A Robust Cold-Start Approach Using Phylogram

In Active Learning (AL) applications, it is crucial to develop a model that efficiently represents data using a limited number of labeled samples. In this context, an oracle is often employed to label samples that are deemed representative and informative. However, one of the primary limitations of AL lies in the need for an initial set of labeled samples, which may not be representative of the entire data space. To address this issue, cold-start strategies are implemented with the goal of selecting an appropriate initial set of samples for labeling, ensuring that the initial model is robust and effective. We propose a novel approach – PhyIL – that harness the advantages of phylogram analysis and community detection techniques to smartly perform the cold-start selection in AL. The method facilitates the choice of samples that best represent the diversity and complexity of the data, which is fundamental for the subsequent success of the active learning process. We evaluated the proposed approach using nine real-world databases and nine traditional ML models. The findings indicate that PhyIL offers an effective and viable solution for the initial sample selection, when compared to the random selection.