In the last decades Machine Learning models have become increasingly used in different areas of interest, leading to the increase of computational and energy resources. In this context, sustainability has become a global imperative, and Green AI is currently used to encapsulate the concept of development of high-performing and eco-friendly models.

We propose in this paper a function for the selection of the best prediction model from a model library, based on a criterion which combines accuracy, latency, and carbon footprint.

Recommendation systems and sentiment analysis are two of the most common tasks that require high resource consumption. The purpose of this study is to create a system that is based on these two applications with a user-friendly desktop interface. Unlike a traditional system that involves running a single model for each functionality, this one will have to choose the optimal model based on the real-time values of carbon emissions intensities from Romania region obtained from an Electricity Maps API and other criteria based on user input, also taking into consideration the type of device on which the model is running. This approach represents a step in the emerging direction of Green AI paradigm, which also represents the motivation of this study.

The proposed system is composed of two main components: (1) image-based recommendation system and sentiment analysis. The recommendation system uses the Fashion Product Images (Small) dataset from Kaggle platform. The candidate pre-trained CNN (Convolutional Neural Networks) models are: ResNet50, InceptionV3 and MobileNetV2. The score function for the model selection takes into consideration the following aspects: the accuracy and latency of K-Nearest Neighbors algorithm for image recommendation of the entire dataset, the documented latency of each CNN model, the resolution of uploaded image by user, the emission factor score and the device where the model runs. The image resolution criterion offers a “bonus” if the image is close to the model target or a “penalty” otherwise. The emission factor score value adjusts the model score based on the emission factor value from the API. Higher values disadvantage complex models, while simpler models gain preference. This logic applies to the sentiment analysis component, too, where the utilized models are BERT (Bidirectional Encoder Representations from Transformers) and VADER (Valence Aware Dictionary and sEntiment Reasoner), here a particular criterion being the text length.

Few results of this study are as following: In the recommendation system, it was observed that for large dimension images, if the emission factor from the API exceeds the value of 360 gCO2/kWh, the selected model is MobileNetV2 in support of sustainability and for the sentiment analysis component, if the length of text exceeds 50 characters and the emissions value are over 300 gCO2/kWh, VADER is the selected model.

According to the official documentation of the utilized models, the results align with expectations, reflecting each model’s complexity and dimension. These key characteristics along with CO2 emission data, guide the selection of the appropriate model based on the platform used to run the app and user input.