Test predictive solutions
The explosion of data science is giving birth to start-ups offering predictive solutions likely to help businesses provide higher customer satisfaction. In this context, the scientific community of the Customer Knowledge Decision-Making Chair has, within its scope, to test innovative predictive solutions with the view to deliver reviews that are not under any commercial constraint. The testing and following review of PredicSis.ai has been done outside any sponsorship of the Chair.
Scores serving Decision
The relationship between a company and his customers is largely based on the ability to suggest the right product to the right customer at the right time. One of the daily issues for data analysts is to estimate the probability of a given behaviour for each customer. Either a behaviour we are looking for – purchase, subscription – or a behaviour we want to avoid – churn, fraud, non-payment. This probability, captured by a score, is a quantitative support to decision-making for experts whether they are involved with marketing, distribution, risk management or any other department dealing with optimising a customer lifetime value.
Predict well, predict fast
In a market always more competitive, the production of a good quality score is not enough: it has to be produced fast! A large part of data analysts’ job today consists of cleaning data, identify attributes to take into account non-linear effects, test numerous potential predictors in order to choose the most relevant model. All these tasks are very time consuming.
Predicsis.ai: bagging of Bayesian classifiers
Naive Bayesian classifiers allow the separate estimation of each attribute’s predictive potential. But are also based on the premise that these attributes independently influence the behaviour to predict, independently to each other. Predicsis aggregate sub-models that comply with this hypothesis in order to find a trade-off between predictive quality on one hand, robustness and capacity to treat large data sets on the other hand.
A friendly interface to build a good quality score
The user access PredicSis.ai’s interface with a simple web browser. He can then choose the number
of aggregates in the model and to exclude or integrate an aggregate.
Reduce upstream data-management workload
The discretization of continuous features is done in an automatic and near-optimal way. Should there be multiple data points for a single users (eg: purchases), these are automatically aggregated, by the machine, using standard operators (counting actions, averaging values, taking the mode, etc.).
Produce reports and communicate with business experts
Adding or removing the variables that are part of a report is easily made via the graphical interface. This allows users to focus more on the specifics of a business rather than the intricacies of ML. The individual contribution of a variable on the final score is easily understandable via the graph produced. Furthermore, the user can add a comment to each feature and export the report in a pdf file to communicate internally.
The scores produced with PredicSis.ai aren’t necessarily better than those built with classical
methods (logistic regression, decision trees, for instance), but for a comparable quality, they are a lot faster to produce. With our test file, we created a score for a likelihood to convert ten time faster with PredicSis.ai than with the logistic regression method while giving the same ROC curve.