Stepping aside from the Cartesian approach may reveal patterns we cannot see with the logical lens. Machine learning techniques provide an opportunity to explore patterns or relationships that aren’t obvious, or where conventional multi-dimensional statistics don’t provide clarity. Alternative approaches may prove useful.
Using neural networks is common in many different domains. The networks can self-learn patterns from known outcomes, then, adapting a model of weights and inferences, can predict what an undetermined outcome will be from available measurements. These techniques continue to be very successful in things like image recognition, classification, point of failure detection.
Here’s an example where neural networks provide a different perspective in a marketing domain for predicting cashflow, and to see effect of adjusting parameters in influencing the outcome.
Successful bids and tenders can come down to a number of different factors, from weighted score responses, through to strength of relationship with the bidder. In this case, the bidder was able to record up to 25 parameters as part of their sales & marketing initiative, for both successful and unsuccessful tenders. The Commercial Director was looking to see if there was a different way to predict the outcome of active bids.
Neural network modeling is based on relatively straightforward components: inputs - in this case the parameters recorded for each tender; outputs - success or fail; and a matrix of neurons - the matrix processing software. It’s the input data that requires most attention. In this study we analysed the parameter lists, and removed duplications (it happens, same data recorded under two headings), and dependent sets (where parameter B is closely associated with parameter A). The data set was quite well populated, but there were gaps where values could not be recorded, so we had to determine the best approach for missing data. Finally, where parameters comprised continuous variables we normalised the data. Discrete values we left as is.
Once the input data had been prepared, we separated the known outcome from the unknown, then separated the known outcome into a training set, and a test data set. This allows the neural network to adapt according to the training parameters, then test how closely the model fits against a test set, i.e. predict the outcome using the network, then check against the actual outcome. This gives a measure of how closely the neural network was able to configure to match the known output. In this study, the match was pretty close to 100%.
Now were were able to feed the unknown outcome data through the network, and predict the output.
In a lot of the cases, the output matched what an expert interpreter had set, but there were significant enough variations to warrant creating alternative income models. Interestingly, the network predicted more tender wins than the expert, purely on the recorded parameters. One advantage of this study and data set is that as new actual results come in, the model can be refined, and applied to new bids.
Process Coordination Specialists