Abstract:
The article is devoted to the development of a method for constructing interpretable survival analysis models. An extended Cox model is used as a basis, where the dependence between traits is specified by a polynome. To solve the problem of optimising the polynomial structure, a modification of the ant-pollinator algorithm is suggested. A key feature is the matrix representation of the goal function, which combines accuracy criteria (c-index), the number of features, and model complexity. Unlike classical approaches, the pheromone in the algorithm is deposited at the vertices of the trait graph. The method was tested on data about critical malfunctions in 5,000 cars. During the experiment, the algorithm demonstrated its effectiveness, accurately restoring the specified dependence of the risk function on features with an average number of iterations of 6.09. The results confirm that the proposed approach allows to simultaneously build accurate predictive models and select significant features, ensuring high interpretability of the results. The prospects for further work are related to the development of a matrix representation of pheromones and testing on real data.
Keywords:survival analysis, ant pollinator method, optimisation methods, matrix goal function.
Fulltext:
PDF file (194 kB)