Fighting against social bots: The issue of identification

Dmitrii Pianov
05/02/2018

An algorithm that interacts with human users on social platforms.
Can be benign or malicious.
Malicious bots are used for variety of applications, including spam, manipulation of public discourse, promotion of people and ideas, etc.
Most of the malicious social bots hide their identity to avoid being detected.

Prevalence of social networks creates an opportunity for adversarial agents to obtain benefit by overflowing social networks with bots.
Breakthrough in computational power and NLP allows bot owners to avoid being detected by traditional heuristics.
There is a curious case of Cynk Technology - a company with no turnover or assets thar risen in value to $4.5 bil. overnight.

Leaving commercial issues aside, bot identification helps to avoid information manipulation and keep society healthy.
Bots are constantly evolving and so must our detecting capabilities.
Social bot detection is a good example why learning, rather than fitting, is an ultimate objective of machine learning.

Early research on bot detection mostly operated on domain knowledge and heuristic measures (Chu et al (2012)).
Recent studies use traditional supervised and unsupervised techniques to classify the accounts (Stukal et al (2017), Kudugunta and Ferrara (2018), Chavoshi (2016)).
Both textual and metadata are used as features.
There are no consensus which one is more important.

I propose a theoretical model to show that the bot network is sustainable.
I use character Bi-GRU RNN to extract features from tweet text.
Using a labeled dataset, I build a model to classify the origin of tweets using both textual features and metadata.

alt text

Bot owners decide on creation of new bots and investment in sophistication.
Sophistication affects the survival rate of old bots.
No investment in sophistication occurs if the cost of creating new bots is low.
Bot model is sustainable if and only if there is an increasing benefit of bots with longer lifetime.

I use the data from Creci et al (2017) that has more than 8 million tweets from variety of human and bot accounts.
Each entry consists of tweet text and metadata (number of mentions, replies, hashtags, etc.).
I sample 60000 tweets (30000 of human users, 30000 of one type of bots).
The tweet text is then fed into Bi-GRU RNN to produce 500-dimensional vector that encodes the characters within the tweet.
For classification, I use cross-validated traditional SVM with embedding vector plus metadata as a input vector.
2 and 4 PCA also estimated for robustness check.

alt text

	Precision	Recall	F1-score	support
Bot	0.85	0.88	0.87	5902
Human	0.87	0.85	0.86	5959
avg/total	0.86	0.86	0.86	11861

alt text

	Precision	Recall	F1-score	support
Bot	0.62	0.92	0.74	4891
Human	0.85	0.46	0.60	5044
avg/total	0.74	0.69	0.67	9915

A model shows 87% accuracy within the same domain of data.
PCA deteriorate the prediction to 50-55% accuracy which suggest no combination of features dominate the prediction importance.
Sampling bots from other domain substantially decrease the prediction rate to 68%.