Research On The Algorithm Of Discovery Of Markov Blanket Based On Logistic Regression And Its Application

Feature selection is an important part of Data Mining and Machine Learningresearch field, selecting correlated features from feature sets and removing featuresof no relation with the target and redundant features. Feature selection can reducethe number of variables needed by the solution of problem efficiently to solve thecurse of dimensionality problem.Bayesian Network combines probability theory and graph theory, describingthe relationships between variables in the data warehouse qualitatively andquantificationally. The Markov Blanket of a variable in the Bayesian networkcontains the parent nodes, child nodes and spouse nodes (the child node’s parentnodes).A variable’s Markov Blanket can shield the influence on the variable fromother variables of the network.In this thesis, firstly, research situation about learning Markov Blanketalgorithm is introduced, and some algorithms have the problem of containingwrong redundant nodes. Regression Analysis is a statistic method of definingcorrelativity between variables, using test of hypothesis to remove variablesweak-correlated with the dependent variable and uncorrelated variables.Useregression analysis method in the learning Markov Blanket algorithm, removing theweak-correlated with the target variable and uncorrelated variables from thecandidate Markov Blanket, and then the Markov Blanket is returned after thecondition test of independence. Compare the results of experiment on classicnetworks using this method and existed learning algorithms, which indicate thatthis method is effective and reliable.There are strong correlations between industry plates in the stock market, andthe real estate industry plays an important role in China’s economy. This thesisprocesses the data of industry plates’ close index from Shanghai stock market usingMarkov Blanket learning algorithm, selecting the industry plates correlate with thereal estate industry, and then from those plates above finds which plate hascausality with real estate plate by Granger causality test, and then uses them toestablish Vector Auto Regression Model with real estate plate to predict its dailylogarithmic rate of return, in the end analyses the model by impulse response and variance decomposition,verifying the influence relationships between real estateplate and correlated plates.