Research On Manifold Learning Based Classifiers And Their Applications

Using data mining technology for earthquake prediction is an interesting field of academic research, which has important academic value and practical significance. This thesis aims to explore a new way to predict time interval between aftershocks and magnitude based on data mining technologies, and to explore that manifold learning dimensionality reduction techniques are applied to anomaly detection of aftershock. With careful study and comparison of several data mining algorithms, this thesis seeks to explore appropriate methods for aftershock time prediction, aftershock magnitude prediction and dimensionality reduction of seismic characters attributes, and to develop the software prototype system. The main contributions are included as follows:A prediction method for time interval between aftershocks based on Adaptive Local Linear ALL is proposed. ALL is an adaptive local linear method based on singular value decomposition, and it could determine current embedding dimension adaptively, which thus overcome the impact of pathological data matrix. The experimental datasets are time intervals between aftershocks with magnitude greater than or equal to 4.0 from Wenchuan earthquake. The evaluation criterions are MRMSE (Mean of Root Mean Square Errors), MMAE (Mean of Mean Absolute Errors) and AE (Absolute Error). Comparing with standard local linear and least square fitting, experimental results show that ALL is an effective prediction method for time interval between aftershocks.For the prediction problem that decision attribute is real value, a modeling method named PR-KNN (Polynomial Regression and K Nearest Neighbor) is proposed, which is based on combination of K Nearest Neighbor and Polynomial Regression. Experimental data are the sequence data of aftershocks with magnitude greater than or equal to 4.0 from Wenchuan earthquake. Time intervals between aftershocks are considered as condition attribute, and aftershock magnitude as decision attribute. The evaluation criterions are RE (Relative Error) and AE (Absolute Error). Comparing with traditional KNN and Distance-Weighted KNN regression algorithm, experimental results show that PR-KNN is a potential method of aftershock magnitude prediction.For the problem that lacking mapping function for test samples from high-dimensional space to low-dimensional space in the supervised manifold learning algorithms, a supervised manifold learning algorithm named PR-SLLE is proposed, which is based on combination of supervised locally linear embedding, K Nearest Neighbor and Polynomial Regression. And this method is applied to anomaly detection. Experimental data are seismic attribute data obtained from Wenchuan earthquake. The evaluation criterions are AR (Accuracy Rate), FR (False alarm Rate) and OR (Omission Rate). Comparing with standard SLLE algorithm, experimental results show that the predicted effect by PR-SLLE and Bayesian classifier is superior to that of SLLE, and also illustrates that PR-SLLE is a feasible and effective dimensionality reduction method.On the basis of above research, an aftershock prediction prototype sub-system was developed which was used as one important part of the software prototype system of data mining based earthquake tendency prediction and assessment. The sub-system includes three modules, the module of time interval between aftershocks prediction, the module of aftershock magnitude prediction and the module of aftershock anomaly detection. The test results show that the prototype system runs well and the aim is to lay the foundation for further study.