Abstract:
The technique based on the use of a hybrid optimization method to solve the feature selection problem for the Cox regression model is proposed. The hybrid optimization method includes two metaheuristic methods: the ant colony optimization and the genetic algorithm. The ant colony optimization used as the basic algorithm that solves the main optimization problem. The genetic algorithm solves the problem of finding the optimal set of parameters for the ant algorithm improving its performance. The method is modified and adapted to solve the problem under consideration. The key feature of adaptation is the deposition of pheromones on the vertices rather than on the edges of the graph, as well as the method for calculating the assessment of heuristic information about each vertex. A fitness target function was constructed that determines the quality of solutions to the feature selection problem and is based on an assessment of the performance of the Cox model with a selected set of parameters. The concordance index (c-index) was used to evaluate the Cox model. The efficiency of the methodology is shown using the implemented program using the example of a database of recidivism. For the database used, the most significant sets of features were obtained that have the greatest impact on the quality of training of the survival analysis model.
Fulltext:
PDF file (268 kB)