Abstract Adapting to drifting data streams remains a key challenge in online learning, where effective model adaptation depends on timely concept drift detection. Most existing approaches respond to drift only after distributional changes occur, reacting to concept drift, limiting their ability to prevent the classifier’s performance degradation. This work introduces a novel methodology to anticipate concept drift and enable proactive adaptation before the data distribution shift negatively impacts the classifier. We propose four proactive adaptation strategies based on the Very Fast Decision Tree (VFDT) algorithm to leverage data trends to estimate proactive changes in the classifier, mitigating or even preventing the performance degradation consequence of the concept drift. We evaluate the proposed methods across four scenarios with diverse data stream configurations. Results demonstrate that proactive adaptation reduces the adverse effects of concept drift and improves classification performance. In particular, the proposed strategies consistently outperformed in settings with incremental drift, underscoring the potential of anticipatory approaches and addressing a notable gap in the current literature.

One of the biggest challenges in learning from data streams is adapting the classification model to new data. Due to the evolving nature of data streams, they are subject to a phenomenon known as concept drift that makes previously learned knowledge and model outdated. Therefore, concept drift must be efficiently detected in order to adapt the classification model. While there exists a plethora of drift detectors, with different mechanisms, selecting the most suitable for a new stream is a difficult task, since apriori knowledge may not be available and changes over time can affect the performance of the detector. This paper proposes a framework that exploits statistical and temporal meta-features from sliding windows to dynamically recommend a suitable drift detector in real-time for unseen chunks of streams according to its properties using Meta-Learning. We performed experiments on 10 real-world data streams and 18 synthetic generated data streams that were subject to concept drift and class imbalance in order to evaluate the performance of the proposed framework. Experiments exposed that the proposed approach was able to enhance the concept drift detection in a variety of scenarios demonstrating robustness to class imbalance and the advantages of dynamically selecting the drift detector.

Choosing the most suitable algorithm to perform a machine learning task for a new problem is a recurrent and complex task. In multi-target regression tasks, when problem transformation methods are applied, this choice is even harder. The reason is the need to simultaneously choose the problem transformation method and the base learning algorithm. This work investigates how to bridge the gap of method/base learner recommendation for problems with multiple outputs. In meta-learning experiments, we use a large number of multi-target regression datasets to investigate whether using meta-learning can provide good recommendations. To do this, we compared the meta-models induced by 3 different ML algorithms, including three variations for each of them, and selected 58 meta-features that we believe are relevant for extracting good dataset descriptions for the meta-learning process. In the experimental results, the meta-models outperformed the baselines (Majority and Random) by recommending the most suitable solution for multi-target regression (for the transformation method and base-learner) with high predictive performance, including real-world applications. The meta-features and the relation between the transformation method and base-learner provided important insights regarding the optimal problem transformation method. Furthermore, when comparing the application of algorithm adaptation and problem transformation methods, our meta-learning proposal was capable of statistically overcoming all competitors, which resulted in a predictive performance using the best choice per problem.

Image segmentation is a key issue in image processing. New image segmentation algorithms have been proposed in the last years. However, there is no optimal algorithm for every image processing task. The selection of the most suitable algorithm usually occurs by testing every possible algorithm or using knowledge from previous problems. These processes can have a high computational cost. Meta-learning has been successfully used in the machine learning research community for the recommendation of the most suitable machine learning algorithm for a new dataset. We believe that meta-learning can also be useful to select the most suitable image segmentation algorithm. This hypothesis is investigated in this paper. For such, we perform experiments with eight segmentation algorithms from two approaches using a segmentation benchmark of 300 images and 2100 augmented images. The experimental results showed that meta-learning can recommend the most suitable segmentation algorithm with more than 80% of accuracy for one group of algorithms and with 69% for the other group, overcoming the baselines used regarding recommendation and segmentation performance.