Using the influence model coefficients and the random walk to predict emotional interactions in parent-child conversations

This study introduces an interactive random walk as a new method for predicting sequences of four different construct states (positive emotion, negative emotion, neutral emotion and silence) of speakers in parent-child conversations. The proposed approach used the emotional transition probability arrays and the Influence Model (IM) coefficients to support the interacting random walk predictions. The interactive random walk was applied to generate sequences of speakers´ states using higher order emotional transition probabilities. The new approach was tested on 63 different parent-child conversations conducted in naturalistic (not-acted) way. The prediction outcomes were visualized using the 2D random walk on a graph approach. The prediction quality was measured using the relative error between the actual and the predicted transition probabilities as well as, the error between the actual and the predicted end-point position on the 2D graph of emotional states. A comparison between the proposed random walk supported by the IM coefficients and the classical approach without the IM coefficients showed that proposed method generally offers improved results in terms of the prediction error and the endpoint position but at the cost of slower convergence rates.