Upgrade to Pro — share decks privately, control downloads, hide ads and more …

A pleb on the web tries to understand psychographic targeting

A pleb on the web tries to understand psychographic targeting

Siva Swaminathan

October 17, 2020
Tweet

More Decks by Siva Swaminathan

Other Decks in Education

Transcript

  1. Psychographic targeting
    Siva Swaminathan

    View Slide

  2. View Slide

  3. Background

    View Slide

  4. How and why did this talk come about?

    View Slide

  5. Diffusion of innovation
    Essentially, an analysis of sociology (individuals and social links).
    Very broadly applicable, for the spread of all kinds of ideas
    (fundamentally what makes us human?)

    View Slide

  6. Legibilizing social interactions
    Electronic social networks which make up fabric of the web, are
    excellent for providing a legible map of sociography,
    allowing for surgical dexterity in targeted messaging.

    View Slide

  7. What spreads effectively need not be
    innovation or even truth
    Reproduction, Variability, Heritability, Differential selection
    (Just needs a good fit into incidental constellation of environmental factors)
    (see arguments by Daniel Dennett, Bret Weinstein)

    View Slide

  8. Methodological individualism
    (factorizable structure of mathematical models)
    Person is described by their interests
    Person is described by their psychography
    Person is described by their sociography

    View Slide

  9. The marketing game

    View Slide

  10. "Half our marketing budget is wasted …
    but we don't know which half!"

    View Slide

  11. Tracing back to Edward Bernays (1928, 'Propaganda'),
    a close relationship between marketing and psychology.
    (Adam Curtis, 'The century of the self')

    View Slide

  12. Innovation diffusion models
    Segment individuals into each stage of adoption
    Innovators, Early Adopters, Early majority, Late majority, Laggards
    Mapping between identifiable/targetable traits & segmentation

    View Slide

  13. Market segmentation & user profiles
    Typically industry/context-specific analysis, sponsored by
    market leaders, or trade associations, or consulting agencies.
    eg:
    Only broad targeting (typically mass media)
    "Subarus for lesbians"

    View Slide

  14. Structured models arbitrary tidbits
    Sub/urban lifestyle choices?
    Preferred car models?
    Which OS & browser?
    Thermostat prefetences?
    etc.
    ≫≫

    View Slide

  15. Personalization: automated, and at scale
    "Direct marketing" (Lester Wunderman, 1960s)
    Convivial Industrial

    View Slide

  16. The technical aspects of
    inferring profiles

    View Slide

  17. Collaborative filtering
    N users, B behaviors Humongous sparse matrix
    Eg: Product recommendations, Ad effectiveness, Topic modeling, etc.
    ⟹ N × B

    View Slide

  18. Low-rank approximations
    Simple algorithms, eg: SVD, NNMF, LDA
    [N × B] ≈ [N × r] ⋅ [r × B]

    View Slide

  19. "Big (wide) data, with simple algorithms
    beats complicated models"
    Humans are not unique snowflakes,
    but well approximated into a handful of types

    View Slide

  20. From how we behave to who we are
    Could we generalize across domains?
    What might be a more intrinsic characterization of people?

    View Slide

  21. Big-5 (OCEAN) model
    Openness, Conscientousness, Extraversion, Agreeableness, Neuroticism
    (rank-5 approximation, SOTA psychology reasearch)

    View Slide

  22. Vacuuming up digital exhaust
    responses in silly quizzes
    links you click on
    time spent on various pages
    location tracking (bluetooth, GPS)
    Facial recognition and body trajectory
    Eyeball tracking
    etc. behaviors

    View Slide

  23. Experiment summary

    View Slide

  24. Models easily beat closest confindants!

    View Slide

  25. More "factors"
    Demographics [who] (eg: age, gender, location, etc)
    Psychographics [why]
    Personality traits: (eg: OCEAN, thriftiness, risk aversion, etc.)
    Psychological states: (eg: mood, emotions, stress level, etc.)
    less temporally stable, but more effective?
    Other factors: Behavioral data [how] & Transactional data [what]

    View Slide

  26. Using inferred profiles

    View Slide

  27. Targeting by psychographics (1)

    View Slide

  28. Targeting by psychographics (2)

    View Slide

  29. "A/B testing"
    Counterfactuals, not just correlations

    View Slide

  30. Other modalities of influence
    Which voice would have the desired influence?
    What wording would have the desired influence?
    Target messages for virality?

    View Slide

  31. GPT3, and automated (personalized)
    content generation
    Eg: Adolos Substack GPT-3 autogenerated

    View Slide

  32. Same methods for political messaging
    Psychometric Profiling: Persuasion by Personality in Elections (link)

    View Slide

  33. Exploring consequences

    View Slide

  34. Possible beneficial uses of
    personalization
    Mental health diagnosis and interventions
    Better matching (employment, dating, entertainment, etc)
    "Nudge" towards healthier outcomes
    Custom insurance premiums and interest rates

    View Slide

  35. Uncomfortable examples
    Target pregnancy
    Uber "rides of glory"
    Large scale emotion manipulation study on Facebook

    View Slide

  36. Fundamentally breaks common
    knowledge
    (eg: Say's law of one price)
    Discrimination (eg: "good" and "bad" price discrimination)

    View Slide

  37. Psyops: Disinformation & Nonlinear war
    Mess with sense-making apparatus
    Spread FUD (war, elections, political polarization, etc)
    Correlate power usage to bring down the power grid!
    "Global village" (McLuhan, 1950s)
    (BLM protests in BLR and NZ, ignoring COVID lockdown!)
    Gerasimov doctrine

    View Slide

  38. Assault on individual agency?
    At what point should we regulate
    targeted messaging like the Imperius Curse?
    (ambient nudges better -vs- direct control)
    Why is advertizing different from rape?
    (Memetics -vs- Genetics)

    View Slide

  39. Thank you

    View Slide

  40. // reveal.js plugins

    View Slide