Individual Neighbor Aware Sentiment Prediction Approach Based on Irregular Time Series

Abstract

The individual sentiment state has become a hot topic in many application fields because it can help analyze and predict their future behavior. In real life, there are usually many situations that an individual miss his/her emotion, such as conflicting emotions with others, feeling insecure or in an uncomfortable situation, etc. The missing values in data sets lead to significant challenges in individual sentiment prediction. At the same time, another difficulty in predicting individual sentiment is that individual sentiment will change with the dynamic changes of time, environment and one's own mood. To accurately capture the characteristics of individual sentiment changes, this information needs to be fully and reasonably combined. In addition, existing methods only model the individual sentiment vector extracted from the text, while ignoring the most important text semantic information modeling. Therefore, to efficiently predict individual future sentiment, we propose an individual neighbor aware sentiment  prediction approach based on  under irregular time series(denote to INSP). Specifically, the INSP consists of three parts: individual and neighbor context-aware network, individual time attention embedded network, and sentiment personality role labels. The  individual and neighbor context-aware network captures individual sentiment changes and the influence of neighbor users. At the same time, an intelligent grouping mechanism is designed to extract the sentiment features contained in the text and remove noise information of the text. The individual time attention embedded network captures the changes in sentiment features of individual  under irregular time series. Individual sentiment personality role labels count the sentiment probability distribution of an individual over a period of time in order to comprehensively understand the individual's sentiment tendencies. INSP's F1 score improve approximately from 1.8% to 15% on X dataset, from approximately 2.4 to 11.2% on Flicker dataset, and from 3.3% to 13.5% on Truth dataset.

Keywords: Sentiment analysis, Sentiment prediction, Neighbor influence, Time series, Personality labels