The performance of soft computing techniques on content-based SMS spam filtering
Content-based filtering is one of the most widely used methods to combat SMS (Short Message Service) spam. This method represents SMS text messages by a set of selected features which are extracted from data sets. Most of the available data sets have imbalanced class distribution problem. However...
Saved in:
| 主要作者: | |
|---|---|
| 格式: | Thesis |
| 语言: | English English English |
| 出版: |
2015
|
| 主题: | |
| 在线阅读: | http://eprints.uthm.edu.my/1496/2/WADDAH%20WAHEEB%20HASSAN%20SAEED%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/1496/1/24p%20WADDAH%20WAHEEB%20HASSAN%20SAEED.pdf http://eprints.uthm.edu.my/1496/3/WADDAH%20WAHEEB%20HASSAN%20SAEED%20WATERMARK.pdf |
| 标签: |
添加标签
没有标签, 成为第一个标记此记录!
|
| 总结: | Content-based filtering is one of the most widely used methods to combat SMS (Short
Message Service) spam. This method represents SMS text messages by a set of selected
features which are extracted from data sets. Most of the available data sets have
imbalanced class distribution problem. However, not much attention has been paid to
handle this problem which affect the characteristics and size of selected features and
cause undesired performance. Soft computing approaches have been applied successfully
in content-based spam filtering. In order to enhance soft computing performance,
suitable feature subset should be selected. Therefore, this research investigates how
well suited three soft computing techniques: Fuzzy Similarity, Artificial Neural Network
and Support Vector Machines (SVM) are for content-based SMS spam filtering
using an appropriate size of features which are selected by the Gini Index metric as
it has the ability to extract suitable features from imbalanced data sets. The data sets
used in this research were taken from three sources: UCI repository, Dublin Institute of
Technology (DIT) and British English SMS. The performance of each of the technique
was compared in terms of True Positive Rate against False Positive Rate, F1 score and
Matthews Correlation Coefficient. The results showed that SVM with 150 features
outperformed the other techniques in all the comparison measures. The average time
needed to classify an SMS text message is a fraction of a millisecond. Another test
using NUS SMS corpus was conducted in order to validate the SVM classifier with
150 features. The results again proved the efficiency of the SVM classifier with 150
features for SMS spam filtering with an accuracy of about 99.2%. |
|---|