DALR: A Dynamically Adjustable Linear Regression Model for Handling Stream Data
Rishita Mondal, Sourab Mandal, Paramartha Dutta, Arindrajit Pal · 2024
In response to the escalating demand for machine learning techniques capable of handling real-time data streams, particularly in applications like stock markets, this research dives deep into the domain of stream regression. The objective is to adeptly navigate challenges posed by evolving data over time, differentiating it from conventional static regression. Existing models, such as Fast and Incremental Model Trees with Drift Detection (FIMT-DD), k-Nearest Neighbours (KNN), and Adaptive Linear Regression (ALR) and some other studies, though competent, exhibit certain limitations like learning based on certain portions of data, using a significantly large memory and not capable of handling each data in every timestamp it arrives. We propose the Dynamically Adjustable Linear Regression Model (DALR) to address these shortcomings. DALR introduces an adaptive regression model that dynamically evolves based on historical knowledge and newly acquired data, providing an innovative solution to the intricacies of dynamic data streams. This endeavor has also highlighted that the DALR algorithm consistently demonstrates strong performance across various datasets according to the RMSE metric, surpassing other state-of -the-art algorithms in most instances.