Making the move from Predictive Modelling to Machine Learning
Making the move from Predictive Modelling to Machine Learning November 16, 2017
Everyone is wanting to find out more about the way machine learning may be utilized in their business. What is interesting though is that a lot See more
Everyone is wanting to find out more about the way machine learning may be utilized in their business. What is interesting though is that a lot of companies might already be using machine learning to some extent without actually realising it.
The transfer out of predictive modelling to machine learning could be easier than you think. But prior to making that move you want to keep two important considerations in mind to make sure that you benefit from all that machine learning has to offer and that your predictive analytics strategy remains a trusted tool which lifts your business, instead of harming it: the effect of Failure and Retaining Frequency.
1. Consequences of Failure
Within the predictive analytics area, trust is important. This is particularly important if you have developed a predictive analytics solution which makes some vital decisions. If those decisions are made seriously, the impact in financial terms could be significant.
Let us consider home loans, for instance. Predictive models will evaluate new loan out applications and - based on various attributes of the applicant - will assign a score which will reflect the probability of default on such loan. If the score is too low, then the loan application is declined.
On the opposite side of this scale, efficiency models are set in place to obtain higher yield for the identical work. The results may vary, but an illustration of a reduced result of failure are a representative to customer model. The model will pick out unsuccessful and successful interactions, and also a 'compatibility model' will be developed and put into the dialler system. If the developed model stop performing, one could revert back to some random allocation of brokers to customers and a few efficiencies could be dropped, but not millions as in the event of defaults on big loans.In regard to other real world applications of high and low results of failure, a very low result of failure could be a university student's algorithm to navigate his robot securely via a group course using predictive analytics to guide it -- the worst that may occur is that the robot falls over. On the opposite end of the spectrum we have the high effect of failure of a self-driving algorithm to maintain the vehicle on the appropriate side of the street.
The result of failure appraisal will have a powerful bearing on the quantity of TLC you employ to the evolution of the model, how visible it is for interrogation and also on how frequently you make it to retrain. This brings us to the character of the machine learning is: the retraining frequency.
2. Retraining Frequency
If you visit the information science competitions website Kaggle, the case problem that they have for new starters is predicting the survivors and non-survivors in the Titanic out of a sampled training dataset.
According to Stealth Technovations, the dataset contains various attributes of their passengers, like the passenger's ticket course, the boarding gate, number of siblings, age, sex, etc.. One could build a model using whatever tool you would like and examine your algorithm on a sample and then determine how you faired against the best of the best. Do not be disillusioned if your model just comes out in 76% precision against the winner's 100%. One can simply Google the real names and see that lived and who did not. However, still, the point of this is that it is very much a static solution -- there is not any retraining of this algorithm using new info.
Contrast this with all the training of this self-driving algorithm which absorbs real-time telemetric information and retrains the model in an ongoing basis. One wouldn't need this algorithm to be retrained in the actual world.
Within financial services, the majority of the models aren't real-time in character. In the majority of scenarios, the result of failure in financial terms generally drops in the high class. For efficiency models which might be considered to be reduced result of failure, we wouldn't urge redevelopment to be done in real-time. But, one still needs the ability to build, assess and examine the model with a certain amount of care, which speaks more to the static rather than real-time retraining approach.
Therefore, if you're already building static models to get your business more efficient, you're well positioned to utilize all you've learnt to springboard your business into the machine learning domain. There are a few wise tricks which will make it possible for you to bring in the benefits of machine learning into your existing models in a secure and reliable manner, but more about this later.