Scientific & Technical roadmap

Planned research and development projects over the next 24 months




Supervised Models - Regression Models

These models aim to predict a numerical variable. Thanks to the MODL approach, regression models will be robust and analytically optimized (no grid-search).



Supervised Models - Web HMI

The MODL approach provides interpretable models which facilitate the interactions with business. A web interface for viewing and interpreting the models will be developed.



Supervised Models - Gradient Boosting Machine

GMB is an ensemble approach which stackes Machine Learning models. Each model learns the error of the previous ones, with the aim of correcting it. Thanks to the MODL approach, it is possible to implement a GBM without parameters, whose stacking of models stops automatically. Thus, the models provided by Edge ML will gain even more precision without sacrificing their robustness.



Auto Feature Engineering - Co-clustering

The objective of co-clustering approaches is to group together the levels of two categorical variables. The MODL approach extends the co-clustering to the  numerical variables and optimizes the choice of the number of groups. This approach can be used as a pre-processing of any supervised model.



Auto Feature Engineering - Tri-clustering and time series

The MODL approach extends the co-clustering to an arbitrary number of variables. The Tri-clustering approach can be applied to time series encoded by the following variables:

  1. identifier of the series
  2. timestamp
  3. value of the data point

Thus, the time series are partitioned into groups and each group is described by the distribution of data points over time. This approach allows one to characterize time series as a preprocessing of any supervised model.



Auto Feature Engineering - Symbolic representation SAXO

This is a "hierarchical" variant of the MODL tri-clustering approach, which alows one to optimally encode a set of time series by sequences of symbols. This approach automatically prepares time series before the learning of any supervised model.