Addressing Drawbacks of MART and Enhancing Predictive Performance*

Boosting, a successful machine learning algorithm, is often used for regression and classification tasks. Boosted ensemble algorithms, particularly when combined with tree base learners, can generate highly accurate and robust predictive models. This study explores the weaknesses of the well-known approach, Multiple Additive Regression Trees (MART), and proposes a new approach called Laplace Additive Regression Trees (LART) that addresses this issue by adding random Laplace noises. The algorithm is then improved upon by combining it with Dropout and multiple Additive Regression Trees, resulting in DLART [1]. In an experiment using a previously studied dataset, we compared the performance of LART, DLART, MART, and DART. The results were promising, as LART and DLART produced significant gains in accuracy for this particular task. For future studies, it would be beneficial to assess the effectiveness of LART on complex data sets and various other machine learning tasks, such as those related to data mining [2]