Research On Intelligent Decision-making Methods Based On Similarity Learning

Decision-making theory plays an important role in management science and Economics. Decision-making means the selection process for achieving the optimal target from a number of options. According to the way of traditional decision-making, the accuracy of decision-making depends much on the capacity of decision-makers. However, the decision-makers must go through long experience to achieve high level decision-making capacity, and suffer from external factors and subjective factors in their decision-making process. Thus, it is important to develop methods for accurate and objectively quantitative analysis on scientific decision-making.Many decision-making problems can be attributed to classic problems in machine learning, such as classification and distance metric. This thesis adopts machine learning methods to solve decision-making tasks. Since the intelligent decision models are constructed on a given data, these models can avoid making irrational decisions due to interference of external factors or lack of experience. Similarity learning has attracted substantial attention in machine learning, and is a method of conforming to human cognition. This thesis focuses on theories and methods of similarity learning, and applies new methods into intelligent decision-making. The main work of this thesis is described in the following.This thesis proposes an intelligent decision-making method based on support vector machine similarity learning. Similarity learning requires constructing pairwise-samples. In the traditional way, every two samples are used to construct a pairwise-sample. Thus, when the data is large-scale in the given task, it is time consuming in constructing pairwise-samples, and there are too many pairwise-samples to be conducive to the subsequent processing. This thesis introduces the k nearest neighbor method to construct pairwise-samples. In this way, it can greatly reduce the number of pairwise-samples. In addition, these constructed pairwise-samples have much value for learning. Experimental results show that this method can not only decrease the running time, but also improve the accuracy.This thesis presents an intelligent decision-making method based on support vector machine similarity ensemble learning. It is well known that ensemble learning can improve the performance compared to single classifiers. This thesis introduces ensemble learning into support vector machine similarity learning, and presents an intelligent decision-making method based on support vector machine similarity ensemble learning. Experimental results show that this method can improve the accuracy and enhance the stability compared to the intelligent decision-making method based on support vector machine similarity learning.This thesis proposes a modified TOPSIS(Technique for Order Preference by Similarity to Ideal Solution) supplier selection method based on cosine similarity. Supplier selection is a multi-attribute decision-making problem, and can be solved by using TOPSIS. In TOPSIS, the Euclidean distance is used to measure the distance of each solution to positive and negative ideal solutions. Unfortunately, the Euclidean distance does not work well when there has a linear correlation between attributes. However, the cosine similarity metric outperforms the Euclidean distance metric in the case of linear correlation. Experimental results on a supplier selection problem show that the new method is simple, effective, feasible and reasonable.