There are many flavors of problems that machine learning can help solve, but I’m most interested in problems of decision making in the enterprise space. In order to help identify the right set of problems, I’ve built a simple framework that helps me gauge good problems.
Before I get into the framework, I should define what I mean by a good problem. A good problem, in this case, is one where solving it provides a step-change in the productivity of the enterprise while enabling humans to do what they excel at.
Length of feedback loop
Humans often get better at tasks when they receive instantaneous feedback. This allow us to form mental models, & predict the outcome of our actions. This is the essence of game play. In this feedback loop is longer than a few days, it is practically non-existent. This makes it impossible for humans to get better at the task, and they end up relying on “rules of thumb” to improve their decision making. Examples of these are sales & recruiting cycles, or performance feedback cycles that last several days to weeks & months. It is impossible for folks at the top of these funnels to improve their decisions based on feedback
In these scenarios, a machine learning based approach essentially incorporates the feedback by bringing insight from the bottom of the funnel to the top of the funnel, & augments human decision making.
Irregularity of environments
Even when rapid feedback is present, if the environment is irregular, it is impossible for human beings to improve their decision making. Most recruiting & sales situations have abundant examples of this. Prospect (or Candidate) Joe under circumstances A & B has high likelihood of closing, whereas if circumstance C is present along with A & B, that deal is a non-starter. These features are impossible to extract for humans, & even when extracted, they are difficult to identify, integrate, & maintain in day to day operation.
In such a scenario, a machine learning algorithm can extract features that can guide a sales rep or a recruiter in their decision making. Further, this machine learning algorithm stays current, is personalized, & integrated into the workflow of workers.
Length of decision-distance
I define decision-distance as the degrees of separation between the ideal decision maker & their proxy. For example, in recruiting the ideal decision maker is the hiring manager or a group of interviewers. However, sourcing is typically carried out by relatively junior recruiters who are not domain experts or come from a different industry. This separation leads to exasperation & waste due to poor decision making.
Here machine learning algorithms that are trained on successful data can bring real time insight of hiring managers (to stick with an example from recruiting) to recruiters who are looking for, and building relationships.
I believe that these 3 principles should be used to examine any decision making problems in order to identify if its a good application for machine learning.