Why this session exists
Session 6 was supervised learning: history came with answers attached. Today the answers are missing on purpose. Nobody labels customers "bargain hunter" or reviews "secretly furious" - yet those structures are in your data, waiting. This session covers the three unsupervised-and-adjacent tools every leader ends up buying: clustering, sentiment analysis, and customer lifetime value. Same rules as always: executive intuition, no math, no code. And the same honest boundary: recommendations, search and GenAI are out of scope by design - they belong to the future AI literacy course.
Unsupervised learning: finding structure without labels 5 min live
No right answers to learn from - so the algorithm does the only thing it can: group what is similar. Feed it customers described by numbers (orders per year, basket size, recency) and it hands back clusters. Whether those clusters mean anything is where you come in.
Livek-means in plain words, and the k question3 min▶
- The whole algorithm: drop k pins on the map of your customers, assign each customer to the nearest pin, move each pin to the center of its crowd, repeat until nothing moves. That is it - no magic, just tidy averaging in a loop.
- Choosing k is a judgment call. The elbow method gives a statistical hint (add clusters until the improvement flattens), but the real test is managerial: every segment must be nameable and actionable, or it is noise. If you cannot brief marketing on segment 4 in one sentence, you have too many segments.
- One-line alternative: hierarchical clustering builds a family tree of similarity (the dendrogram) - you cut the tree at the number of branches you want. Same ends, different route.
- RFM × clustering - the power combo. Session 5's recency, frequency and monetary scores make superb clustering inputs. Rules gave you the grid you designed; clustering finds the shapes you did NOT think to look for - the "big basket but fading" pocket no grid cell captured.
Tesco Clubcard, run with dunnhumby, is the founding legend of retail clustering. Loyalty-card baskets were clustered into shopper types - price-sensitive families, fine-food loyalists, convenience top-uppers - and the segments reshaped store ranging, promotions and mailings. The famous result: coupon redemption rates competitors could not touch, and a chairman who admitted dunnhumby knew Tesco's customers better than Tesco's own veterans did. The clusters did not sell anything - the actions built on them did.
Sentiment analysis: reading text at scale 4 min live
Your customers write to you constantly - reviews, tickets, tweets - and no human team can read it all. Basic sentiment analysis is the triage nurse: crude, fast, and good enough to tell you where it hurts.
LiveThe basic recipe, and where it honestly breaks3 min▶
- Step 1 - the lexicon: count positive and negative words from a dictionary. "Love, fast, perfect" versus "broken, refund, never". Crude, transparent, surprisingly useful.
- Step 2 - bag-of-words + a simple classifier: treat each review as a bag of word counts and let naive bayes learn which words make a review likely-negative ("refund" is a lot of evidence, "the" is none). Still no grammar, still fast, noticeably smarter.
- What it buys you: triage at scale. Thousands of reviews and tickets sorted by pain, themes ranked by frequency and negativity, releases compared before and after. Direction, not poetry.
- Where it breaks: sarcasm, negation and context. "Great, delayed again" reads as positive; "not bad at all" reads as negative. When the stakes need real understanding, that is the AI literacy course's turf - do not buy a hedge trimmer and expect a forester.
Airlines mine tweets and reviews to route service recovery in real time - a spike of angry mentions on one route pages the ops team before the morning report exists. Product teams run the same trick per release: sentiment by theme, this version versus last, tells them whether "battery" complaints actually fell after the fix. Nobody reads 10,000 tweets; everybody reads the ranked themes.
CLV: the metric that prices every other decision 6 min live
Customer lifetime value answers the only question marketing budgets really turn on: what is a customer worth over their whole relationship with you? Get CLV roughly right and acquisition spend, retention offers and even the churn threshold from Session 6 all price themselves.
LiveHistoric, predictive, and the honest middle3 min▶
- Historic CLV - what a cohort actually spent. True but backward-looking; your 2019 customers tell you about 2019.
- Predictive CLV - what customers WILL spend, modeled. Forward-looking but only as good as its assumptions.
- Cohort-based CLV - the honest middle. Track what each acquisition cohort spends over time (Session 5's curves) and project the curve forward gently. Most businesses should live here before buying anything fancier.
- Name-drop tier: BG/NBD + Gamma-Gamma (Fader & Hardie) - the classic probabilistic models your DS team would reach for to predict who is still "alive" as a customer and what they will spend. Names to recognize, not math to learn.
- CLV:CAC with the 3x rule of thumb - and its sibling, payback period. Below 1x you lose money on every customer you win; hovering near 5x or above, you may be growing too timidly.
Harrah's casino built its Total Rewards program on CLV - and the data said something nobody expected: the glamorous high-rollers were not the most valuable customers. Mid-tier locals - teachers and retirees visiting weekly - out-valued them over a lifetime, and the entire marketing budget was reshaped around them: offers, comps, even floor layout. It became the flagship case in Davenport's Competing on Analytics, and the lesson generalizes: until you compute CLV, your gut is spending the budget on the wrong customers.
Self-studyProduct analytics: the product team's funnel + cohort toolkit2 min read▶
Product teams run Sessions 5-7 thinking against the product itself. Three terms complete your vocabulary:
| Metric | The question it answers | The exec use |
|---|---|---|
| Activation | Did the new user reach the first moment of real value? | The retention lever that pays earliest - fix activation before buying ads |
| Feature adoption | Who actually uses the thing we shipped, and how often? | Kills the "we launched it so it worked" illusion in roadmap reviews |
| Path analysis | What routes do users actually take (versus the one we designed)? | Finds the detours and dead ends - the funnel's free-range cousin |
All three are funnels and cohorts wearing product clothes - you already know how to read them.
Run k-means with your own hands ★ 10 min · everyone clusters
Sixty customers plotted on two axes: orders per year and average basket. You are the algorithm now - assign, move, repeat.
Look at the raw cloud first. Can you SEE groups before any algorithm runs? (Usually: sort of. That fuzziness is why we bother.)
Click assign - every customer joins its nearest pin. Then click move - each pin walks to the center of its crowd.
Alternate assign, move, assign, move until nothing changes. Count your rounds - convergence usually arrives embarrassingly fast.
Now the human part: name the three segments that emerged. In this data they are roughly bargain-occasional, frequent-regular, and big-basket-rare. Write one action per segment.
Price a customer with the CLV calculator ★ 8 min · everyone prices
Every slider is one lever of the workhorse formula. Watch which levers move the verdict - the answers surprise most rooms.
Read the baseline customer: AOV $80, 4 orders a year, 30% margin, 3-year lifespan, CAC $95. Check the verdict - roughly 3x, healthy.
Challenge 1: marketing wants to double acquisition spend, pushing CAC to $150. Move the slider - is the engine still healthy?
Challenge 2: instead, invest the same energy in retention and stretch lifespan from 3 years to 4. Compare the two end states side by side.
Say the conclusion out loud: which lever would you fund this quarter, and what would you tell the team that owns the other one?
Segment reality check ★ 4 min · everyone audits
Write down the 3 customer segments your company ACTUALLY treats differently today - different offers, different service, different anything.
For each: is it defined by data (RFM scores, cluster membership) or by folklore ("enterprise means big logo")?
The uncomfortable test: if every segment gets the same email, you have zero real segments - and Session 5's RFM grid plus today's clustering is your gap list, in priority order.
Pick the one segment you would treat differently FIRST if the clusters existed, and note what the treatment would be. That is your ask for the data team.
This week ◐ 25 min total
- Finish the segment reality check from Apply-along 3 and share it with whoever owns marketing - the same-email test starts good arguments.
- Ask your team for CLV by acquisition cohort - even a rough historic version. If the answer is "we don't compute that", you have found this quarter's cheapest high-leverage project.
- Read the self-study card above: activation, feature adoption and path analysis - fifteen minutes that makes every product review legible.
- Optional deep end: Davenport, Competing on Analytics (the Harrah's chapters), and Fader's Customer Centricity for the CLV worldview in full.
Three questions before you go 🎯 ◐ 90 seconds
1 · What does k-means need from a human being before its clusters are worth anything?
Unsupervised means no labels needed - not no humans needed. Choosing k is a judgment call, and a cluster without a name and an action is noise. The algorithm finds similarity; meaning is your department.
2 · "Great, my flight is delayed again." Why does basic sentiment analysis misread this tweet?
Lexicon and bag-of-words methods count words without reading meaning, so sarcasm, negation and context defeat them. They remain great for triage at scale - and that failure line is where the AI literacy course begins.
3 · Your CLV:CAC ratio comes back at 0.8. What is the data telling you?
Below 1x, each acquired customer is a small loss you then scale with the marketing budget. The rule of thumb: ~3x is healthy; far above 5x may mean you are underinvesting in growth you could afford.
Concepts covered & their origins
The working 80% of each idea, credited to its canon - the deep end stays optional and the boundary stays honest.