Learning regularized hoeffding trees from data streams

Learning from data streams is a hot topic in machine learning that targets the learning and update of predictive models as data becomes available for both training and query. Due to their simplicity and convincing results in a multitude of applications, Hoeffding Trees are, by far, the most widely used family of methods for learning decision trees from streaming data. Despite the aforementioned positive characteristics, Hoeffding Trees tend to continuously grow in terms of nodes as new data becomes available, i.e., they eventually split on all features available, and multiple times on the same feature; thus leading to unnecessary complexity. With this behavior, Hoeffding Trees lose the ability to be human-understandable and computationally efficient. To tackle these issues, we propose a regularization scheme for Hoeffding Trees that (i) uses a penalty factor to control the gain obtained by creating a new split node using a feature that has not been used thus far; and (ii) uses information from previous splits in the current branch to determine whether the gain observed indeed justifies a new split. The proposed scheme is combined with both standard and adaptive variants of Hoeffding Trees. Experiments using real-world, stationary and drifting synthetic data show that the proposed method prevents both original and adaptive Hoeffding Trees from unnecessarily growing while maintaining impressive accuracy rates. As a byproduct of the regularization process, significant improvements in processing time, model complexity, and memory consumption have also been observed, thus showing the effectiveness of the proposed regularization scheme.