Cyber Bullying Detection Using Social Network Analysis

Social Netwrok and Textual Features

 

In Figure 1, a comparison of the average values obtained for different features in the two categories as follows: mean(cyberbullying)−mean(non−bullying)\ mean(cyberbullying). As can be seen, the values for 11 out of 18 features showed a difference of at least 20%. These differences indicate that both social and textual features could indeed be useful for classification via various machine learning techniques.

 

 

Machine Learning

 

We used 70% of the total number of messages (both bullying and non-bullying messages) as our training data set. The remaining 30% were used as the testing data set. For accuracy, we had 30 groups of training and testing data sets which were randomly produced. Filter ’SMOTE’ and ’Randomize’ were used to keep the number of cyber bullying and non-bullying instances equal in the training data set. The testing data set was kept the same while testing across different rounds.

 

We used Weka 3.0 as the implementation tool for this work and tried multiple well-known classification algorithms including J48, Naive Bayes, SMO, Bagging and Dagging. 

 

In table 2, we found a clear trend of increasing classifier performance as we shift from ‘textual features’ models to ‘composite’ models in terms of both ROC and TP rates. In terms of the performance of different classification algorithms, Naive Bayes demonstrates much better results than other methods in TP rates. For ROC area-under-the-curve, Dagging has a better performance than the other classification algorithms.

 

To take ROC data and its best performance algorithm Dagging as an example, in Figure 3, it is obvious that the classifier with ‘all features’ (Standard Error = 0.0078) has a higher value than classifiers with ‘social features’ (Standard Error = 0.0087) and ‘textual features’ (Standard Error = 0.0079).