GC Intelligence: Predictive Modeling Security Overview

GC Intelligence: Predictive Modeling

Predictive modeling helps GiveCampus make smart recommendations for key workflows throughout your use of the platform, such as the dollar amount to recommend to a constituent on a giving form sent, or which constituents to engage for a particular type of ask.

This document summarizes key frequently asked questions regarding how we generate predictive modeling data and insights for use by our customers.

Q1: Who are your third-party vendors for Machine Learning? 

We do not use any third-party vendor(s) and all our work is conducted in-house. We do leverage open-source code libraries to implement well-known algorithms, such as Neural Networks, Gradient Boosted Decision Trees, Random Forests, and Support Vector Machines. 

Q2: Is my data mixed with other customer’s data or used to train any models?

With respect to our Predictive Modeling service: our ability to build an accurate predictive model is predicated on training the model against a vast amount of anonymized data from over one thousand colleges and universities. However, one customer’s data is never exposed to another customer.

Q3: How do you evaluate the effectiveness of your models?

With respect to our Predictive Modeling service, we gauge effectiveness based on real-world results (e.g., did the model reliably predict the desired outcome) and metrics. For example, for classification models used for GC Smart Segments, we use the following:

• AUROC (Area Under the Receiver Operating Characteristics Curve): AUROC (or AUC) is one of the most commonly-used measures of performance for machine learning classifiers. AUROC is derived from the ROC curve, which plots how well a model can strike a balance between maximizing true positives and minimizing false positives. It is represented as a percentage from 0-100%, where 50% is the expected performance of a “random” model.
• Precision: Of the constituents that the model predicted to be in the positive class (e.g., likely to convert), what percentage were actually observed to be in the positive class (converted)?
• Recall: Of the constituents that were observed in the positive class (e.g., converted), what percent did the model predict to be in the positive class (e.g., likely to convert)?