Approaches to Hyperparameter Tuning in the Task of Avatars Clustering in Online Social Media

The growing popularity of social media platforms such as VK is accompanied by an increase in the volume of graphical digital traces left by users (for example, profile pictures or avatars). Such data are valuable to researchers who harness machine learning models to develop solutions for various applications including marketing, banking scoring, public health, and information security. Modern approaches have evolved into flexible and versatile tools that must be adjusted to the peculiarities of specific datasets. In this study, we address the problem of hyperparameter tuning aimed at enhancing the clustering quality of VK user avatars. The underlying algorithm, common in image analysis workflows, consists of the sequential application of a neural network-based feature extraction model, a dimensionality reduction step, an outlier detection phase, and finally a clustering step. In the present work, we investigate the pipeline comprised of the methods CLIP-UMAP-LOF-HDBSCAN. For hyperparameter optimization, we evaluate several strategies, including grid search, random search, gradient-based optimization, Bayesian optimization, and an evolutionary approach. An experimental study was conducted on a dataset comprising more than 50,000 VK avatars. The results indicate that Bayesian optimization yielded the most effective clustering, achieving stable groupings with silhouette values exceeding 0.56. In addition, the optimal hyperparameter settings derived from this investigation can be used to obtain superior clustering on similar datasets.