. SVM, LR, and KNN followed ANN, respectively. The worst result was obtained in the KNN analysis. The smallest LR (−) value was obtained in the ANN analysis. Similarly, SVM, LR, and KNN followed ANN, respectively.
Youden Index: It gives an overall assessment of the performance of the test and is used for comparing multiple tests. The test result is desired to be close to 1.
In the study, Kappa coefficient and Youden index values LR, KNN, SVM and ANN were found to be 0.30, 0.20, 0.42, 0.64, respectively. Although the values are the same, these analyzes are included in the study because there are differences in terms of interpretation.
It is sometimes difficult in a study to decide which of the methods used had the best performance because while the sensitivity analysis result is high in some methods, the specificity rate may be high in others. For this reason, combined criteria such as correct classification rate, LR, and odds ratio, which are obtained by combining sensitivity and specificity values, are used to compare the performance of different methods. The results are as follows.
Sensitivity shows the percentage of actual positives identified as positive in the results of the developed test. The ratio obtained as a result of the test is desired to be close to 1. Among the performed analyses, the ANN analysis identified actual females as females with 80% success. The success of the KNN analysis was 60%.
Specificity indicates the percentage of actual negatives identified as negative in the results of the developed test. The ratio obtained as a result of the test is desired to be close to 1. As with sensitivity, the ANN analysis achieved the most successful results in specificity as well predicting actual males as males with 85% success.
Positive predictive value gives the correct prediction of women, and it was 0.62 with KNN, 0.64 with LR, 0.74 with SVM, and 0.86 with ANN.
The negative predictive value gives the probability of negative assignments made by the applied test being actual negatives and is desired to be close to 1. As a result of the analysis, correct classification ratios were 0.58 with KNN, 0.66 with Logistics, 0.68 with SVM, and 0.78 with ANN.
Error, which is the rate of incorrect classification, was the highest in KNN at 40%, and the lowest in the ANN method at 18%.
The optimal cut-off point values that distinguish male and female status can be determined by the ROC curve analysis (36). With ROC analysis, the correct prediction rate of a test is measured by the area under the curve (AUC). The AUC value indicates the overall accuracy of the test. The values of 0.90-1.00 indicate perfect accuracy, 0.80-0.90 good accuracy, 0.70-0.80 moderate accuracy, 0.60-0.70 poor accuracy, and below 0.60 imply that the test is not useful (37).
The ROC curve (graph) is obtained connecting the sensitivity results obtained according to all cut-off values marked on the y-axis and the specificity results marked on the x-axis. The value, denoted as AUC at the end of the analysis, represents the “area under the curve,” and the determination value increases as it approaches 1. The AUC value obtained as a result of the analysis varied between 0.60 and 0.88. Herein, the best explanation was approached with the ANN analysis. Consequently, when the measurement values of a person are known, there is an 88% probability of determining the correct sex of that person.
DISCUSSION
In forensic handwriting examinations, if more than one person is examined, the ability to determine the sex of the writer is of great benefit. This situation will contribute to the reliability of the study by acting as an additional determination factor and thus will reduce the workload by eliminating individuals of different sexs and will set the ground for faster results. For this purpose, classifying the handwritings of males and females is necessary. Several studies exist on this subject in the literature (1,5,13,18,23,36,38-40). In these studies, the probabilities of handwriting belonging to a female or male participant were determined using various texts and measurement techniques. Generally, this rate remains at approximately 70%.
In the present study, 50 female and 50 male individuals were asked to write 11 sentences in the Turkish language containing the letters “b, d, f, g, h, k, p, t, y.” In the examination, it was determined that the ascender and descender parts of the letters “b, d, h, and k” were statistically significantly higher in males than in females. No such difference was detected in the letters “p,” “f,” “y,” and “g.” The extensions of these two types of letters are made considerably long in both males and females compared to their bodies.
The measurement process was conducted by three different researchers, and no statistically significant difference was found among their measurements, which shows that the measurement process is repeatable.
The data obtained in the measurement of ascender and descender parts were used to determine the probability of handwriting belonging to a female or male participant using the data mining techniques, i.e., KNN, SVM, LR, and ANN. Moreover, the method that made the best predictions was also investigated.
Although the success rate in studies on sex prediction using SVM varies between 48.9% and 77.98% (5,13,16,23,38,40), the correct classification rate obtained in this study was 71%.
Despite the success rate in studies on sex prediction with ANN varies between 55% and 74.7% (1,5,13,39), a correct classification rate of 82% was obtained in the present study.
In the sex prediction study using global features conducted by Ibrahim et al. (22), the ROC value obtained from the feature type with the highest accuracy value was found to be 0.658 whereas the ROC value in the prediction made using local properties was 0.534. In the analyses performed in the present study, the ROC was found to be 0.88 with ANN and 0.77 with SVM.
The success rate was reported to be 74% in a study conducted by Sesa-Nogueras et al. (21) to predict sex with dynamic features in handwritings written on tablets. Similarly, in the study conducted by Liwicki et al. (15) on handwritings collected online using different SVM types, the maximum correct classification rate obtained was 62.9%, which is approximately 10% below that obtained in the present study. It has been reported by Erbilek et al. (38) that sex classification was performed with 75% accuracy through handwritings collected online using the SVM classifier.
Study Limitations
One of the limitations of this study is that it has printed text. Additional work is needed for cursive text. In this way, it can be evaluated whether there is a difference between cursive and printed letters.
One of the study’s limitations is that the slant and other spatial-geometric features were not included in the analysis.
For this reason, research on sex estimation is carried out in Handwriting examinations.
It will be helpful to compare the results by repeating similar studies, taking into account the slope and without taking into account the slope in future studies.
Another limitation is the precision of the measurement technique. In this study, a model was established with LR, SVM, k-NN and ANN methods. Among these methods, ANN came to the fore with the best prediction rate. The validity of the ANN model can be determined by comparing it with the words of known sexs to be obtained from a Forensic case. Since this study was in an experimental environment, the model was run on the same words. In this study, it was not investigated whether the correct classification was made by having the same people write different words with the letters used in the model. This can be examined in another study.
The success rate of the method used in the present study will further increase with the inclusion of additional dynamic features in handwritings written on tablets.
CONCLUSION
In this study accuracy rates are 0.65, 0.60, 0.71 and 0.82 for Logistics, KNN, SVM and ANN respectively. The results showed that the model developed using ANNs achieved significant success in sex prediction. The results obtained from this study were higher than those obtained in other studies in the literature. The biggest difference of this study from other studies is that it can predict with higher accuracy without pattern recognition; however, not all prior studies employed pattern recognition. The high accuracy rates achieved in this study without pattern recognition indicate that better rates will be achieved when pattern recognition is used. In other studies with artificial neural networks, databases such as IAM and KHATT are generally used and the accuracy rate varies between 60-80%. In this study, 82% accurate classification rate was obtained faster with a simpler measurement method. In our study, 9 letters, 27 variables (initially, in the middle, at the end), which are the extensions, were selected by choosing those that differ according to sex. For this reason, it showed a better performance than the others. High accuracy rates without pattern recognition in this study indicate that higher rates will be achieved when pattern recognition is used. In future studies, a more detailed distinction can be made by adding variables such as age and hand used to the model. Based on studies using similar methods, the result we obtained is promising but in need of improvement for its application to forensic cases.
ETHICS
Ethics Committee Approval: Approval for the current study was granted by the Istanbul University-Cerrahpaşa, Social and Human Sciences Ethics Committee (approval no: 2019/71, dated: 15.11.2019).
Authorship Contributions
Concept: G.Ç., Design: D.Ö.K., G.Ç., Data Collection or Processing: D.Ö.K., Y.K., T.Ü.K., Ö.K., İ.D., Analysis or Interpretation: D.Ö.K., Y.K., Ö.K., İ.D., G.Ç., Literature Search: D.Ö.K., Y.K., T.Ü.K., İ.D., Writing: D.Ö.K., Ö.K., İ.D., G.Ç.
Conflict of Interest: No conflict of interest was declared by the authors.
Financial Disclosure: The authors declared that this study received no financial support.
