Discussion

Received results of predicting whether a person is lying or telling the truth were surprisingly good. Some models obtained accuracy more than 90%, which is incredibly good result. On the contrary, in the original article highest average accuracy was 63% (combining gaze, EEG and audio/video data). [1]

The reason for our good results is the way our train and test data were formed. Those were created randomly choosing from the whole data set. Meaning that one persons observations could end up in both test and train data. In the real world different people’s EEG data can vary a lot, so model trained on one subset of people might not preform well on completely new people. Therefore, it would have been better if we had chosen train and test such that there is no intersection of people.

In addition, one of the team members tried to guess based on the videos whether the person was telling the truth or not. Every recorded trial for 21 different subjects was watched. Guessing whether a person is lying wasn’t an easy task and therefore were the answers quite random. Final accuracy was ~67.5%, which is higher than expected. The reason behind it is that every observed person had to describe same pictures (some more, some less pictures). This affected team member’s predictions a lot as different people described some objects similarly which indicated that they were talking about the original picture (meaning telling the truth).