Today’s blog is brought to us by Dr. Paul Pallath, Chief Data Scientist & Senior Director with the Advanced Analytics Organisation at SAP
Change is inevitable. This is an undisputable fact of life. Businesses have constantly changed and adapted to the ever changing requirements of their customers and/or markets. Businesses who were resistant to change, or did not adapt to the changing landscape, have gone out of business. We see this constantly where big multinationals are putting a big thrust towards innovation and new ideas, and the influx of funds available for start-ups. There are examples of start-ups in the landscape of well-established players, constantly challenging the landscape by disruptive ideas.
In recent years, several organizations all over the globe have embarked on or have taken the initial steps in the direction of digital transformation and improvising business process around the digital core. Predictive is one of the core elements in the digital transformation journey that empowers the business to gain competitive advantage.
Predictive analytics are most powerful when they are operationalized and are part of the business process, and they enforce a lot of changes in more ways than one. For example, for customer targeting, predictive models help choose the customers that are most likely to respond when reached out to, which helps businesses save on marketing and increases return on investment.
Predictive models have influenced the business by proposing how, when, where and which customer should be targeted, influencing change in behaviour of the customer.
As a consequence, the behaviours captured in the new data have now been influenced by the predictive model.
Similar observations can be made for several business problems that have found interventions via predictive models.
Predictive Models and Concept Drift
The concept that was modeled by the predictive models has enforced changes in the “knowledge” captured in the data. Patterns in the data evolve and the predictive models that have been created in the past will become obsolete over time. This is also termed as concept drift. Since the predictive models don’t have endless life, and need to be monitored constantly, any changes in the concept drift necessitates the regeneration of predictive models on new data.
Concept Drift and Historical Data
Concept drift is not only true of new incoming data, but it is also true for historical data. While building predictive models for solving a business problem, one could find different segments in the data that exhibit different behaviour. This could be due to several reasons like segments based on geographies, products, customer demographics/profile, type of business and so on
Thus, a “one-size-fits-all” approach for predictive modelling is hardly possible and not desirable for concept drifts induced because of these reasons. This implies there could be several predictive models (sometimes in the order of thousands) for addressing a business problem.
How to Adapt Models without Impacting Business
Thus change is inevitable in the predictive context—and so it’s important that businesses be able to monitor, evaluate, and automatically retrain or decommission the models without impacting the business.
SAP Business Objects Predictive Analytics 3.0 enables businesses to tackle these problems via Predictive Factory capabilities. These factories operationalize and schedule several thousand models in real time, detect data and models deviations, and provide capabilities to automate several aspects that are important for operationalizing the models.
To learn more about predictive analytics, read the other posts in our Predictive Thursdays blog series.