Analysis of international life expectancies with manifold learning and neural networks

Abstract Manifold learning can be utilised to obtain a low-dimensional representation of data and recover the underlying structure of the data. In this paper, we apply three manifold learning techniques, including Laplacian Eigenmaps, t-SNE, and UMAP, to analyse life expectancies at birth for a larger number of populations. We categorise the populations into appropriate clusters based on manifold learning embedding and then employ neural networks and vector autoregressive models to perform multi-population modelling for each cluster. This approach offers a more informative description of the changes in international life expectancies, allowing for the potential co-movements between related populations. There are a number of major findings in this study. First, we observe that the more developed nations exhibit homogeneous temporal patterns within their respective clusters. Comparatively, we notice that the developing nations demonstrate a greater extent of heterogeneity amongst them. Moreover, the proposed approach enhances the overall accuracy of forecasting life expectancies of multiple populations.