The big data mining forecasting model based on combination of improved manifold learning and deep learning

In this paper, we use the combination of Local Linear Embedding (LLE) with Continuous Deep Belief Networks (CDBN) as the input of RBF, and construct a mixed-feature RBF model. However, LLE depends too much on the local domain which is not easy to be determined, so we propose a new method, Kernel Entropy Linear Embedding (KELE) which uses Kernel Entropy Component Analysis (KECA) to transfer the non-linear problem into linear problem. CDBN has the difficulty in confirming network structure and lacks supervision, so we improve the situations by using the kernel entropy information obtained from KECA, which is called KECDBN. In the empirical part, we use the foreign exchange rate time series to examine the effects of the improved methods, and results show that both the KELE and the KECDBN show better effects in reducing dimensionality and extracting features, respectively, an also improve the prediction accuracy of the mixed-feature RBF.