Data Con LA 2022 - Real world consumer segmentation
0 likes29 views
Jaysen Gillespie, Head of Analytics and Data Science at RTB House 1. Shopkick has over 30M downloads, but the userbase is very heterogeneous. Anecdotal evidence indicated a wide variety of users for whom the app holds long-term appeal. 2. Marketing and other teams challenged Analytics to get beyond basic summary statistics and develop a holistic segmentation of the userbase. 3. Shopkick's data science team used SQL and python to gather data, clean data, and then perform a data-driven segmentation using a k-means algorithm. 4. Interpreting the results is more work -- and more fun -- than running the algo itself. We'll discuss how we transform from ""segment 1"", ""segment 2"", etc. to something that non-analytics users (Marketing, Operations, etc.) could actually benefit from. 5. So what? How did team across Shopkick change their approach given what Analytics had discovered.
![We transformed by z-score and/or by bucket number
Pythonistas: make sklearn.preprocessing.StandardScaler()your friend
• Age in years (18, 29, 44, 68, etc.)
• Gender, encoded as 1/0
• Race, encoded as 1/0 variables
• Household income in $ (20000, 50000,
etc)
• # of children (0, 1, 2, etc.)
• ZIP Code (90210, 10023, etc.)
• Years of eduation (12=HS, 16=Bachelors
Degree, etc.)
11
Raw Data
• Age [z-scored]
• Gender, encoded as 1/0 [OK]
• Race, encoded as 1/0 [OK]
• Household income in [z -scored]
• # of children/3 [0, 0.33, 0.66, 1.0]
• ZIP Code [dropped]
• Years of education [0, 1, 2, 3]](https://image.slidesharecdn.com/session8jaysengillespierealworldconsumersegmentation-220906175058-c6819db3/85/Data-Con-LA-2022-Real-world-consumer-segmentation-11-320.jpg)
Ad
More Related Content
Similar to Data Con LA 2022 - Real world consumer segmentation (20)
More from Data Con LA (20)
Ad
Recently uploaded (20)
Ad
Data Con LA 2022 - Real world consumer segmentation
- 1. User Segmentation (Real World) Jaysen Gillespie Head of Analytics &Data Science
- 2. 2 Commercial Notes • Over 1000 global employees • Based in Warsaw, Poland • Leader in using Deep Learning to solve online advertising challenges • A “DSP” for those of you in ad -tech/programmatic Technical Notes • Over 1B bid requests per day • Billions of ads served each year • Millions of clicks and site visits generated each year
- 3. How well do we really know our users? 3 • Marketing, Operations, and other teams have general ideas about the kind of people who use the app • Analytics had ideas about “best users” that different from views held by other teams
- 4. Why would we care about this (as analysts)? grow tribal knowledge; challenge existing thought; drive business impact • Product Strategy • What do we build? What do our key user segments need? • Marketing Strategy • Which types of users should we acquire? • What’s the best way to find them? • Finance • Which user profiles lead to better financial performance for the business? 4
- 5. We started with one -dimensional knowledge Let’s build some basic facts • Everyone knows at least one tool for this job • Excel • SQL • R / Python • Other languages (C++, etc.) 5
- 6. Univariate review of users is a good start And it can help give ideas for more complex analysis • Your most common user is not necessarily your most valuable user! 6 34% 23% 13% 11% 8% 11% 18-24 25-34 35-44 45-54 55-64 65+ New Users by Age, Jan 2022 3% 8% 13% 20% 25% 31% 18-24 25-34 35-44 45-54 55-64 65+ User Activity by Age, Jan 2022
- 7. But it’s hard to get the full picture, when you are looking at one color at a time • Let’s move from ETL & EDA to modeling 7
- 8. We wanted an “unsupervised” learning approach so that we don’t need the answers first Cluster Analysis is a classic choice • Linear and Logistic Regression • Decision Trees • Support Vector Machines • Random Forests 8 Supervised (Classification and Regression) UN-Supervised (Clustering) • K-means clustering • Hierarchical clustering • Association Rules (buy X -> buy Y) • Principal Component Analysis Deep Learning • Could be either • We use supervised to predict which ads will “work”
- 9. Quick reminder: K-Means clustering creates groups based on the centroid of group members Credit: Wikipedia 9 • Randomly pick N points as initial cluster centroids • Assign each point to closest proposed centroid (remember this!) • Find new centroids based on group assignment • Wash. Rinse. Repeat. (Until convergence)
- 10. K-means relies on a distance formula, so make sure to normalize data in some way Which features (raw data fields) will drive the analysis? • Age in years (18, 29, 44, 68, etc.) • Gender, encoded as 1/0 • Race, encoded as 1/0 • Household income in $ (50000, etc.) • # of children (0, 1, 2, etc.) • ZIP Code (90210, 10023, etc.) • Years of eduation (16=Bachelors, etc.) 10 Raw Data
- 11. We transformed by z-score and/or by bucket number Pythonistas: make sklearn.preprocessing.StandardScaler()your friend • Age in years (18, 29, 44, 68, etc.) • Gender, encoded as 1/0 • Race, encoded as 1/0 variables • Household income in $ (20000, 50000, etc) • # of children (0, 1, 2, etc.) • ZIP Code (90210, 10023, etc.) • Years of eduation (12=HS, 16=Bachelors Degree, etc.) 11 Raw Data • Age [z-scored] • Gender, encoded as 1/0 [OK] • Race, encoded as 1/0 [OK] • Household income in [z -scored] • # of children/3 [0, 0.33, 0.66, 1.0] • ZIP Code [dropped] • Years of education [0, 1, 2, 3]
- 12. If “unsupervised” doesn’t require a training set, why don’t we use it for everything How will I know… • Qualitiative evaluation measures 12 Supervised (Classification and Regression) UN-Supervised (Clustering) • “Looks good to me!”
- 13. Limited design decisions make k -means clustering user friendly Just tell me how many clusters you want please 13 Less is more • How much gain in consistency by adding another cluster? • Ask the business: how many segments do you think (want to) exist? Elbow Method • Where the curve flattens out 0 500 1,000 1,500 2,000 2,500 3,000 3,500 4,000 4,500 5,000 1 2 3 4 5 6 Sum Squared Distance vs # Clusters Case Study 1 Case Study 2
- 14. Excellent – so we have a super awesome output with 4 -clusters Is this where we break out the $.drinks.favorite? 14 How do we know if this is of any use?
- 15. The real fun is just about to begin… 15 Full-stack analysts shine in the interpretation phase Data acquisition, cleaning and other preparation EDA Run Mode l Analytical Interpretation Business Interpretation Business Execution (Not always Analytics’ Job)
- 16. We can now find all sorts of averages or other summary data points “by cluster number” Let’s learn how the K -means algo made decisions 16 • Which input features differ most across segments? • Are there features that don’t seem to matter? • Did we validate that our choice of encoding didn’t fail us? • Do we sense a primary feature or two that drove segmentation? • Should we re-run with N-1 or N+1 segments? Analytical Interpretation
- 17. Marketing and business leadership teams should love the insights from cluster analysis ”Segment 2’s average income is 1.6 standard deviations below the mean!” 17 • Can we better align our verbiage with our audience? • Let’s re-skin what we did in terms of how our audience thinks and speaks
- 18. Instead of specific facts, we found that a holistic view of each segment earned a warmer reception Listing the dominant value of key attributes appealed to business users Now we are starting to get the “look and feel” of each segment (Actual table had 8 -10 key dimensions) 18 Segment Age Income Education Marital Status 1 Boomers (55%) Some GenX/Mill. Top 50% (75%) Some college Married (77%) 2 GenX (35%) Mill (32%) Top 50% (92%) College grads Married (63%) 3 GenX/Mill. (65%) Bottom 50% (88%) Little college Unmarried (59%) 4 Mill. (38%) GenZ (22%) Bottom 50% (68%) Some college Unmarried (82%)
- 19. Even a rudimentary attempt at working in our audience’s discipline can work magic Who came up with these names? The Data Science team? 19 • First-draft names often stick • Analytics can own internal collateral explaining the work done • Projecting n-dimensions into a 2D graph is a crowd pleaser AGE INCOME Senior Stables Mid-Age Thrivers Mid-Age Laggers Young & Struggling
- 20. Joining segmentation to average KPIs (in -app actions, purchase behavior, etc.) rounds out the profiles 20 • Analytics created internal collateral with segments and usage/purchase data • Reasonable stopping point for hand -off to Marketing • Python code makes updating segmentation on a regular basis (1x/year or 4x/year) easy to do
- 21. At this point, we can hand off implementation to other teams (or ourselves) 21 • Marketing to create personas; change targeting for new customer acquisition; brief ad agency • Finance (or Analytics) to understand the total lifetime value (LTV) for each user segment • Product to connect feedback with each user segment • Analytics to develop reporting based on user segments
- 22. Jaysen Gillespie Head of Analytics and Data Science [email protected] THANK YOU