HUMAN CAPITAL AND WELL-BEING METER BASED ON FEDERATED LEARNING POWERED SENTIMENT ANALYSIS

Abstract

The true well-being of a person does not only mean money, but it is the overall quality of life, the social, emotional, and psychological state of a person. All these emotions are usually reflected in common speech. Though a mere textual analysis will inform us of the emotional tone by sentiment analysis, a whole picture of well- being means an appreciation of the context in the emotion. This is offered in our well-being meter, which incorporates a capital prediction component that links the feeling to the area of life capital such as financial or social, that it was caused by. Such a general approach is necessary to have a comprehensive perspective. Nevertheless, the main risk of privacy is the old-fashioned and centralized means of analyzing this form of sensitive data. This paper proposes FedBiLSTM-Net, a secure approach based on Federated Learning (FL) and a Bidirectional Long Short-Term Memory (Bi-LSTM) network. This strategy stores the raw and sensitive text in local machines, and the clients share model updates with a central server. This decentralized data ensures privacy and integrity of data. We constructed a dataset of 336,029 text records based on both the public and clinical datasets such as CLPsych Shared Task and DAIC-WOZ datasets, to include sentiment and other types of capital. We adopted Bi-LSTM into our FL model after ensuring that it outperformed in the first test. We trained the model on 20 different independent data groups (non-IID clients) in 10 rounds using the Flower framework and FedAvg. It considers sentiment and capital prediction simultaneously, which makes it less complicated. The last model had a centralized accuracy of 98.59%. These findings indicate that a privacy-aware model can be trained using Federated Learning to effectively learn intricate well-being patterns using a diversified, distributed dataset. Our model FedBiLSTM-Net provides a viable privacy-first well-being existence tracker.