With the growing number of machine learning (ML) algorithms, the selection of the top-performing algorithms for a given dataset, task, and evaluation measure is known to be a challenging task. The human expertise required for this task has fueled the demand for automatic solutions. Meta-learning is a popular approach for automatic algorithm selection based on dataset characterization. Existing meta-learning methods often represent the datasets using predefined features and thus cannot be generalized for various ML tasks, or alternatively, learn their representations in a supervised fashion, and thus cannot address unsupervised tasks. In this study, we first propose a novel learning-based task-agnostic method for dataset representation. Second, we present TRIO, a meta-learning approach based on the proposed dataset representation, which is capable of accurately recommending top-performing algorithms for unseen datasets. TRIO first learns graphical representations from the datasets and then utilizes a graph convolutional neural network technique to extract their latent representations. An extensive evaluation on 337 datasets and 195 ML algorithms demonstrates the effectiveness of our approach over state-of-the-art methods for algorithm selection for both supervised (classification and regression) and unsupervised (clustering) tasks.

The widespread use of machine learning algorithms and the high level of expertise required to utilize them have fuelled the demand for solutions that can be used by non-experts. One of the main challenges non-experts face in applying machine learning to new problems is algorithm selection - the identification of the algorithm(s) that will deliver top performance for a given dataset, task, and evaluation measure. We present AutoGRD, a novel meta-learning approach for algorithm recommendation. AutoGRD first represents datasets as graphs and then extracts their latent representation that is used to train a ranking meta-model capable of accurately recommending top-performing algorithms for previously unseen datasets. We evaluate our approach on 250 datasets and demonstrate its effectiveness both for classification and regression tasks. AutoGRD outperforms state-of-the-art meta-learning and Bayesian methods.