When it comes to predictive analytics, it’s no secret that majority of predictive models, machine learning or otherwise, never get deployed (See www.kdnuggets.com/2022/01/models-rarely-deployed-industrywide-failure-machine-learning-leadership.html for a good summary of model deployment research). Operational silos, technology challenges, and lack of leadership are often the main hurdles. While we may not be able to help organizations on the leadership front, we can definitely help on addressing the first two hurdles. Our decision management platform, SMARTS™, allows business analysts and data scientists alike to define, test, deploy, and manage operational decisions. This not only includes business rules, but also any predictive models that drive these decisions.
Standards-Based Model Import and Deployment
Traditional methods for deploying predictive models typically involve data scientists handing off a model to IT for manual coding and implementation in operational systems. This process is time-consuming (and near impossible with machine learning models), error-prone, and risks crashing the whole system.
SMARTS™ uses the Predictive Model Markup Language (PMML) standard to import and deploy predictive models. PMML is an XML standard for the interchange of predictive models developed by the Data Mining Group (dmg.org). Organizations can move PMML models among different tools and applications without the need for custom coding. In addition, organizations can develop PMML models in commercial or open-source data mining tools (such as SAS, SPSS, KNIME, R or Python) and deploy them directly in a SMARTS™ decision service.
With SMARTS™, a data scientist can export a PMML model from their data-mining tool of choice and hand it off to a business analyst. Next, the business analyst can import that model into the SMARTS™ environment to write business rules to transform model insights into business decisions. These model-driven business decisions can be tested and deployed as a decision service (batch or real-time) in a scalable operational environment. As a result, organizations can deploy predictive models faster, adapt to market changes, and reduce overall IT costs.
SMARTS™ supports a wide range of predictive models expressed in PMML 4.2 including:
- Neural Networks
- Support Vector Machine
- Random Forest
- Tree
- Scorecard
- Ruleset
- Naïve Bayes classifiers
- Clustering
- Regression (General, Multinomial, Cox, Linear/Log)
- Mining (for example, Random Forest Models)
Operationalize Predictive Analytics in Many Applications
SMARTS™ is paving the way for more organizations to unleash the power of predictive analytics in their operational decision-making. Examples include:
- Banking: payment fraud, risk management, loan approvals, credit scoring, collections, customer acquisition and retention, up-sell/cross-sell
- Insurance: automated claims processing, rating engines, claims fraud, document configuration, customer retention, up-sell/cross-sell, best-fit recommendations
- Healthcare: health evaluations and recommendations, appropriate-use criteria, readmissions and outcome management
- Retail: up-sell/cross-sell, campaign optimization, social media management, next-best offer
- Manufacturing and Utilities: quote generation, predictive maintenance, asset and inventory management, warranty claims
Improve Predictive Models through Experimentation
SMARTS™ decision analytics capabilities allow both business analysts and data scientists to measure and improve the quality of predictive models in the context of the actual automated decisions:
- Pre-deployment, users can set up and run large-scale simulations to compare different decision strategies and identify the winning strategy to deploy.
- Post-deployment, users can set up and run Champion/Challenger experiments to continually test new strategies on live data.
Cloud and On-premise Deployment
SMARTS™ offers flexible deployment options that works with your organization’s infrastructure:
- For organizations looking for a cost-effective, scalable solution, SMARTS™ is available in a cloud environment as a fully hosted SaaS solution.
- For organizations that need to manage their data within their own physical environment or need to execute behind the corporate firewall, SMARTS™ is available on-premise as a fully-configured virtual-appliance that is installed in your virtualization-hosting environment.
Learn more about SMARTS™ AI & ModelOps capabilities.