Deception detection with machine learning: A systematic review and statistical analysis.

Area of Science:

• Computer Science
• Artificial Intelligence
• Psychology

Background:

• The field of Machine Learning (ML) for deception detection has seen significant growth over the last decade, resulting in a complex landscape of methodologies and findings.
• Identifying clear trends, successful strategies, research gaps, and opportunities for novel contributions in this domain can be challenging for researchers.

Approach:

• This literature review systematically analyzed 648 articles retrieved from major databases (ACM, IEEE, Scopus, Web of Science) following the PRISMA protocol.
• A final corpus of 81 documents was summarized using mind maps, with metadata encoded for statistical analysis using Python and Jupyter Lab Notebooks, all publicly available.
• The five most prevalent ML techniques identified were Neural Networks, Support Vector Machines, Random Forest, Decision Tree, and K-nearest Neighbor.

Key Points:

• Deception detection performance varied widely (51%-100%), with 19 studies achieving accuracy above 0.9.
• While monomodal approaches exist, bimodal and multimodal strategies are increasingly trending, often yielding higher accuracy in deception detection.
• Research predominantly focuses on English language data (75%), with limited exploration of linguistic features across diverse languages and cultures.

Conclusions:

• A critical limitation is the scarcity of labeled datasets derived from real-world scenarios for training ML models.
• There is substantial opportunity for developing novel ML approaches for deception detection, particularly those focusing on languages and cultural contexts beyond English.
• Future research should prioritize the creation of new, labeled, multimodal datasets to advance the accuracy and generalizability of AI-based deception detection systems.