TinyML Based Optimization of Nonlinear Models for Housing Price Prediction on Resource Constrained Embedded Systems

The expanding adoption of machine learning in low‑power devices and embedded systems has created a growing need for models that maintain acceptable predictive accuracy while exhibiting low computational complexity and memory consumption. In the context of TinyML, developing such lightweight models has become an essential requirement. In this study, a lightweight framework for housing price prediction is proposed using the Ames Housing dataset, with the objective of enabling efficient deployment of learning models on resource‑constrained embedded systems. Following data preprocessing steps including data cleaning, feature selection, and feature normalization, several linear and nonlinear learning models—Ridge Regression, Random Forest, Decision Tree and a Multilayer Perceptron neural network (MLP) —are trained and evaluated. To adapt nonlinear models to the hardware constraints of embedded environments, several optimization techniques are applied, including reducing the number of input features, limiting the depth and number of trees in ensemble models, structural pruning, and model quantization. The performance of the models is compared in terms of Mean Absolute Error, Root Mean Squared Error, coefficient of determination, and also measured by model size and inference time to examine their suitability for memory‑limited systems. The experimental results indicate that the optimized models achieve competitive predictive accuracy despite their computational complexity, making them suitable candidates for deployment in embedded systems and TinyML applications.