|Rosella Machine Intelligence & Data Mining|
Nine Steps Predictive Modeling Guide for Risk Management
Developing and deploying predictive models is not a trivial job. It requires a methodical approach. In this page, nine step quick guides for of predictive model development are described.
1. Data Cleaning and Transformation
The first step to predictive modeling involves data cleaning and transformation. Data may contain bogus values, synonymous values, outliers, etc. Those values need to be standardized and cleaned. In addition, values may need to be transformed into new variables. Note that data is cleaned and transformed from database systems.
2. Import Data into CMSR Data Miner / Machine Learning Studio
Once data is cleaned and transformed, the next step is to import data into CMSR Data Miner / Machine Learning Studio. Imported data will be used for variable analysis and predictive modeling.
3. Variable Relevancy Analysis
Once data is imported into CMSR Data Miner / Machine Learning Studio, variables are analysed to determine whether certain variables contain information that can predict customer tendencies. Categorical bar/histogram charts and correlation analysis tools are used. For more, read Variable Relevancy and Factor Analysis.
4. Predictive Modeling
Once relevant variables are identified, neural network predictive models can be configured and trained using data imported into CMSR Data Miner / Machine Learning Studio. Only those variables identified as relevant are used in modeling. Multiple models with differing configurations may be used. Testing of model accuracy is performed using (historical) data stored in database.
5. Model Integration - Aggregation
Multiple predictive models are integrated as single models using RME-EP (Rule-based Model Evaluation - Event Processing) rule engine. Integration can take various forms such as maximum, minimum, average, etc. Those combined score values are tested using historical data (in database) and analysed.
6. Model Integration - Classification
Next step model integration may involve classification. Predicted values from neural networks are hard to understand for casual users. Those numbers are translated into more easily understood vocabularies such as “very high risk”, “high risk”, “medium risk”, “low risk”, etc. This classification is done with RME-EP. RME-EP models need to be tested for accuracy with historical data in database.
7. Preparing Model Documentations
Once models are integrated and tested, end-user model documentations are prepared. Model documentations are HTML files that can include charts relating to models. Model documentation should contain information what end-users will need to know.
8. Deployment on MyDataSay
Models along with model documentations can be deployed on Android devices using MyDataSay Android App. MyDataSay can be served as trial before deploying on Web.
9. Web and Database Integration
Models along with model documentations can be deployed on web using Rosella BI Server on Java/Jakarta EE Application Server. In addition, BI server provides database and other web application integration using JSP programs. For example, predictive information can be integrated into other web-based business applications using HTML IFRAME tags and JSP pages.