Triclusters of close values for the analysis of 3D data

The paper deals with the problem of triclustering in multivalued triadic contexts in terms of one multidimensional extension of formal concept analysis; triclustering can be viewed as a search for dense subtensors in three-dimensional tensors over the field of real numbers. Two methods are proposed for solving this problem, namely, NOAC — a version of the OAC triclustering method for numerical data based on delta operators — and a triadic version of the $ k $-means method with an improved metric based on Manhattan distance and proximity predicates in each of the three dimensions. Numerical experiments are carried out both on real and synthetic data and confirm the superiority of the NOAC method in terms of the performance criteria for the resulting triclusters.