Sentiment Analysis via Trustworthy Label Enhancement for Consumer Electronics Applications

Abstract

Consumer electronics are becoming increasingly popular in our daily life. Enabled with Artificial Intelligence (AI) of Things (AIoT), consumer electronics can autonomously analyze user data and learn user preferences. AIoT has empowered various personal consumer applications, such as healthcare and recommendation, in which user sentiment analysis is necessary. This paper studies sentiment analysis by analyzing user data generated from consumer electronics, especially image data. Considering that the images generated by consumer electronics generally have blended emotions, we apply the Label Enhancement (LE) technologies to enhance the emotion labels into fine-grained emotion distributions. To match the need of real-world AIoT scenarios, we put forward in this paper the first trustworthy LE algorithm, called LE-Weighted k-Nearest Neighbors (LE-WkNN). Theoretical analysis shows that the enhanced emotion distributions by LE-WkNN are guaranteed to approach the ground-truth ones, which has strong theory guidance. Second, we train a convolution neural network to learn the enhanced emotion distributions. Finally, we conduct experiments on three large-scale emotion datasets. The experimental results validate that LE-WkNN accurately enhances the emotion distributions and our model achieves the best performance for sentiment analysis. Overall, LE-WkNN is trustworthy and has great potential for consumer electronics applications.