fcm.zip_FCM + wsn_WSN_Wsn fcm_fcm wsn_it

Distance Contents 1 Introduction 2 Usage 3 Adding distance functions 1 Introduction The main purpose of distance-module is to provide a flexible way to add new distance functions without a need to modify the algorithms that use them. This implementation supports using centroid vectors as models of clusters. Therefore e.g. Gaussian mixture models aren't supported by the interface. All distance functions are hidden behind DistanceInfo object, which stores information about which functions to call. Distance can be laculated between two data points, two models, from data point to model and from model to data point. I.e. the distance function need not be metric and distance from model to point need not be the same as from point to model. 2 Usage If you are using some algorithm to perform clustering and you only need to pass the distance function to the algorithm, do as follows: First, include "DistCrit/distance.h". This file is the only one you need to include regardless of which distance function you use. Second, create DistanceInfo object with diNew-function. It takes information about the type of the distance (used for built-in types) and alternatively function pointers if you want to use your own distance function. For description of these, see next chapter. If you use any criteria as defined in criteria-module, you must initialize the criterion with DistanceInfo-object. Note that criteria may have some specific distance function that they have been designed to work with, so usually it is best to use ciCriterionDefaultDistance-function to set the distance function, unless you know what you are doing. Likewise certain distance functions may have been ddesigned for certain criterion, so using them elsewhere is not recommended unless you knwo what you are doing. Third, pass the distance object to the desired algorithm along with other necessary information. And in the end, destroy the DistanceInfo-object with diDelete. If you need to use the distance functions in your own code, there is macro diDistance, which takes data set, model, partitioning, indices of start and end and information about the direction. That's all there is to it. Note: only use the DistanceInfo object for the data set it was initialized with. 3 Adding distance functions First, you need four functions, each of which takes as parameter: TRAININGSET CODEBOOK PARTITIONING integer specifying where the distance is calculated from integer specifying where the distance is calculated to integer specifying the direction The direction is possibly redundant. Using it, you can write one function, which determines with the last parameter how to calculate the distance. If you provide all four specialized functions, the overhead of determining the direction is eliminated, since each of the functions is called only for the specific case. You can also provide one function for the most common case and use single function for the rest of the cases etc. The main point is that a function for all four cases is required, even if it is the same function. Direction also gives indication what the indices mean. If the distance is calculated from model then the first index refers to cluster, otherwise to data point. Likewise for the second index. Once you have the necessary functions, you can either add them to the distance-module or use them as user-defined distance function. The first step involves adding new enumerant to distance.h and corresponding pieces of code to diNew-function so that the DistanceInfo object contains pointers to correct function of the type. Using your distance function as user-defined function only requires that you pass the function pointers to diNew and pass DT_User as the type. When creating DistanceInfo-object, you also pass the TRAININGSET that is used to the function. The intention is that in some cases, a special version of a distance function can be used, e.g. for data of known dimensionality that is used very often. This allows some optimizations when the dimensionality can be treated as constant. Therefore you should use the DistanceInfo-object only for the TRAININGSET it was initialized with. When implementing the distance-function, keep in mind that for the average clustering algorithm, it is the function that is called most often, therefore any unnecessary work will quickly slow down the program. This is one of the reasons the four functions are available, to eliminate the switch statement which would be relatively costly for low-dimensional data.