Ting Yu

VQSVM: A case study for incorporating prior domain knowledge into inductive machine learning

The paper reviews the recent developments of incorporating prior domain knowledge into inductive machine learning, and proposes a guideline that incorporates prior domain knowledge in three key issues of inductive machine learning algorithms: consistency, generalization and convergence. With respect to each issue, this paper gives some approaches to improve the performance of the inductive machine learning algorithms and discusses the risks of incorporating domain knowledge. As a case study, a hierarchical modelling method, VQSVM, is proposed and tested over some imbalanced data sets with various imbalance ratios and various numbers of subclasses.

A Hierarchical VQSVM for Imbalanced Data Sets

First, a hierarchical modelling method, VQSVM, is introduced, and some remarks are discussed. Secondly the proposed VQSVM is applied to a nonstandard learning environment, imbalanced data sets. In cases of extremely imbalanced dataset with high dimensions, standard machine learning techniques tend to be overwhelmed by the large classes. The hierarchical VQSVM contains a set of local models i.e. codevectors produced by the vector quantization and a global model, i.e. support vector machine, to rebalance datasets without significant information loss. Some issues, e.g. distortion and support vectors, have been discussed to address the trade-off between the information loss and undersampling rate. Experiments compare VQSVM with random resampling techniques on some imbalanced datasets with varied imbalance ratios, and results show that the performance of VQSVM is superior or equivalent to random resampling techniques, especially in case of extremely imbalanced large datasets.

Classify Unexpected News Impacts to Stock Price by Incorporating Time Series Analysis into Support Vector Machine

The paper discusses an approach of using traditional time series analysis, as domain knowledge, to help the data-preparation of support vector machine for classifying documents. Classifying unexpected news impacts to the stock prices is selected as a case study. As a result, we present a novel approach for providing approximate answers to classifying news events into simple three categories. The process of constructing training datasets is emphasized, and some time series analysis techniques are utilized to pre-process the dataset. A rule-base associated with the net-of-market return and piecewise linear fitting constructs the training data set. A classifier mainly built by support vector machine uses the training data set to extract the interrelationship between unexpected news events and the stock price movements.

Financial prediction using modified probabilistic learning network with embedded local linear models

In this paper, a model is proposed which combines multiple local linear models with a novel modified probabilistic neural network (MPNN). The proposed model is shown to provide improved regularization with reduced computation utilizing semiparametric model approach and efficient vector quantization of data space. In this paper, the proposed model is shown to generalize better with reduced variance and model complexity in short-term financial prediction application.

Incorporating prior domain knowledge into a kernel based feature selection algorithm

This paper proposes a new method of incorporating prior domain knowledge into a kernel based feature selection algorithm. The proposed feature selection algorithm combines the Fast Correlation-Based Filter (FCBF) and the kernel methods in order to uncover an optimal subset of features for the support vector regression. In the proposed algorithm, the Kernel Canonical Correlation Analysis (KCCA) is employed as a measurement of mutual information between feature candidates. Domain knowledge in forms of constraints is used to guide the tuning of the KCCA. In the second experiments, the audit quality research carried by Yang Li and Donald Stokes [1] provides the domain knowledge, and the result extends the original subset of features.