Determining the “what’s best in the next best action” or relevance in the moment with ML/AI in combination with heuristics and rules.
Recommenders are tied to product and offering sku-level recommendation based on user probabilities (i.e. recommend this type of television) of transactions what to recommend user. Recommenders input data to Next Best Action which adds contextual insights (i.e. can this user afford this television).
Don’t do AI/ML for AI sake. Think about what are the outcomes implementing algorithms? revenue growth, revenue growth with constraints / risk assessment. Think for customer retention: level 1 is Outcomes – how much is it worth to keep them; level-2 Codify Rules what would do you offer that can be presented for customers to say yes or no to?; level-3 Make Predictions how to talk to the customers and what to say. Generate a script of actions based on next best actions. From outcomes, to rules, to predictions and data to make predictions that would keep a customer.
In the industry the pendulum swing from all-Rules based to all-ML to now a combination of ML below heuristics, which provides unparalleled performance. It’s not a technical integration but a process of outcomes, (eligibility/risk/etc) fit to rules, aligned to a categorical actions library, for each action automatically calculating the right one in the moment calling the right model (risk model, propensity model, eligibility model, self learning model) and updating the algorithms in real time based on outcome.
Can you have an algorithm behind every action and treatment? Can you execute it in the moment so you aren’t retrieving a probability from a db? Can you recalculate on the fly? Complimenting data science models with machine learning algorithms and heuristics for optimal implementation.
Compliment above with responsible AI is it: fair, robust, transparent, empathetic. How do you enforce in implementation?
Combination of heuristics and AI, is the final recommendation fair after passing the data through the entire process?
You don’t want predictions in a database to evaluate and learn. How do you do at scale and fast enough for customer engagement. in a fraction of second recalculate the next best action ready?
Do you need real time? 1) no for static profile, do you have access to all the data real time (e.g. birthday) . 2) for in the moment profile, what you browsing, clicking, not clicking, etc. – when you click No, informs 1000 predictions and updates. A score from a database versus real time is like talking a minute old is like talking with a minute delay in conversation.
Somethings (features) you can precalculate, this would fit a feature store for making raw data more predictive that can be calculated real-time or stored in a db.
Ai is two way informative, predicting for users and informing for organizations what data is more effective and what isn’t.
Supervised labels that comes from historic encounters versus hand labeling.
Deep learning gets a lot of attention, but many bread and butter uses are not. DL is opaque, ML is less black box.
ModelOps, type of algorithm, data, features, use case labelled, compliance assessment.
david marcus 