Abstract:
Prognostic models play a key role in decision-making in various scientific and applied areas, allowing to predict the outcome of events based on the input parameters of the studied object. One of such areas is survival analysis — a set of statistical methods designed to estimate the probability of occurrence of terminal events. The paper addresses the problem of building predictive models based on a hybrid metaheuristic ant-pollination algorithm. The efficiency of parallelization of this algorithm is examined, which allows to reduce the computational time in training the models. The work demonstrates that the proposed survival analysis model building method can be effectively parallelized. Parallelization speeds up the model training process, which is especially important when intermediate models are repeatedly trained. Experiments show that the dependence of parallelization efficiency on the number of computational nodes is non-linear: increasing the number of nodes increases the speed of the algorithm, but with a certain limit. The results show that parallelization on a few computational nodes is the most effective in terms of reducing computation time.
Keywords:parallelization, optimization, survival analysis, Cox model, ant algorithm.
Fulltext:
PDF file (219 kB)