A Multi-way Agglomerative IB Algorithm

By compressing one variable to the "Bottleneck" variable and maximally preserving the mutual information which is among the variables, we can find out some potential data structure model from a great deal of data. Along with the embedded applications of method in areas such as image analysis, machine learning, pattern recognition and so on, the shortcomings of IB method have emerged, because of the original formulation of the IB principle concentrates on compressing one variable, X, while preserving the information it maintains about some other, relevant, variable Y. This formulation is inherently a-symmetric, especially in co-occurrence data. IB method cannot exploit the collaborative relations among the various ways of data.To solve the problem, this paper proposes an algorithm called multi-way aIB algorithms which based on the multi-way IB idea and aIB algorithm, the aIB algorithm employs a greedy agglomerative method to find a hierarchical clustering tree in a bottom-up fashion. This algorithm fully exploits the collaborative connections of data ways which can analysis the data of each way after choosing the appropriate analysis mechanism. Experiment results show that the Multi-way aIB algorithm can significantly improve the accuracy of data analysis and the results can be interpreted easier. Multi-way IB can not only be applied aIB algorithm, but also equally suitable for other data analysis algorithms.