Over the years, we’ve worked on many different types of A.I. projects, from building complex analytics engines from scratch to optimizing established systems.
This week I want to talk about startups and how they can achieve quick wins.
I’m not talking about the normal eight to twelve months of development, but rather two to four months, resulting in solutions companies can use to automate a process normally done by a human, such as matching a customer to an appropriate product or answering a question during a purchase.
This is possible! But there are several challenges to this approach. Let’s dive into what to expect, how to execute an effective short-term project, and what to look out for.
Setting Your Expectations:
Data science projects usually take a long time, especially for companies or divisions that don’t already use predictive analytics. To get an idea of all the work involved, take a look at the ten-step data science process we generally use:
- Business understanding
- Data assessment and understanding
- Choose model candidates
- Data preparation
- Pipeline construction
- Modeling
- Evaluation
- Deployment
- Monitoring and maintenance
- Optimization
Just getting from step #1 to step #5 can take 12 months, and that’s before any data science actually happens (in step six).
Most of the time goes to building a rock-solid pipeline to funnel reliable clean data to the data scientists, then later to deploying the predictive models for the world (or your company) to use.
But what if we could circumvent a lot of that work and make something that shows a quick win and actually works?
It’s entirely possible, but as you could imagine, we’d be skipping over a lot of important steps. Let’s take a look at an example to illustrate this.
Illustrative Example
Let’s say we’re working on a telecom company that wants to know when a subscriber will churn, and go to a new carrier. One way to fast-track the data science process to measure this is to have a stakeholder meeting early on and ask this very important question: If you knew which subscribers would churn tomorrow, how would you use that information?
Would capturing this data in a spreadsheet be enough? Or does there need to be a “churn score” present in some legacy system or dashboard?
Do you see where I’m going? If delivering results in a spreadsheet is sufficient, this will be a lot less time consuming than e.g. integrating results into an established platform. We’re likely talking about months of time saved in this case.
Let’s assume the spreadsheet is sufficient. And we’ll also assume that it’s straightforward to pull the data each day to make our churn calculations. Sounds like we’re done. So we go build it, and each night we analyze the previous day’s data and deliver a spreadsheet report of those customers most likely to churn. This only took a couple of months to deliver.
What could possibly go wrong? Here are a few of the downsides to quick-win data science projects along with questions to ask:
- Did you spend any time educating stakeholders about how to use the churn data you’re giving them? Skepticism in the ability of an algorithm to make good decisions can be a real blocker to data science success.
- Is the data you’re grabbing each day changing in any way? Without timely work to implement checks for consistency, the data pipeline may not be delivering accurate churn results.
- What happens when things change over time? Without monitoring, the churn model might (and will!) become less accurate over time, providing results that deliver less and less business value.
- The “fix” for churn will increase the workload for those responsible for reaching out to customers and feel more like a temporary band-aid rather than a long-term solution. This will weaken the long-term success of quick-fix projects and lowers the perceived value of important analytics projects.
Conclusion
What’s the verdict? Are quick-win data science projects worth it? Would a quick-win help convince management that data science is worth the spend in the long run?
Each case is different, but picking a project with a high payoff and keeping an eye out for downstream gotchas can be a way to show early value and motivate bigger data science projects in the future. Just know this walks a fine line between utility and futility.
Here’s a related article that helped motivate today’s Tech Tuesday. I hope you enjoy it.
Be well,
-Zank
Of Interest
Ten Common Uses for Machine Learning Applications in Business
Machine learning applications are unlocking value across business functions. Here are 10 examples of how machine learning applications are being used in business:
https://searchenterpriseai.techtarget.com/feature/10-common-uses-for-machine-learning-applications-in-business
The Next Wave Of A.I. is Even Bigger
Artificial intelligence is a broad area, covering diverse fields such as image recognition, natural language processing (NLP), and robotics. A.I. technologies are also developing at what sometimes seems like a frenetic pace, so that it can be difficult to keep up to speed with everything that is happening. Have a look at this overview of some of the major near-term trends in A.I.:
https://www.nextplatform.com/2020/05/15/the-next-wave-of-ai-is-even-bigger/
Scientists Have Discovered Vast Unidentified Structures Deep Inside the Earth
Scientists have discovered a vast structure made of dense material occupying the boundary between Earth’s liquid outer core and the lower mantle, a zone some 3,000 kilometers (1,864 miles) beneath our feet. The researchers used a machine-learning algorithm that was originally developed to analyze distant galaxies to probe the mysterious phenomenon occurring deep within our own planet:
https://www.vice.com/en_us/article/ep4zvw/scientists-have-discovered-vast-mysterious-structures-deep-inside-the-earth