Use of Compression Analytics to Detect Deception

This paper describes the development and application of compression-based machine learning (ML) combined with part of speech (PoS) analysis to detect deception. Multiple data sets, including research associated with papermills, social media, and online reviews were used to test these methods. Our research is especially concerned with achieving reliable performance with minimal parameters and data preprocessing. The core idea draws from scientific literature suggesting that subtle, yet predictable linguistic cues can be indicative of intentionally deceptive writing. Part of speech parsing was chosen as a means to represent how the person was communicating as opposed to the content of the writing. The contribution of this paper beyond existing literature is demonstrating the robustness of the application of compression-based ML across a variety of data sets and using explainability to validate that the results are consistent with the underlying psychology. These methods have relatively few parameters and are easy to apply while still providing the best performance among the methods we tested. We present the results of studies over six data sets. In addition to the classification results, we also perform an explainability analysis that shows what parts of speech contributed the most to the performance. for the compression-base ML. These results are largely consistent between data sets and match the underlying psychology research.