ML and TOC for getting 20/80 business solutions12.08.2019 17:35

Backlog prioritization requires understanding of relative importance. Each group of tasks in the task manager belongs to some strategy like organic acquisition, advertisement leads, conversion optimization and others. We may consider productivity of system, operational costs as well. A set of metrics/strategies is the dimension of a business model. Reduction of dimension and relative weighting, separate budgeting is sufficient. In general case, unite economy relations between profit margin and metrics are violated because of non-linearity. It«s impossible to separate acquisition and conversion, because the quantity of acquisition may affect its quality and vice versa. However decomposition of tasks requires a factor analysis (FA). FA requires a linear decomposition. We meet a contradiction.
Machine learning (ML) theory proposes a new approach to reduce dimension and get weights. ML combines feature weighting with nonlinear output functions like logistic function or neural network. The approach is based on practical question — what set of metrics is enough to predict the business goals (offers, revenue) with acceptable accuracy like 90%? ML proposes an iterative process. It was applied in decision trees class but may be generalized at ensemble models. At first we analyze the influence of a full group of metrics — high dimension task. Second iteration includes throwing away one of N metrics with minimum affect on prediction power. Prediction is made inside the historical data — out-of-sample approach. N combinations of features are tested with the same (!) prediction model. It is a transition from N to (N-1): N=>N-1. Cross validation and out of sample testing techniques lies inside the model. Part of mixed historical data is used to get prediction for the out-of-sample part of data. Iterative process goes on until the threshold of 80% is reached.

It seems that we sacrifice accuracy to simplify our model and number of sufficient metrics. It is not a true in ML and nonlinear world. Large number of features means unstable prediction and over fit sometimes. Actually dimension can be reduced with increase of accuracy — yes, it«s true. More complex system may be described by smaller number of parameters due to the nonlinear tie-up. Non-linearity means simplification, not complication. This idea was remarked by Eliyahu Goldratt — founder of theory of constraints (TOC). If acquisition (number of sessions, leads) and conversion are closely connected — it is a problem on the one hand. Unit economy is useless. But on the other hand we may use a single parameter and a single strategy to boost sales goals in nonlinear CJM and delayed conversion. In fact we may select a parameter/strategy which is more valuable and more cheap — simple tradeoff. We get 20/80 return due to the non-linearity.

The example of web strategy evaluation is given here. Two groups of metrics are considered — traffic and conversion. It«s shown that a single traffic measure /number of sessions/ is enough to describe commercial offers without (!) loss of accuracy. In the considered case, traffic attraction through SEO is cheaper than other strategies as well. However the process of evaluation should be restarted periodically. Dynamic market environment should be considered. The basic constraint — factor of highest business efficiency — is always wandering through the supply chain or customer map. The input features may be replaced by operational costs, revenue, fixed costs and others. It doesn’t matter because the approach may be applied in any nonlinear problem. At general case all business problems are nonlinear in nature and can’t be decomposed.

Thanks for Karma.