When people use machine learning without investing enough time in planning, it’s no surprise that the results will be inferior.
Project managers and other stakeholders who fall into this situation are likely to see a lack of accuracy in the algorithm, or “bad intelligence” provided by the machine learning program. This can cause all kinds of problems, and it can send the human interpreters of the information in the wrong direction when trying to meet business goals and desired outcomes.
Different types of machine learning bias can make ML programs produce bad insights and dirty data can cause people to make bad business decisions. When this happens, leaders have to deal with the sunk costs of investing in machine learning projects that don't turn out well. (Read also: 4 Myths About Starting a Machine Learning Project.)
They may tend to blame the computer or artificial intelligence (AI) in general, when the problem was really caused by weak algorithms, biased data, poor data quality. (Read also: Quality Data: Why Diversity is Essential to Train AI.) Then there's the problem of overfitting and underfitting, which can lead to one of two things – either the results will be too narrow to be useful, or the results will be too broad to be of any use. This is a lot of what experts are talking about when they discuss fitting and dimensionality in machine learning.
Companies that invest in ML projects too quickly are likely to see poor outcomes. Developing the requirements for implementing machine learning models should be an iterative process in which stakeholders start out by defining how machine learning will support a business goal. Once a machine learning model has been selected, it's important to make adjustments on a continual basis. For all of these reasons, it's a good idea to proceed with adequate scope work and orientation for an ML project.
- PHASE I: Develop a plan for how ML will support a specific business goal.
- PHASE II: Seek out a commercial off-the-shelf (COTS) ML service that will help you quickly build a prototype.
- PHASE III: Test the ML prototype in a simulated (sandbox) environment to ensure outcomes are as desired.
- PHASE IV: Monitor the results on a continual basis and take note of learnings that can be used in future projects.
