2016 DI-Detecting predatory conversations in social media by deep Convolutional Neural Networks

[SarahSalamati]

Main problem: Law enforcement organisations can take proactive measures through early identification of predatory activities in cyberspace thanks to automatic detection of predatory communications in chat logs.

Existing work: To address this issue domain, this research provides a classifier based on Convolutional Neural Network (CNN). In terms of classification performance, as assessed by F1-score, the proposed CNN design surpasses existing classification approaches often used in this area, such as Support Vector Machine (SVM) and conventional Neural Network (NN). In addition, our experiments show that using existing pre-trained word vectors are not suitable for this specific domain.

Inputs: Conversation in chat log in social media

Outputs: The duty is essential for the quick detection of prospective predators who harm children and young people. It demonstrated how text mining techniques might reduce search space by highlighting the discussions that are more likely to be predatory.

Data Set: PAN-2012 competition

Method:

Experimental Setup: Graphical Processing Units (GPUs) take use of the parallel nature of neural network-based learning models by speeding execution since parallelism is implemented at the hardware level. Therefore, we made use of a high-performance computing cluster that comprises several Tesla K80 GPUs from NVIDIA. One K80 GPU with 24 GB of memory and 2496 processing cores was used to run the programmer.

Baselines: It starts with the timeline of applying classification algorithms to the OPI in the past and ends with a brief introduction to corresponding frameworks that are commonly used in practice. Online predator identification with classification Pendar (2007) --- used a weighted k-Nearest Neighbor (k-NN) classifier (Kontostathis, 2009)-- ChatCoder1 (Mcghee et al 2011)- ChatCoder2 The system used a rule-based approach in conjunction with decision trees and k-NN classifier as an instance-based learning method (Eriksson and Karlgren, 2012) --- Entropy-based Classification (Kang et al., 2012) -- k-Nearest Neighbor (Villatoro-Tello et al., 2012) - traditionnel Neural Networks (Morris, 2013) -- Support Vector Machine Escalante et al. (2013)- using chained-classifiers based on adapting a psychological hypothesis Bogdanova et al. (2014) -- enriched the lexical features Cano et al. (2014)- by adding high-level features such as emotions, neuroticism, and psychological aspects. Finally, Ebrahimi et al. (2016) - one-class SVM classification algorithm.

Results:

Code: The code of this paper is available.

Presentation: There is no available presentation for this paper.

[SarahSalamati]

@hosseinfani please review the summary of the article.

[hosseinfani]

Hi @SarahSalamati , The summary is copy-paste in some parts. Please revise. Also, why this method outperforms is not clear.

Btw, what are the gaps of this work. How can we improve this work?

[SarahSalamati]

Hi @hosseinfani As I discussed about the article, I found many articles that finally we decided to go through the Thesis of Ebrahimi Etl, now I am working on the Thesis.