II.2 AI-Based Interaction Analysis between Humans (and other Living Creatures)
The analysis process to predict human behavior from communication logs through machine learning, NLP, and social network analysis follows four steps. These four steps are introduced in this section. We will not discuss general principles of machine learning, NLP, and social network analysis, as there are many excellent books about these topics. Rather, this section will outline how to use these techniques for predicting human behavior by analyzing archives of traces of human-to-human and human-to-other-living-creatures interaction such as e-mail or GPS sensor data. The aim is to find general patterns of human behavior indicative of future actions. Learning about these patterns, and then analyzing past behavior and comparing it with desirable behavior (“the best against the rest”) will change future behavior towards better performance and happiness.
The general process consists of four steps (Figure 20):
- Collecting the data from body sensors and communication archives. This means for example tracking heartrate values, or accelerometer values of a smartwatch, or collecting the mailbox of a person.
- Converting the data into features or variables suitable for machine learning. This means for example converting the x/y/z coordinates of an accelerometer into an energy vector, or calculating network centrality measures of the e-mails of all people in a mailbox.
- Determining the “ground truth” of a training sample, for example collecting FFI personality characteristics of a large enough sample of people through surveys, or the HR performance ratings for employees where high-performing networking behavior should be identified. Or label a subset of e-mail messages manually by assigning an emotion rating to each message. A sample suitable for machine learning should consist of at least 100 cases, better many more, and the values should be evenly distributed.
- Feeding the features which have been calculated in step (2) into a machine learning system such as Rapidminer, Knime, or a python framework such as scikit-learn, predicting the dependent or target variable from step (3) above using the data from step (2) to build a machine learning model.
Now the four steps will be discussed in more detail.