Marketers, Get Used To Disappointment.
"Who are you?!"
"No one of consequence."
"I must know."
"Get used to disappointment."
I've been meaning to write a short post about the game-changing impact of machine learning on consumer marketing. Watching The Princess Bride again recently with my kids provided the perfect moment of inspiration. EB
With the rise of mobile-first engagement, voice interfaces and IoT, consumers are increasingly comfortable with and expecting individualized experiences. These require their personally identifiable information (PII) is shared with and utilized by brands and service providers. Privacy and security are still paramount of course, but convenience, efficiency, and real value delivered across touch points by innovative companies have changed consumer thinking.
For marketers and data analysts already working to bridge the gap between first-party customer data and third-party anonymous audiences (for profile enrichment and lookalike prospecting), machine learning represents a new and exciting power tool. But as with any new approach, hype drives a disconnect between expectations and reality. And capitalizing on ML means rethinking the nature of customer identity and experience -- and approaching the customer in a fundamentally new way. One that, counter-intuitively at first, requires embracing uncertainty and imperfection.
For non-engineers and non-data scientists, initial exposure to AI and ML usually comes from (i) competitor claims and (ii) the business press. No surprise that it is nearly impossible to separate ground truths from marketing fiction. I've seen the game changing benefits of machine learning for consumer brands first-hand, as well as its limitations. There are plenty of technical primers on what machine learning is (and is not) and how it works, so I'm skipping the fundamentals here and jumping straight to the implications for marketers and analysts.
Machine learning is statistics -- with a feedback loop. OK, so I lied. This is a bit of a primer, written in my own words as a layman (get used to disappointment!). But important if you're going to view and adopt the points below as credible best practices.
Basically machine learning is data analysis through software algorithms that assigns probabilities to less-than-exact (but also sometimes exact) outcomes. Algorithms are trained by comparing calculated values against prevalidated training data. The learning part of machine learning means the algorithms become more accurate and/or accommodating as more data is encountered, without the need for constant code rewrites. In practice, both the application and the training data need to be highly domain-specific for the algorithm to perform well.
Once trained, ML infers or estimates future values (like whether two customers are the same person) -- and the confidence, or certainty, of these estimates can range from 0 to 100% for a given record or data point. As new data is added over time, confidence may increase or decrease depending on the new information. I thought "John Smith" was the same person as "John A. Smith", but now that I know his middle initial is "B.", my confidence score is much lower (and my estimate more accurate).
Machine learning expands your universe of actionable data. In practice, marketers throw out the vast majority of customer identity and behavior signals they collect because it is impossible to match or assign this data with 100% confidence. These perfect matches are often referred to as "deterministic" or "validated". The cost of certainty is high, burdened by manual data cleansing and processing. Without either perfect data or machine learning, the risk of customer disappointment means marketers limit themselves to use cases tailored to validated data. But this changes once we have transparent confidence scores for individuals and attributes, and we can unlock use cases built upon less-than 100% data confidence. Think of all that data I can utilize when my approach allows for uncertainty.
Machine learning demands and rewards the uncertainty trade-off. So machine learning requires both a comfort level and practical, working approach to uncertain data. A deterministic (exact) query would profile and segment customers as "female residents of San Francisco that converted online in the last 30 days with a high product affinity in shoes based on her last 3 purchases." This query might return a segment representing 1% of current customer count. The same query written for a probability-driven (or probabilistic) approach might read, "80%-likely female customers, 72%-likely residents of San-Francisco, 50%-likely to have converted online in the last 30 days, and with a 64% chance of having high affinity for shoes." This segment might reach 5% of current customers. You've just increased your addressable market for an offer by 5X or 500% by exposing and embracing the uncertainty inherent in your data. It's machine learning, working in the background, that estimates and infers the underlying connections and their probabilities.
Machine learning contributes the most when you embrace uncertainty. With machine learning and estimated outcomes, the risk of getting the answer wrong increases greatly. For this reason marketers must tailor each message and offer (and the customer experience these are intended to support) for the specific data and confidence scores available. You won't be refunding purchases or changing airline tickets with anything less than 100% certainty. On the other hand, the value of presenting a discount promotion or a personalized product recommendation even if it's wrong some of the time, can unlock huge value for your customers. Your baseline (for both customer experience and marketing performance) is the limited market of customers where perfect information exists today. Machine learning allows you to throw off those constraints and leverage more (and eventually most) of the data you have regarding your customers' identity, preferences, and behaviors. You just have to get used to disappointment.