Today's attribution methodologies range from simple rules based to basic logistic regression models. Individual models have strengths and weaknesses, but a one size fits all algorithm can be inaccurate, limiting and ultimately less effective.
Conversion Logic’s solution uses a proprietary Ensemble Method, which trains multiple state-of-the-art machine learning algorithms, and combines their predictions together to deliver the most accurate, unique and flexible attribution model for each client. It is a highly dynamic machine learning framework with higher predictive accuracy than any one single algorithm.
XC LogicTM is powered by predictive machine learning models to provide the most accurate insights at a granular level. The Ensemble Method consists of four phases: 1) Data Collection and Segregation 2) Training 3) Validation and 4) Ensemble Model selection
Predictive performance scores in the Training and Validation phases are key to ensuring each client has the most accurate model. Validation is executed on a 20% holdout of the complete data set. The attribution model predicts variations for this holdout data set and historical patterns guide machine learning algorithms to predict variations.
This process ensures the models' ability to adapt, learn and perform accurately when applied, versus over-fit to learnt data patterns. Validation performance of the client’s Ensemble Model is monitored regularly (daily to weekly) and resolution steps are immediately initiated in case of deterioration in model performance.
Collected media and conversion data is divided into a training set and hold-out validation set.
Training data is modeled against several algorithms individually and in combinations.
Predictive performance of each model is evaluated on hold out validation set; Conversion Logic uses various metrics.
Ensemble determines optimal combinations and ratios of each model's predictions to use based on validation performance.