Machine Learning That Supplements Human Judgement
I often get asked how SpringML’s machine learning is different from other software companies that do machine learning and predictive analytics. Here are three ways we’re different.
- Predictions from our models are interpretable. We explain why the model thinks an opportunity is likely to win or provide details on why a forecast for a particular month is weak. We believe that unless such details are provided the predictions will not be trusted by our users in sales operations or sales management. This is especially important because unlike some areas where machine learning models are used to make automated decisions (e.g. display an ad or not), our models are used to supplement human judgement.
- The primary factor that impacts effectiveness and accuracy of any predictive model is the treatment and transformation of data (aka feature engineering or data engineering). Our models are built to consume structured and unstructured data from a customer’s Salesforce system. We do not combine data across customers. Opportunity, Opportunity History, Account, Product and Activities are some of the objects our models consume. This data is then processed in-memory and meaningful features and signals are extracted and passed as input to the model.
- There are fairly robust techniques in Machine Learning to test the accuracy of a model (e.g. cross-validation). However, nothing can replace comparing predictions made with actual outcomes once they’re available. We continuously monitor and measure our models implemented for our customers, which leads to better results. For example, forecasts made by our models consistently come within 90% of actuals.