A1426
Title: Kernel metric learning for mixed-type fuzzy clustering
Authors: John Thompson - University of British Columbia (Canada) [presenting]
Daniel Krasnov - McGill University (Canada)
Abstract: Fuzzy clustering algorithms, such as the popular fuzzy C-means algorithm, extend hard clustering to allow for a degree of uncertainty in the cluster assignments through a fuzzifying parameter in the objective function. However, challenges remain in selecting an optimal fuzzy parameter, incorporating mixed continuous and categorical data types, and balancing variables based on their importance to fuzzy clustering. A kernel weighting approach is proposed in the objective function, where bandwidth selection controls variable importance and reduces the influence of selecting a fuzzy parameter. Methods are discussed for selecting bandwidths through kernel metric learning, as well as bandwidth optimization in the objective function. The kernel metric is incorporated into the fuzzy C-means objective function, and it is shown that the bandwidth and the fuzzifying parameter are correlated. It is found that the overall effect of fuzzifying parameter choice on cluster center values and fuzzy adjusted Rand index is mitigated by selecting bandwidths that optimize the objective function. This method is applied to simulated and real mixed-type benchmark datasets to demonstrate improvements in clustering performance against current fuzzy clustering algorithms.
