The Application Research Of Immune Algorithm In Data Mining

The main purpose of Artificial Immune System (AIS) is to extract special informationprocessing mechanisms contained in biological immune system, and then to study and designthe corresponding models and algorithms, which can be used to solve many kinds of complexproblems. Currently, AIS has been a new branch in the field of computing intelligence, whichhas already shown a powerful capability in the information processing and problem solutions,such as data mining, machine learning, automatic control and fault diagnosis, etc. Mostresearches of data mining based on AIS put emphases on data clustering, data concentrating andclassification.De Castro and Von Zuben proposed the clonal selection immune algorithm according tothe basis of clonal selection principle, which inherited many attributes of the biology immunesystem such as self-organization, self-learning, self-identification and memorizing. It can solvecomplex machine learning problems like pattern recognition and multi-modal optimization.According to the basis of clonal selection immune algorithm and hierarchical clustering, animproved dynamic clustering algorithm is presented, which needs no pre-knowledge. Firstly,the same size of antibodies as the antigens are initialized randomly. Secondly, the process ofantigen recognizing, antibody restraining and merging is performed based on antibody affinityto complete a round of clustering. Thirdly, in order to do some motivated mutating, themutating locations of antibodies are determined by using aiNET immune network model, andthe mutating rate is dynamically adjusted inversely proportional to the generation number of theimmune evolution. After that, the similar antibodies are merged. Then it repeats the aboveprocesses until the ending condition is satisfied. In hierarchical clustering, once the step ofmerging or splitting is carded out, the clusters could not be revised; while in our approach, thedynamic mutating can adjust the clusters after merging or splitting and yields better clusteringresults.In my dissertation, I had used a group of artificial datasets and some data from the UCIdatabase as the testing data. The experimental results demonstrated that my proposed algorithmcan classify the multi-dimensional data correctly, and obtained better clustering results and higher performance than the traditional ones.