Predictive Analytics Exam

Interaction (definition) - answer-Effect of one variable on target variable depends on the value/level of another variable. GLM Description (idea) - answer-GLMs relate a function of target mean linearly to a set of predictors. Target itself is not transformed. Recursive binary splitting (idea) - answer-divide feature space recursively into a set of non-overlapping regions of relatively homogeneous observations until a stopping criterion is reached. Random Forests (idea) - answer-Reduce overfitting and variance of base trees by combining the results of multiple trees fitted to different bootstrapped samples of training data in parallel. Reduce correlation between different trees by taking a random sample of predictors as candidates for each split. Boosting (idea) - answer-In each iteration, fit a tree to the residuals of the preceding tree and a scale-down version of current tree's predictions is added to the previous predictions. Each tree focuses on observations that the previous tree predicted poorly. Backward step-wise selection - answer-start with the full model with all features and sequentially remove features until no improvement in selected measure. Forward Stepwise Selection - answer-start with the null model with no features and sequentially add features until no improvement in selected measure. BIC vs. AIC - answer-BIC is more conservative and results in simpler models. Binarization - answer-Replacing a value with 0 or 1 based on threshold. pros: drop individual factor levels statistically insignificant with respect to the base level. cons: more steps in the stepAIC procedure and model may become harder to interpret. Regularization (idea) - answer-reduce overfitting by shrinking the size of the coefficient estimates, especially those of non-predictive features. Reduction occurs by optimizing the log-likelihood function adjusted by a penalty term that reflects the size of the coefficients. Principal Component Analysis - answer-a statistical method to simplify the description of a set of interrelated variables. Its general objectives are data reduction and interpretation; there is no separation into dependent and independent variables; the original set of correlated variables is transformed into a smaller set of uncorrelated variables called the principal components. To reduce the dimensions of a dataset via a set of linear combinations of the original features, known as PCs, generated to capture as much information in the dataset, with respect to variance, as possible. Cluster Analysis - answer-a technique used to divide an information set into mutually exclusive groups such that the members of each group are as close together as possible to one another and the different groups are as far apart as possible. Pros of GLMs - answer-excel in accommodating a wide variety of distributions for target. model equation clearly shows how target mean depends on the features, which coefficients providing a readily interpretable measure of the directional effect each feature exerts on the target. Simple to implement. Cons of GLMs - answer-unable to capture non-linear relationships unless manually factored in advance for some link functions, the coefficients may be difficult to interpret Pros of regularized GLMs - answer-via the use of model matrices, binarization of categorical variables is done automatically and each factor level is treated as a separate feature to be removed. For lasso, variable selection is automatically done by cross validation. Cons of regularized GLMs - answer-coefficient estimates are more difficult to interpret since variables are standardized limited/restricted model forms allowed by the glmnet function Pros of Single Trees - answer-If there are not too many buckets, trees are easy to interpret because of the if/then nature of the classification rules and their graphical representation. Excels in handling nonlinear relationships (interactions) categorical predictors are automatically handled without the need for binarization variables are automatically selected as part of the model building process. Variables that do not appear in the tree are filtered out and the most important variables show up at the top of the tree Cons of single decision trees - answer-more prone to overfitting than GLMs even with pruning strongly dependent on training data. predictions are unstable with high variance leading to lower user confidence. Pros of ensemble trees - answer-much more robust and predictive than base trees by combining the results of multiple trees Cons of ensemble trees - answer-opaque and difficult to interpret because many trees are used. Require variable importance measures such as partial dependence plots. Difficult to implement because of the huge computational burden with fitting multiple base trees. Offset - answer-To incorporate a measure of exposure. Larger exposure leads to larger Mu. Weight - answer-To incorporate a measure of precision. More precise observations receive higher weights. Unbalanced Data - answer-One or more groups are predominant which is problematic because model places more weight on the majority group and tries to fit those observations well when the minority group is of interest. Introduction to executive summary - answer-This report provides a preliminary investigation into the factors that contribute to the target variable. As requested, we have applied predictive analytic techniques to develop a predictive model based on the data provided to identify important factors associated with the target and accurately predict the target. Significance of a cutoff - answer-All classification problems take predictors as inputs and return the predicted probability of the binary target. To get the predicted class requires that we specify a cutoff. Predictions that exceed this cutoff will be predicted as the positive class. While ideally we want both sensitivity and specificity to be high, we need to strike a good balance between the two by selecting the cutoff carefully. Note on Ethical Considerations - answer-Given our model will be used to predict the target, the use of demographic variables such as gender and age, and other senstive information migh raise discriminatory or privacy concerns and should be reviewed prior to implementation of the models. For the purposes of this analysis, these variables were reta