You are on page 1of 7

Deep Q-Reinforcement Learning with

Particle Swarm Optimization to


Strengthen Information Trust in social
media

Document Submitted by,


Udhayavannan A.
2022207033

Guided by,
Dr. S. Valli,
Professor,
College of Engineering, Guindy,
Anna University, Chennai.

Signature of Guide
Problem Statement
In today's world, where we have lots of information and news shared on social
websites, there's a big need for a system that can reliably check if news articles and
their sources are trustworthy. The problem at hand is to develop an effective
methodology to evaluate the trustworthiness and reliability of news content shared on
a social website by consider various factors, including the behaviour of users,
temporal patterns, and the influence of potential spam or malicious entities.
Phase 1:
On social websites, some users act dishonestly. They might pretend to be
someone else, or promote annoying stuff. To make the website safer and more reliable,
we need a way to spot these dishonest users and separate them from the honest ones.
These malicious actors engage in various deceptive activities, including spreading
false information, impersonating legitimate users, and promoting spammy content. To
mitigate the adverse effects of spam users and ensure the credibility of user
interactions, there is a need for an effective system to identify and categorize these
spam accounts reliably.
Phase 2:
In the world of social media, we see a lot of news and information. But not all
of it is trustworthy. Some news might be true, while others are not. The challenge is to
develop a systematic approach to assess the credibility of news content on a social
website. Using the insights from Phase 1, the goal is to provide users with accurate
and trustworthy information about the reliability of news articles and their respective
sources, thus enhancing the overall quality of user engagement with news content.
Related Works
Lingam [1] have proposed adaptive Deep Q-Learning for detecting Social Spam Users
who post spam content manipulate their behaviour by pretending as a normal user, and
perform phishing attacks. Experimentation was conducted using Social Honeypot
dataset. Precision, True Positive Rate, False Positive Rate, F-measure are the
performance metrics used for the project.

Josephine [2] have proposed Hybrid Particle Swarm Optimization with Deep Learning
Driven Sarcasm Detection (HPSO-DLSD) technique to identity sarcasm, a positive
sentence with negative meaning, on social platform which was considered to be
challenging problem in Natural Language processing domain. Experimentation was
conducted on Twitter and headline datasets. Recall, Precision, F-score, Accuracy are
the performance metrics used for the project.

Yang [3] proposes the idea of ranking meta path between 2 different users in a social
network by considering the direct influence, indirect influence and global influence of
user on others. It also assigns weightage to the speed range of influence parameter.
Experimentation was conducted on 4-area and Yelp dataset. Degree Centrality,
PageRank are the performance metrics used for the project.

Masood [4] proposes the user model evaluation with twitter accounts where the tweets
and features of various users are extracted and analysed. The classification process is
categorized as Fake content-based spammer detection, URL based spammer detection,
Spam in trending topics and Fake user identification using statistical data.
Experimentation was conducted on CREDBANK and PHEME dataset.

Wang [5] proposes a novel approach SentiDiff to merge textual information with
sentiment diffusion data within a supervised learning framework. The exploration of
sentiment diffusion was particularly focused on sentiment reversals where tweets and
their retweets exhibit differing sentiment polarities. Experimentation was conducted
on dataset obtained from Beijing Intelligent Starshine Information Technology
Corporation, a leading big data collection and mining service provider in China. Area
Under the Precision-Recall Curve (PR-AUC) was the performance metrics used for
the project.
Overall Architecture

Detailed Architecture Phase 1:


Detailed Architecture Phase 2:
Explanation
Phase 1 - Deep Q-Learning for User Behavior Analysis:
Data gathered from Twitter dataset extracted into social attributes, user
attributes, and tweet attributes. Social attributes are utilized to construct a user's social
graph, providing insights into their network and connections.
User attributes undergo Principal Component Analysis (PCA) to identify
essential behavioral patterns, while tweet attributes undergo text normalization,
tokenization, and stemming to extract meaningful information. These attributes
collectively form the state (S) in the Deep Q-Learning framework.
Deep Q-Learning model is then built to classify users based on their behavior,
effectively distinguishing between honest and deceptive users. Finally, the model's
performance is evaluated to ensure its effectiveness.

Phase 2 - Enhancing Deep Q-Learning with PSO:


PSO is a powerful optimization technique that enhances the Deep Q-Learning
workflow by facilitating the exploration of the Q-value space more effectively. Deep
Q-Learning model assesses user behaviour and distinguishes between honest and
deceptive users based on learned Q-values. PSO allows the model to adapt and find
optimal solutions more efficiently allowing the model to identify and converge
towards the most suitable Q-values for accurate decision-making.
References
1. Lingam, Greeshma & Rout, Rashmi & Somayajulu, Dvln, "Adaptive deep Q-
learning model for detecting social bots and influential users in online social
networks". Applied Intelligence, 2019, pp.3947–3964, doi: 10.1007/s10489-
019-01488-3.
2. A. Josephine, S. K. Maharana, M. A. A. Walid, T. T, M. S. Alam and M. Tiwari,
"Hybrid Particle Swarm Optimization with Deep Learning Driven Sarcasm
Detection on Social Media," 2022 International Conference on Automation,
Computing and Renewable Systems (ICACRS), Pudukkottai, India, 2022,
pp. 687-693, doi: 10.1109/ICACRS55517.2022.10029167.
3. Y. Yang, L. Zhou, Z. Jin and J. Yang, "Meta Path-Based Information Entropy
for Modeling Social Influence in Heterogeneous Information Networks," 2019
20th IEEE International Conference on Mobile Data Management (MDM),
Hong Kong, China, 2019, pp. 557-562, doi: 10.1109/MDM.2019.00119.
4. F. Masood et al., "Spammer Detection and Fake User Identification on Social
Networks," in IEEE Access, vol. 7, pp. 68140-68152, 2019,
doi: 10.1109/ACCESS.2019.2918196.
5. L. Wang, J. Niu and S. Yu, "SentiDiff: Combining Textual Information and
Sentiment Diffusion Patterns for Twitter Sentiment Analysis," in IEEE
Transactions on Knowledge and Data Engineering, vol. 32, no. 10, pp. 2026-
2039, 1 Oct. 2020, doi: 10.1109/TKDE.2019.2913641.

You might also like