I’ll start out this discussion track with a cartoon from Marketoonist Tom Fishburne and a general assumption that machine learning/AI is something you will look at, explore or invest in at some point in 2019.
According to most analysts’ reports, this is one of the hottest topics and investment areas for digital marketers. What gets lost in evaluation is the general knowledge gap: knowing what your options are based on the data you have and the real questions you hope to answer.
A Post-It note on my desktop lists 10 key things any leader should do when starting a new company. One is contextual and subtle but so important: “Match Strategy to Situation.”
In this world of choices and tons of digital disruptions, what is your strategy this year? And, how are you leveraging advances to create scale and competitive advantages?
Is 2019 the year of machines? Email has long been a manual production process that is fraught workarounds and many opportunities to make mistakes. But the data trove continues to build with each interaction and purchase on file.
Building models to identify high-value customers or new purchase propensities are often the top-level views. But when you get in the trenches with email, how far can you really take it? Does it require multiple vendors? Multiple data sets? A data scientist to explain the options and risks?
I believe in the value of machines to accelerate decisions, automate manual processes and rethink how we do things in the spirit of enabling digital transformation that sets a company apart. I believe optimizations can happen, machine enabled, and some that are purely exploratory in designing a user experience that includes email and other channels.
So, in the spirit of innovation and thinking differently, what role will machines play in creating that competitive advantage in 2019? Can you add innovation without operational disruption? Or, do you need a catalyst to disrupt legacy ways of doing things that seem to work, but aren’t setting you apart.
I guess the crux of the question is this: Are you satisfied and safe? Or, are you striving to do things differently, better, faster, smarter?
So, not to let this get too far out of scope, advances in machine learning available today for email marketers center on four areas, yet it’s a much broader use that I’ve listed below to help normalize how you think about data, application, and methods used. We’ll discuss these in more detail next week:
- Predictive analytics (supervised machine-learning algorithms)
- Recommenders (content/product)
- Optimization/testing engines
- Natural language processing
Most people in the email marketing space know about a few of these because they’re offered by ESPs or tangential providers that specialize in AI/Machine learning for email.
Many have product recommendations to help with identifying what people are buying, what they might buy and general classification systems that help you address bundling, discounting and offer/upsell strategy.
Others offer content recommenders that will help you personalize email and your site with the content that is most ideal to effect a purchase, a click or for general entertainment. Others are using natural language processing and other methods to derive “language” options to help you speak to an audience.
Other predictive analytics go much farther in certain industries to help identify customer segments based on value, propensity to buy, biases, effects of discounting, etc.
Machine learning is a synergistic exercise between humans and machines. Machine learning in practice requires human application of the scientific method and human communication skills. Successful organizations have analytic infrastructure, expertise and close collaboration between analytics and business subject-matter experts.
To fully address the topic, here are four fundamentals on the general methods most adapt:
- Supervised learning
In supervised learning, the machine learns by example. Humans provide examples of the desired inputs and outputs. The “machine” (aka the algorithm) uses this input to determine correlations and logic that it can use to predict the answer. This is like giving students an answer key and asking them to show their work. In supervised learning, sample Q&As are provided. The machine fills in how to get from A to B. Once it identifies a logical pattern, it can apply that pattern to solve similar problems.
Common techniques used:
- Baysean statistics
- Decision trees
- Neural networks
- Random forests
- Regression analysis
- Support vector machines
- Risk assessments
- Image/speech and text recognition
- Personalizing interactions
- Fraud Detection
- Semi-supervised learning
Semi-supervised learning is used to address similar problems as supervised learning. However, in semi-supervised learning the machine is given some data with the answer defined (aka “labeled”). It is predominately used in cases where there is too much data or too many subtle variations in the data to provide a comprehensive set of examples.
In this case, the given inputs and outputs provide the general pattern the machine can extrapolate and apply to the remaining data.
- Image recognition/classification
- Speech recognition
- Web content classifications
- Unsupervised learning
In unsupervised learning, the machine studies data to identify patterns. In this case, there is no answer key. The machine determines correlations and relationships by parsing the available data.
Unsupervised learning is modeled on how we humans naturally observe the world, drawing inferences and grouping like things based on unconstrained observation and intuition. As our experience grows (or for the machine, when the amount of data it is exposed to grows), our intuition and observations change and/or become more refined.
Common techniques used:
- Affinity analysis
- Clustering – K-means
- Nearest-neighbor mapping
- Self-organizing maps
- Singular value decomposition
- Market basket analysis
- Anomaly/intrusion detection
- Identifying “like” things
- Reinforcement learning
In reinforcement learning, the machine is given a set of allowed actions, rules and potential end states. In other words, the rules of the game are defined.
By applying the rules, exploring different actions and observing resulting reactions, the machine learns to exploit the rules to create a desired outcome. Thus, it determines what series of actions in which circumstances will lead to an optimal or optimized result.
Reinforcement learning is the equivalent of teaching someone to play a game. The rules and objectives are defined clearly. However, the outcome of any single game depends on the player’s judgment, how adjust an approach in response to the environment, skill and actions of the opponent.
Common techniques used:
- Artificial neural networks (ANN)
- Learning automata
- Markov decision process (MDP)
Machine learning in practice requires human application of the scientific method and human communication skills. The recipe is not as simple as “add data and stir.” Humans, above and beyond the data scientist programming the algorithm, are required to answer questions like these:
- What are we trying to predict?
- Are resulting correlations predictive? Causal? Are there inherent biases?
- Are results in line with expectations? Do we need to address exceptions?
- What is the predictive value, and can it be generalized?
- Can we apply the model and results in real life?
- What is the proper response?