Investigação de métodos para estimação de pesos de relevância para agrupamento de dados relacionais com múltiplas visões

Abstract

Multiview data clustering is characterized as the challenge of grouping similar data and separating dissimilar data by using different representations, such as attribute vectors, feature matrices, graphs, or images, whichtogetherprovidecomplementaryinformation.Thisapproachisappliedin real-world scenarios such as social network analysis, medical image classification, recommender systems, and multimedia data integration. Data clustering is one of the most relevant tools for data analysis, and its importance is related to the faithful representation of real-world data, with the combined use of multiple views providing greater accuracy when compared to the use of a single view. The experiments show that the performance of multiview clustering algorithms strongly depends on the distance metrics and weight estimation strategies. While MRDCA-RWL is more sensitive to the adopted distance metric, methods such as MVKKM and MVSC exhibit a strong joint dependence on both distance metrics and weights, reinforcing the importance of selecting appropriate parameters. Given the above, this work addresses the need for a proper interpretation of the use of multiview data in the context of contemporary applications, aiming to investigate different methods for estimating relevance weights. These weights are then applied to the different views in the context of relational data clustering, discussing both the quality of the results and the algorithmic efficiency in comparison with the complexity of competing solutions. Based on this, the experimental results demonstrate the influence of the methodological choices on the behavior of the algorithms, with MRDCA-RWL exhibiting more stable behavior, whereas MVKKM,MVSC,andTW-KMdemonstrate sensitivity to configuration changes.