Étude de variabilité par bootstrap résiduel pour une méthode de subspace clustering

Combining dimensionality reduction and clustering techniques is effective for clustering high-dimensional data by simultaneously identifying low-dimensional subspaces and their corresponding partitions. This work aims to assess the sensitivity of the partition and the subspace to sampling fluctuations. To achieve this, we first propose a way to simulate these fluctuations using different bootstrap approaches. This initial step, for instance, provides a means of visualizing the sensitivity of the results through the construction of confidence ellipses around the centroids. We then propose a method to quantify this sensitivity. The proposed procedure is evaluated through simulations on different data structures. The results show that, with the proposed bootstrap approach, it is possible to correctly assess the sensitivity of the method, provided that the chosen bootstrap strategy matches the structure of the data.