Artificial Life and Evolutionary Computation by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9783030450168
Publisher: Springer International Publishing

4.2 Troll Detection

In this section, we compare the results of troll detection obtained through classifiers trained using the original dataset with those obtained after zI-based preprocessing of the same data.

The zI analysis identified 161 groups of trolls and 143 groups of non-trolls. Considering the stochastic nature of HyReSS, it should be noticed that its outcome was stable over five repetitions, repeatedly finding the same groups. The preprocessed dataset is composed of the centroids of each group of trolls and non-trolls detected using the zI-based approach. This preprocessing step aims at polishing the original dataset by substituting groups of users behaving similarly to the prototype which is closest to their dynamical behavior. After preprocessing, the final dataset contains 304 instances (the 161 centroids corresponding to the groups representing trolls, along with the 143 centroids of groups representing ordinary users).

The effect of this preprocessing on the detection of troll and non-troll users has been tested using four different classifiers: Random Forest (RF), Naïve Bayes (NB), Sequential Minimal Optimization (SMO), and K-Nearest Neighbors (KNN). Table 1 shows the accuracies obtained by training the classifiers on the original and on the preprocessed dataset.

To take into consideration the intrinsic stochastic nature of Random Forest and Sequential Minimal Optimization, we repeated all experiments using such classifiers five times; in the table, we report the average accuracy over the five runs.Table 1.Troll detection accuracies obtained by training different classifiers on the original dataset and on the preprocessed one.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.