EcoSentiNet: A Multi-Objective Deep Reinforcement Learning Model for Energy-Optimized Sentiment Classification

The growing computational requirements of transformer-based sentiment analysis models have intensified the energy efficiency and sustainability concern especially in the mobile, Internet of Things, and edge computing systems. In order to overcome this issue, the paper presents EcoSentiNet, which is a multi-objective Deep Reinforcement Learning (DRL) model that aims at energy-efficient sentiment classification. The EcoSentiNet is balanced with accuracy, latency and power consumption dynamically by varying the depth of neural network, computation granularity, and the timing of an inference. The framework uses a DRL controller that is a policy gradient based and learns the best inference strategies in different hardware and workload conditions. In contrast to the classical static models, EcoSentiNet is able to continuously assess the energy-performance trade-offs via the reward scheme that combines both accuracy and energy cost, guaranteeing the inference quality over long periods of time under limited energy budgets. The experiments on the benchmark datasets, including IMDb, Twitter Sentiment140, and Amazon Reviews, demonstrate that EcoSentiNet can reduce energy usage and minimize latency significantly (40 and 25 percent, respectively) without further deteriorating the accuracy in relation to the traditional deep sentiment models. The suggested framework proves sustainability of the NLP in terms of intelligent energy-performance co-optimization and moves the field forward to the next stage of Green AI real-time sentiment comprehension.