The modeling and multivariate platforms in JMP put powerful analytic techniques at your fingertips. See how to make the most of the model visualization and data discovery tools. This session covers descriptive and predictive modeling, modeling when you know the relationship between X and Y, and using the Multivariate Platform to understand the structure of the X.
Modeling Roles
- Descriptive Models: Contract data to describe patterns and trends
- Predictive Models: Use patterns and trends to make predictions about new data
If you know the model form...
- Fitting a known equation vs. polynomials
How to Build Predictive Models and Compare Models in JMP Pro
- When the observations are less than 1,000, use K-Fold cross validation
- k-Fold Cross Validation
- k is the number of randomly partitioned subsets of the data
- 1 partition is used for validation, k-1 are used for training
- The process is repeated k times, the results are combined
- If n is greater than 100, K should be between 5 and 10
- If n less than or equal to 100, number of samples in each k fold group should be about 1 or 2. So, each training set is very close to the size of your data and only a few records are withheld for validation
- if n > 1,000, then use training, validation and test data partition using simple or stratified random sampling.
Training, Validation and Test Data Partitioning
- Create a Validation Column (0,1 for training and validation. 0,1,2 - for training, validation and test)
- Use the Validation Column in Fit Model, Neural and Partition platforms
- Generally, don't train on more than 1 million observations
- Use at least 30% of the data for validation
- For rare events, consider oversampling or upweighting rare observations, so the model doesn't just fit the non-rare data
- Compare R2, Root Averaged Squared Error (RASE) and Average Absolute Error (AAE)
- Visualize and Compare Models using the Model Comparison platform
Building Models with Collinearity
- Need to understand the multivariate nature of the X
