Analysis And Application Of Telecommunications Data Based On Support Vector Machine And Decision Tree

Both the rapid development of modern communication technology and the popularity of intelligent mobile terminals make people demand increasingly high quality of telecommunications services. By improving service quality, telecom operators can win customers, whose user experience of services is a basis of developing efficient strategies to improve the quality of service. It is difficult to get an accurate and comprehensive user perception by traditional ways like user satisfaction survey, service satisfaction score, a random return visit by telephone, etc. Data of Measurement Reports (MR) collected by the base station controller objectively reflects the wireless environment of the whole network, and it also reflects the "feel" of the terminal on the micro level. User experience from a statistical analysis on MR seems more intuitive, simple and one-sided, even it is a representation. User experience resulting from a data mining method applied to the MR has some important practical significance and application value. The wireless networks can be optimized in time to improve call quality if parameters that affect the user experience mined from MR can be positioned, which in turn improves the user experience.In this paper, the problem of mining user experience was converted into a prediction and classification problem, which solved by a multi-classification method based Support Vector Machine (SVM) and Decision Tree (DT). SVM is one of the important methods to solve classification problems in today’s machine learning. It is based on statistical learning theory, optimization algorithms and kernel method. It has features like global optimization ability to generalize, ability to avoid the "curse of dimensionality" and so on. It also has the advantages of small sample, high-dimensional data pattern recognition. DT is a predictive model and represents a mapping between object attributes and object values. Easy to understand and fast classification are its advantages. Because of the big data feature of MR, a pure SVM multi-classification method caused a long training time, a low accuracy, a slow prediction and other disadvantages. In this paper, a new multi-classification method based on SVM and DT was constructed by absorbing the structure of DT. The new method divided a multi-classification problem into multiple binary classification problems, which solved by SVMs, and selected the positive class and negative class to train depending on the dissimilarity. At the end of the current training, the positive class and negative class were merged into a cluster as a new class to participate in the next dissimilarity computing and training. Repeat it to the end that all classes were merged into one cluster. The structure of the classifier generated by this method is a binary DT.Based on the theory of SVM and analysis of DT’s structure feature, a study and on multi-classification method based on SVM was conducted and applied it to telecommunications data. The main tasks include:First, a detail data preprocessing was done after a careful analysis of MR, including data cleaning, data reduction, data normalization. And MR data was made categories based on received signal level (RXLEV) and received signal quality (RXQUAL).Then, errors accumulation and local optimal solution were found in the multi-classification method based on SVM and DT during the study. It tends to decrease classification accuracy and results in a bad classification. With a careful analysis of it, a new multi-classification method which conducts a multi-classification model based on Huffman tree and SVM is proposed in this paper. The new multi-classification method can effectively reduces errors accumulation and avoid local optimal solution.Finally, three experiments with MR data sets respectively verified the traditional SVM multi-classification method, the multi-classification based SVM and skewed DT and the multi-classification based SVM and Huffman Tree. The experimental results showed that the new method has a superior effect than the traditional SVM multi-classification method in training time, classifying time and classification accuracy. With a careful analysis of the experimental results, some corresponding network optimization schemes for different classification results were proposed.