Research On Improved Ant Colony Algorithm Based Data Classification

Data Classification is an important branch in Data Mining. With the fast developments of Information Technology and Internet, traditional classification methods cannot meet people's need. As a swarm intelligence algorithm developed in recent years, Ant Colony Algorithm obtains good results for large-scale combinatorial optimization problems. Furthermore, Ant Colony Algorithm has been applied to solve classification problems, and presents a great potential and a wide development perspective. Nevertheless, its performance is not fully developed because of the disadvantages in its own mechanism. Therefore, it is meaningful both in theory and application to improve it when applied to data classification. To begin with, the comparison study between TSP and the process of rules extraction is conducted. Then the model of TSP is applied to construct the model of data classification, which is processed by means of Ant Colony Algorithm and acquires the classification rules. Furthermore, the disadvantages in the internal mechanism of Ant Colony Algorithm are analyzed and thus improved. The main works of this paper are as follows:1. The concept of "contribution arcs" is introduced to avoid the effect of the "short-sight" of ants. Based on this concept, the "contribute function" was presented to improve the selection strategy of ants, which results in that ants can have more comprehensive knowledge about global search space;2. The "contribution function" was also applied to the smoothing of pheromone trails to dynamically update the pheromone trails according to "contribution" when algorithm is close to convergence. This improvement can keep the diversity of search space, and in turn enhanced the ability of algorithm to explore the global optimal solutions;3. Due to the "blindness selection" of ants, the concept of "information segment" was introduced to dynamically decrease the dimension of search space to make algorithm more efficient, but not decreased the quality of solutions.Finally, a prototype system to solve the TSP and data classification was developed, and accordingly the performance of improved algorithm was evaluated. The experiment result indicated that compared with AntMiner and CN2, the classification rules extracted by improved algorithm are simpler to be understood and can predict the data with unknown classifications more exactly.