Enhanced Twitter Bot

Detection: A Hybrid Model

Guide:

DR. KARTHIKEYAN D
INTRODUCTION:
• Twitter has become a pivotal platform with over 1.5 billion registered accounts, reflecting its widespread use and influence
across the globe.
• Montly active users are 365 Million.
• With recent changes making Twitter verification more accessible, it's now easier than ever to create a verified, monetized account.
• Using this feature bots are created with verified accounts, it has become a major issue as they contribute to spreading
misinformation, spamming, and manipulating public opinion.
• These bots are often designed to behave like real users, making them difficult to detect using traditional methods.
PROBLEM STATEMENT:
• The ease of creating monetized and even verified
accounts has exacerbated the problem, allowing bots to
imitate human behavior and post content that can
influence public opinion or generate illicit revenue.
• Existing bot detection methods often fall short in
accuracy and adaptability, failing to keep pace with the
evolving tactics of bot creators.
• This research aims to address this need by developing an
enhanced detection framework that integrates models
like Bot-DenseNet, SVM, and Random Forest to improve
detection accuracy and maintain the integrity of Twitter’s
ecosystem.
Literature Review:
S:NO Title, Author and Journal Merits Demerits

1 Title: Twitter bot detection The literature review effectively demonstrates the utility of traditional The review primarily focuses on traditional machine
using supervised machine machine learning models, including Decision Trees and K-Nearest learning techniques, missing advanced methods like
learning Neighbors, for Twitter bot detection. It introduces the 'Bag of Bots' deep learning. Its basic feature engineering and
Words model, enhancing detection accuracy. The clear methodology reliance on a single Kaggle dataset limit the study's
Author:
and evaluation metrics offer valuable, reproducible insights for generalizability. It also lacks a comparison with
A Ramalingaiah , S Hussaini, S improving bot detection systems. emerging techniques and does not address real-time
Chaudhari detection capabilities, reducing its practical

Year: 2021 applicability.

Journal:

Institute of Physics

2 Title: The paper provides a thorough literature review on Twitter bot The paper primarily serves as a literature review and
detection, offering a detailed categorization of bot types and detection does not present new empirical findings or
Bot Detection in Twitter: An
techniques. It effectively summarizes various approaches, including advancements in detection methods. It acknowledges
Overview.
graph-based, feature-based, and hybrid methods, and highlights the limitations such as the challenges of detecting
Author: potential of advanced models like BERT and GANs. This overview is sophisticated bots and outdated crowdsourcing
Salvador Lopez-Joya(B) , J. valuable for understanding the current landscape and identifying gaps methods but lacks specific recommendations or novel

Angel Diaz-Garcia, M. Dolores in bot detection research. solutions. The practical applicability of the reviewed
Ruiz, and Maria J. Martin- methods and their effectiveness against evolving bot
Bautista. strategies is not fully addressed.

Year: 2023

Journal:

Research Gate
3. Title: Twitter bot detection using deep learning. The research offers a comprehensive The research faces limitations due to the small
approach to bot detection by integrating LSTM dataset size (fewer than 3k accounts), which
Author: Adam Kenyeres , György Kovács.
and BERT models for tweet classification and constrains the training of deep learning
Year: 2022 combining these with classical machine models. Data augmentation techniques did
Journal: learning methods like Adaboost for account not yield expected improvements, possibly
classification. The study effectively utilizes due to complex patterns in the data.
Research Gate
advanced techniques such as data Additionally, hyper-parameter optimization
augmentation and hyper-parameter might have been suboptimal, affecting model
optimization, demonstrating significant performance, especially with augmented data.
improvements in accuracy with a combined
approach. The use of late fusion methods and
probability-based techniques enhances the
robustness of the classification system.
4 Title: Twitter Bot Detection Using Neural The paper presents BOTLE, an advanced Despite its strengths, BOTLE shows some
Networks and Linguistic Embeddings neural network model that leverages multiple performance gaps compared to top models
linguistic embeddings (word, character, POS, and incurs higher computational costs due to
Author: Wei and Nguyen
and named-entity) for effective Twitter bot the extensive use of embeddings. This
Year: 2023 detection. The model excels in recall, F1- increased resource demand may be
Journal: IEEE score, and accuracy, outperforming existing problematic for deployment on low-power
solutions and eliminating the need for manual devices. Additionally, the model's
feature engineering. Its efficiency and performance improvement comes with a
scalability make it a practical choice for real- trade-off in runtime and memory usage.
world applications.
5 Title: Twitter Bot Detection using NLP The paper introduces an innovative The GCN model's performance was
and Graph Classification approach to Twitter bot detection by limited by insufficient training data and
integrating classical machine learning challenges with real-time processing.
Author: Warada Jayant Kulkarni
with graph convolutional networks Additionally, high-dimensional feature
Year: 2023 (GCNs). It effectively combines vectors from BERT and spaCy can
Journal: SJSU Scholar traditional NLP techniques with impact model efficiency. The GCN
advanced graph-based methods to approach, while promising, fell short in
enhance detection accuracy. The use accuracy compared to classical
of BERT and spaCy embeddings methods, highlighting the need for
alongside classical ML models like KNN, more comprehensive training and
SVM, and Random Forest demonstrates diverse datasets.
high accuracy and F1 scores,
showcasing a versatile methodology.

6. Title: A Deep Learning Approach for The study introduces Bot-DenseNet, a The study encounters challenges with
Robust Detection of Bots in Twitter comprehensive solution for detecting overfitting, particularly with text
using Transformers Twitter bots by integrating both encoding methods like Pooling and
metadata and advanced NLP features. LSTMs, which affects validation
Author: DAVID MARTÍN-GUTIÉRREZ ,
By leveraging state-of-the-art performance. Additionally, the model’s
GUSTAVO HERNÁNDEZ PEÑALOZA,
Transformers like BERT and RoBERTa, performance may be inconsistent with
ALBERTO BELMONTE-HERNÁNDEZ ,
the model significantly improves bot different bot types or languages not
ALICIA LOZANO-DIEZ and FEDERICO
detection across multiple languages. covered by pre-trained NLP models.
ÁLVAREZ.
Its ability to combine text-based and Further exploration of newer
Year: 2022 metadata features into a unified vector Transformers and non-parametric tests

Journal: IEEE representation enhances accuracy and is needed to address these limitations
robustness, addressing limitations of and improve generalizability.
previous methods.