Sensitivity analysis - to evaluate relative importance of marketing explanatory variables (random forest model, importance functions, caret R package and learning vector quantization (LVQ), neural networks or other). In this research it is used sensitivity analysis created in SAS JMP tool, which combines drag and drop easy approach and multiple statistical methodologies without necessity to perform R code. Coeficients close to number + or – 1 are demonstrating the most important dependencies among metrics. The similar observation can be seen in scatterplot diagrams below. Each pair of variables has a bivariate normal distribution, each ellipse contains around 95% of the points. The narrowness of the ellipse reflects the degree of correlation of the variables. If the ellipse is round, the variables are uncorrelated. If the ellipse is narrow and diagonal, the variables are correlated. As expected visits are highly correlated with impressions and links. Employee count is related to sales and capital. There is a visible dependency of KW search count and position on the Swiss market.
Color map on p-values -Map is showing significance of correlations on a scale from 0 to 1. Review of p values
Dependent resampled inputs Research uses dependent resampled inputs for 1 response variable, Sales in this case. Prediction profiler for all variables is displayed on figure below, it uses neural network for predictive modeling and predicts a response based on the responses of the k nearest neighbors. When you fit the neural network, k-fold cross validation is used. This partitions the data into training and validation sets at random. Also, Monte Carlo sampling is used to calculate the factor importance indices. Prediction profiler cells have been reordered by magnitude of the total effect. Expected significant impact is created by User ID, Category, Web, KW and City. Interesting insights are provided by importance of explanatory metrics on sales. Value visitor, Site count, Employee Count, Position on Swiss market. New BCG segmentation on the other side does not show significant impact. Based on Summary report from SAS JMP prediction profiler, in the case there is an interest to maximize sales, it would be interesting to focus in marketing on the main objective. Validation method used: KFold Number of Folds: 5 Hidden Nods in decision tree: 3 Random seed: 0
Comments