Report

Predicting Income from Census Data using Multiple Classifiers Presented By: Arghya Kusum Das Arnab Ganguly Manohar Karki Saikat Basu Subhajit Sidhanta CSC 7333 PROJECT, SPRING’ 13 LOUISIANA STATE UNIVERSITY Agenda Objective Data Methods Artificial Neural Network Normal Bayes Classifier Decision Trees Boosted Trees Random Forest Results Comparisons Observations CSC 7333 - Dr. Jianhua Chen 2 Objective Analysis of Census Data to determine certain trends Prediction task is to determine whether a person makes over 50K a year. Analyze the accuracy and run time of different machine learning algorithms CSC 7333 - Dr. Jianhua Chen 3 Data • 48842 instances (train = 32561, test = 16281) • 45222 if instances with unknown values are removed (train = 30162, test = 15060) • Duplicate or conflicting instances : 6 • 2 classes : >50K, <=50K • Probability for the label '>50K' : 23.93% / 24.78% (without unknowns) • 14 attributes : both continuous and discreet-valued. The Attributes • • • • • • • • • • • • • • Age Workclass fnlwgt Education Education-num Marital-status Occupation Relationship Race Sex Capital-gain Capital-loss Hours-per-week Native-country Data SnapShot Artificial Neural Network • Sigmoid function is used as the squashing function. • No. of Layers = 3 • 256 nodes in first layer. Second and third layers have 10 nodes each. • Terminate if no. of epochs exceed 1000 or rate of change of network weights falls below 10-6. • Learning rate = 0.1 Normal Bayes Classifier • The classifier assumes that: • Features are fairly independent in nature • the attributes are normally distributed. • It is not necessary for the attributes to be independent; but does yield better results if they are. • Data distribution function is assumed to be a Gaussian mixture – one component per class. • Training data Min vectors and co-variance matrices for every class Predict Decision Trees Regression tree partition continuous values Maximum depth of tree = 25 Minimum sample count = 5 Maximum no. of categories = 15 No. of cross validation folds = 15 CART(Classification and Regression Tree) is used as the tree algorithm Rules for splitting data at a node based on the value of variable Stopping rules for deciding on terminal nodes Prediction of target variable for terminal nodes CSC 7333 - Dr. Jianhua Chen 9 Boosted Trees • Real AdaBoost algorithm has been used. • Misclassified events Reweight them Build & optimize new tree with reweighted events Score each tree Use tree-scores as weights and average over all trees • Weak classifier classifiers with error rate slightly better than random guessing. • No. of weak classifiers used = 10 Trim rate Threshold to eliminate samples with boosting weight < 1 – trim rate. Trim rate used = 0.95 Random Forest • • • • • Another Ensemble Learning Method Collection of tree predictors : forest At first, it grows many decision trees. To classify a new object from an input vector,: 1. It is classified by each of the trees in the forest 2. Mode of the classes is chosen. All the trees are trained with the same parameters but on different training sets Random Forest (contd.) • No. of variables randomly selected at node and used to find best split(s) = 4 • Maximum no. of trees in the forest = 100 • Forest accuracy = 0.01 • Terminate if no. of iterations exceed 50 or error percentage exceeds 0.1 Results Unknown data included Method Correct Classification Class 0 false positives Wrong Classification Class 1 false positives Time Accuracy Neural Network 13734 2547 1339 1208 719 0.84356 Normal Bayes 13335 2946 1968 978 3 0.819053 Decision Tree 13088 3193 1022 2171 5 0.803882 Boosted Tree 13487 2794 1628 1166 285 0.828389 Random Forest 13694 2587 864 1723 51 0.841103 Unknown data excluded Method Correct Classification Class 0 false positives Wrong Classification Class 1 false positives Time Accuracy Neural Network 12711 2349 1804 545 545 0.844024 Normal Bayes 12226 2834 1945 889 3 0.811819 Decision Tree 12017 3043 983 2060 4 0.797942 Boosted Tree 12260 2800 1510 1290 221 0.814077 Random Forest 12621 2439 850 1589 48 0.838048 CSC 7333 - Dr. Jianhua Chen 13 Comparisons (unknown data included) Accuracy Time 0.85 800 0.84 700 0.83 600 500 0.82 400 0.81 300 0.8 200 0.79 100 0.78 0 Neural Network Normal Bayes Decision Tree Boosted Tree Random Forest Neural Network Class 0 false positives 2500 2000 2000 1500 1500 1000 1000 500 500 0 0 Normal Bayes Decision Tree Boosted Tree Decision Tree Boosted Tree Random Forest Class 1 false positives 2500 Neural Network Normal Bayes Random Forest Neural Network Normal Bayes Decision Tree Boosted Tree Random Forest Observations Removing non relevant attributes improves accuracy (Curse of Dimensionality) Some attributes seemed to have little relevance to salary. For example: Race, Sex. Removing the attributes improves accuracy from by 0.21% in decision trees. For Random Forest, accuracy improves by 0.33% For Boosted Trees, accuracy falls slightly by 0.12% For ANN, accuracy improves by 1.12% Bayes Classifier – Removing co-related attributes improves accuracy. Education-num highly related to Education. Removing education-num improves accuracy by 0.83% CSC 7333 - Dr. Jianhua Chen 15 Thank you!!! CSC 7333 - Dr. Jianhua Chen 16