Machine Learning: Are You Ready? A 7-Part Checklist
Machine learning is all the rage. But while the topic is top of mind in boardrooms and the media alike, it is not always clear how machine learning is best applied. Or what it takes to implement. While the tools themselves are getting easier to use, machine learning projects are not just add data to algorithm and stir.
Ultimately, machine learning is a synergistic exercise between man and machine. Machine learning in practice requires human application of the scientific method and human communication skills. Successful organizations have the analytic infrastructure, expertise and close collaboration between analytics and business subject matter experts to translate these synergies into ROI.
Is your organization ready? Here are seven signs your organization is ready to forge forward with machine learning:
#1 Articulated a Problem That Needs Solving
Like any other technology, machine learning works best when there is a clear problem statement and outcomes defined. Routine or repetitive decision points that are high-volume, require rapid response and/or where potential actions are defined but dependent on variable inputs are good candidates for machine learning.
It works especially well for applications where applicable associations or rules might be intuited but are not easily described by logical rules, when accuracy is more important than interpretation or interpretability, or when the data is problematic for traditional analytic techniques.
Because machine learning is time and data intensive, a critical evaluation of whether existing analytic models/approaches or alternate solutions may apply is also in order. This ensures potential value is commensurate with input effort.
#2 Established an Experimental Mindset
Machine learning is an iterative, experimental process. Although core algorithms are increasingly commoditized, every project must be customized based on the business context and data.
As with any good experiment, some hypotheses will be proven false. New data may need to be procured or created. Or the problem statement recast based on what is found. As a result, decision makers and team members alike must adopt a test-and-learn mentality for machine learning to succeed.
Use a gated, iterative process that provides the flexibility and agility to quickly assess progress to determine whether an alternate approach is warranted or when enough is enough.
#3 Enlisted a Collaborative Data Science Team
To invest in machine learning and see results, you can’t just invest in the technology. You also need to make sure you have the right people in place to guide the systems and allow them to create the most impact.
Yes, machine learning expertise is a requirement. Equally important is a dynamic teaming model that engages diverse experts with business, data and technical expertise. This includes data experts that can assess and onboard requisite data assets, business experts to provide context, assess implications (business, social, moral) of proposed actions or new product or service offerings and IT personnel who deploy and maintain the technical ecosystems.
Don’t forget about resources who can translate between the language of the “quant” (i.e. those who speak math) and that of the business.
#4 Developed a Robust Data Strategy and Ecosystem
Machine learning runs on data. A lot of data. Establishing a data process to effectively identify, acquire, provision and access high-quality data and information assets is critical.
To that end, governance policies and the data ecosystem must support exploratory environments (often referred as sandboxes) as well as production environments. This requires a multi-tiered approach to balance access and agility without sacrificing security, privacy, or quality.
The introduction of non-traditional (big) data sources including unstructured text, voice, pictures and so on may also require new data management capabilities.
#5 Assessed the Organization’s Risk Tolerance
From agreeing on the criteria for what is “good enough” to understanding how to validate and develop models, machine learning often challenges traditional approaches to quality assurance and risk management. Why? At some point, the training wheels or, in this case, the training data must come off.
Real validation comes from testing the performance of the machine against new data. Very often, this entails putting the system into operational practice. This may range from performing “A/B testing” in production to confirm a model will incent desired customer behaviors to taking a self- driving car on the road with a human overseer ready to take the wheel in a crisis. Considering your organization’s perception and tolerance of the risk (which will vary based on the application at hand) allows appropriate guardrails and change management to be proactively addressed and promotes adoption, rather than resistance to requisite business changes.
#6 Committed to Adapting Established Business Processes
Whether automating an existing decision point or enabling a net new product or service offering, machine learning is disruptive. Assessing potential implications to existing business processes, functions and roles is key. This doesn’t mean architecting the entire change before you start. But a quick gut check can mitigate the potential for costly moot exercises. To plan for the plan, begin by asking: “If we answer this question or provide this hypothesis, what can we do with the information? How might this impact existing processes? Are we willing and able to make the requisite changes?” Machine learning can provide new answers to business questions, but it’s just noise unless the organization is prepared to put insight into action.
#7 Committed to Adopting New IT Practices
Once deployed, iterative modeling and tuning of the machine learning model must continue. The cadence at which updates are required is unpredictable, not conforming to traditional scheduled deployment windows. As a result, deploying machine learning requires fundamentally different QA and deployment models, skill sets and service levels than traditional IT DevOps practices. In this case, operationalization is not a merely keep the lights on activity. Maintaining the model is a critical, ongoing process that requires as much due diligence as initial model development.
About the author: Kimberly Nevala is the Director of Business Strategies for SAS Best Practices.