The Research And Application Of Multiple Kernel Prediction Model Based On Statistical Learning Theory

With the development of economy, science, and technology, the ask of statistics be-comes more complex, and the size of the date set involved is on the increase. To cope with these challenges, we develop several fast kernel machines in the dissertation, which are successfully used in time series forecasting. Compared with the existing methods, the resulting methods can obtain better performance. For each machine, we make a clear theoretical analysis and carry out a series of experiments to evaluate its performance.A main problem for traditional kernel learning method is to take what kind of optimization algorithm after a learning model is established. Based on this, in order to solve LSSVM, a new technique for the selection of working set in sequential minimal optimization (SMO)-type decomposition methods is proposed. By the new method, we can select a single direction to achieve the convergence of the optimality condition. A simple asymptotic convergence proof for the new algorithm is given. Experimental comparisons demonstrate that the classification accuracy of the new method is not large different from the existing methods, but the training speed is faster than existing ones.In order to overcome the difficulty:the traditional kernel learning method needs to choose the specific kernel function, the researchers put forward multiple kernel learning (MKL) method for multi-source data or heterogeneous data. The kernel functions for MKL are the combination of kernel ones. The traditional MKL is based on (?)1norm, but the sparse solution will reduce the prediction accuracy of the model.(?)1-norm multiple support vector regression is rarely observed to outperform trivial baselines in practical applications. So to allow for robust kernel mixtures that generalize well, we adopt-norm multiple kernel support vector regression (1≤p<∞) as a stock price prediction model. The optimization problem is decomposed into smaller sub-problems and the interleaved optimization strategy is employed to solve the regression model. Experimental results show that our proposed model performs better than (?)1-norm multiple sup-port vector regression model.For the complex, time-varying data, online MKL has got the favor of the researchers, and now it has become a hot research topic in the field of machine learning. Finally, an online multi-kernel prediction framework is proposed, and the corresponding algorithm and theory analysis is given. The optimized algorithm is based on the fusion of two online learning ones. Considering the calculation cost increases gradually in the process of online learning process, Weighted Ran-dom Sampling strategy is adopted. It can reduce the computational cost and keep the forecast accuracy. The test results in standard time series data set show that the online multi-kernel sup-port vector regression prediction model obtains a good forecasting effect with large computation costs. After the random strategy is adopted, the learning accuracy has hardly fallen at all, but can significantly reduce learning time.To solve the problem of data analysis with learning theory will bring new vitality to s-tatistics, this dissertation has made a beneficial attempt in this aspect. The application of these achievements is not only confined to the forecast, the ideas and relevant theory, technology can also be extended to other learning areas. Learning from data prompts the development of s-tatistics, and in turn, the development of statistics provides more theoretic support for learning. The theory of data analysis and processing method for statistical practice has a certain guiding significance.