Asteroid families are groups of asteroids that are the product of collisions or of the rotational fission of a parent object. These groups are mainly identified in proper elements or frequencies domains. Because of robotic telescope surveys, the number of known asteroids has increased from ${\simeq}10\, 000$ in the early 1990s to more than $750\, 000$ nowadays. Traditional approaches for identifying new members of asteroid families, like the hierarchical clustering method (HCM), may struggle to keep up with the growing rate of new discoveries. Here we used machine learning classification algorithms to identify new family members based on the orbital distribution in proper (a, e, sin (i)) of previously known family constituents. We compared the outcome of nine classification algorithms from stand-alone and ensemble approaches. The extremely randomized trees (ExtraTree) method had the highest precision, enabling to retrieve up to 97 per cent of family members identified with standard HCM.

Asteroid families are groups of asteroids that share a common origin. They can be the outcome of a collision or be the result of the rotational failure of a parent body or its satellites. Collisional asteroid families have been identified for several decades using hierarchical clustering methods (HCMs) in proper elements domains. In this method, the distance of an asteroid from a reference body is computed, and, if it is less than a critical value, the asteroid is added to the family list. The process is then repeated with the new object as a reference, until no new family members are found. Recently, new machine-learning clustering algorithms have been introduced for the purpose of cluster classification. Here, we apply supervised-learning hierarchical clustering algorithms for the purpose of asteroid families identification. The accuracy, precision, and recall values of results obtained with the new method, when compared with classical HCM, show that this approach is able to found family members with an accuracy above 89.5 per cent, and that all asteroid previously identified as family members by traditional methods are consistently retrieved. Values of the areas under the curve coefficients below Receiver Operating Characteristic curves are also optimal, with values consistently above 85 per cent. Overall, we identify 6 new families and 13 new clumps in regions where the method can be applied that appear to be consistent and homogeneous in terms of physical and taxonomic properties. Machine-learning clustering algorithms can, therefore, be very efficient and fast tools for the problem of asteroid family identification.

No abstract