AI, Society, and Human Behavior

Research Methods in Context

Carina I Hausladen

Topics

Four cutting-edge topics at the frontier of computation social science:

1. Measuring Bias in AI
2. Social Choice for LLM Alignment
3. Clustering Multidimensional Time Series — Modeling Human Behavior
4. Modeling Social Dilemmas through Reinforcement Learning

• Research Skills

  • Design your own research question

  • Replicate, extend, or reinterpret topics we discuss

• Applied Methods

  • Analyze real data using computational tools

  • Code in teams to explore your question

  • Build a GitHub repository for open, replicable research

• Communication & Impact

  • Write a short research-style paper

  • Present your insights to others

  • Discussion & active participation

Skills

Activities & Assessment

• 10:00 – 11:30:
  • Carina introducing topics and methods
• 11:30 – 12:30:
  • Lunch (together)
• 12:30 – 14:00:
  • 30 minutes discussants
  • 30 min explaining of core concepts to each other
  • 30 min coding

1. Reading Response

1. Reading Response

• Your response should answer the following:
  1. What is the core idea or contribution?
  2. What questions would you like to ask in class?
  3. What parts of the paper are interesting to you and why?
  4. How would you replicate or extend the paper?
      
• These responses are not graded.
• Responses are contributed via Overleaf.

• Serve as a discussant for one paper (only once!)

• Probably in pairs of two

• Deliver a brief (~7–10 min) presentation, focusing on:

  • Summarize the core idea of the paper

  • Does it introduce an interesting dataset we could utilize?

  • Is there an analysis worth replicating? How could this work be extended*?

    • *who did recently cite this paper?

  • Encourage discussion with your classmates

• Graded (20%)

• Deadline: Thursdays, 10 PM

2. Discussant Role

3. Group Project

• Group Project, delivered as
  • presentation (30%)
  • paper (50%)
• The paper should have around 8 pages and 4,000-8,000 words, and should be structured like a paper.
  • You should include a 'contributions' section outlining what group member did what.
• You should link a Github repo with the code you developed.

Code Clinic

• Checking analysis choices; assessing whether additional statistical tests are needed.
• Do the figures make the point?
• Does your GitHub repository support replication?

Writing Clinic

• Good writing
  • specifically focusing on abstract, figure captions, title
• Good presentations: what makes a talk effective

Presentation and Paper

• Writing is thinking
  • Ideally, the core of your paper is in a good shape before the presentation
  • When do you want to hand in your final paper? 
• Your presentation should also include a short introduction to your GitHub repository

1. Ethics of AI

January 9

Plan for today

First Session

• 40'
  • 15' Introduction
  • 25' Defining Bias
    
    
    —5' break—
     
• 45'
  • Bias Metrics  (JN Tutorials)
    • 15' WEAT 
    • 15' Probability Based
    • 15' Generated Text

Second Session

• 40'
  • Bailey 2022
  • Bai 2025
    
    
    —5' break—
     
• 40' 
  • Khan 2025
  • Hausladen 2025

A Career Track

📚 Academia

• Bias & fairness is a core research area

• Survey papers regularly reach thousands of citations
  (e.g. Mehrabi et al. 2019 >8,000 citations)

• Dedicated top-tier venue: ACM Conference on Fairness, Accountability, and Transparency (FAccT)

• Strong presence at NeurIPS, ICML, ICLR, ACL, EMNLP

• Interdisciplinary work = high visibility + funding relevance

🏭 Industry

• Major companies run dedicated fairness teams

  • Apple, Google, Meta, Microsoft, IBM, ...

• Common job titles:

  • Responsible AI Scientist

  • Fairness / Bias Engineer

  • Algorithmic Auditor

  • Trustworthy ML Researcher

• Regulation (EU AI Act, audits, compliance) → growing demand

Are Emily and Greg
More Employable than
Lakisha and Jamal?

 Bertrand & Mullainathan (2003)

Leon Yin et al. (2024)

https://www.linkedin.com/pulse/unveiling-hidden-bias-lessons-from-derek-mobley-v-workday-bareket-v7off/

(2024)

https://business.linkedin.com/talent-solutions/hiring-assistant?ss=1

Why do you care about fairness and bias? 

Defining Bias for LLMs

1. Fairness Definitions
2. Social Biases
3. Where Bias Enters the LLM Lifecycle
4. Biases in NLP Tasks
5. Fairness Desiderata

* much of following slide content is based on  "Bias and Fairness in Large Language Models: A Survey"

1. Fairness Definitions

2. Social Biases

3. Where Bias Enters the LLM Lifecycle

4. Biases in NLP Tasks

5. Fairness Desiderata

5' break

Bias Metrics

1. Embedding Based
2. Probability Based
3. Generated Text

1. Embedding Based Metrics

man

work
salary

man

home
family

women

work
salary

women

home
family

2. Probability Based Metrics I

Log Probability Bias Score
(LPBS)

$$LPBS = \log\left(\frac{P(\text{she}\mid context)}{P(\text{she}\mid prior)}\right) - \log\left(\frac{P(\text{he}\mid context)}{P(\text{he}\mid prior)}\right)$$

2. Probability Based Metrics II

1. mask one word at a time
2. calculate e.g. P('she' | context)
3. calculate log(P)
4. sum all log probabilities

3. Generated Text Based Metrics

It's your turn!

1_metrics_weat.ipynb

 

github.com/carinahausladen/konstanz-ai-behavior-2026

Papers

• cosine similarity between static word embeddings (fasttext / glove)

  • embedding-based

• WEAT

  • embedding-based

• LLM Word Association Test (LLM-WAT)

  • generated text-based → distribution

• LLM Relative Decision Test (LLM-RDT)

  • generated text-based → distribution

• WEAT

  • embedding-based

• coreference confidence

  • probability-based

• coreference confidence disparity

  • probability-based

• accuracy

  • probability-based

January 16

Plan for today

First Session

• 30' Datasets
  • 10' Introduction
  • 25' JN
    
    —5' break—
    
     
• 50' Fairness and Causality
  • 15' Simpson Paradox
  • 35' Social Perception of Faces

Second Session

• 40'
  • Conizer 2024
  • Qui 2024
    
    
    —5' break—
     
• 20' 
  • Kirk 2024
• 20'
  • research ideas for the PRISM dataset

Datasets for Bias Evaluation

Counterfactual Inputs

• Masked Tokens (WinoBias)
• Unmasked Sentences (CrowS-Pairs)

Prompts

• Sentence Completions (BOLD)
• Question-Answering (BBQ)

WinoBias

Zhao et al. (2018)

CrowS-Pairs

Nangia et al. (2020)

Bias in Open-Ended Language Generation Dataset (BOLD)

Dhamala et al. (2021)

Bias in Open-Ended Language Generation Dataset (BOLD)

Bias Benchmark for QA (BBQ)

Parrish et al. (2022)

Bias Benchmark for QA (BBQ)

It's your turn!

2_metrics_maskedtoken

3_metrics_pll

4_metrics_generatedtext

5_datasets

github.com/carinahausladen/konstanz-ai-behavior-2026

Fairness and Causality

Limitations of Observational Data

In the 1970s, UC Berkeley was sued for alleged gender bias in graduate admissions.

Bickel, Hammel & O'Connell (1975)

What is Simpson’s Paradox?

Simpson’s Paradox happens when a trend seen in aggregated data reverses or disappears when the data is broken into groups.

Simpson’s Paradox can make AI systems look biased or fair depending on how you slice the data.

What does it mean for AI Fairness Audits?

What is the cause of this discrepancy?

• Alternative hypotheses:
  • Less competitive departments were less welcoming to women?
  • Some departments had a track-record of unfairly treating women and this was known to
    applicants?
  • Some departments advertised programs in a way that discouraged women from applying? ...
• For more detailed analysis of this example, see Pearl and Mackenzie, The Book of Why: The New Science of Cause and Effect 2018

Counterfactual Fairness

• Discovering the cause of the discrepancy requires a more advanced inference framework
• Counterfactual inference:
  • estimate probability an individual would have been admitted if they were a man instead of a woman
  • Bertrand & Mullainathan (2003) / correspondence studies 

Red Car Scenario

• Insurance companies charge more for red cars

• Causal model:

  • aggressive drivers (unobserved) u

  • like red cars x

  • tend have more accidents y

• What if, people of some races c prefer red cars?

• Fairness test: If we changed the person’s race in the model but kept their underlying aggressiveness the same, the prediction should not change.

Kusner et al. (2018)

Carina I. Hausladen, Manuel Knott, Colin F. Camerer, Pietro Perona

Social perception of faces in a vision-language model

2. Social Choice and LLM Alignment 

January 16

Papers

Ideas for the PRISM dataset

How does the choice of aggregation rule reshape who benefits from alignment?

• Test multiple aggregation rules:
  Utilitarian (mean), Thiele-style proportional scoring, Rawlsian (floor-maximizing), inequality-adjusted welfare, etc.

• For each rule, select the top-$K$ models.

• Compute user welfare under access to the selected models (e.g., random-choice lower bound; best-choice upper bound).

• Compare welfare across socio-demographic groups:
  Gender, ethnicity, age

• Report outcomes, e.g.

  • Mean welfare

  • Bottom-decile welfare (10th percentile / bottom 10%)

  • Welfare gaps between groups (e.g., max–min group mean; or pairwise differences)

January 23

Plan for today

First Session

• 40' Social Choice Recap
  • 10' Why do we care about pluralism?
  • 10' Social Choice Theory in a nutshell
  • 20' JN: Different aggregation methods with PRISM dataset
    
    —5' break—
     
• 40'
  • 10' research idea: X community notes
  • 30' preference elicitation methods & legitimacy

Second Session

• 20'
  • Clara presenting Köster et al.
• 70'
  • Guest Lecture: Joshua Yang
  • Proportional-based fairness (PB)
  • LLM Voting Paper in depth

Why do we care about Pluralistic Alignment?

Common mindset in AI, economics, policy:

• Optimization

• Quantification

• Preference satisfaction

→ Preference-based utilitarianism

Problems

• Can justify injustice to minorities

• Preferences can be racist, sexist, distorted

• Ignores reflection on what preferences should be

• Reduces ethics to prediction + optimization

→ Ethics ≠ just aggregating preferences

Tasioulas (2022)

The Risk of Narrow Ethics in AI

Three core principles:

1. Pluralism

2. Procedures

3. Participation

Humanistic Ethics

• Values are often incommensurable​

  • No single master value

• Some choices have multiple reasonable options
  • career choices, sentencing decisions
• Implication:

  • Algorithms cannot always optimize correctly

  • Human judgment remains essential (reciprocity)

• "Noise" is not always bad — it can reflect legitimate diversity

1. Pluralism

We care about:

• How decisions are made (throughput legitimacy)

• Not only what result is produced

Examples:

• Cancer diagnosis → outcome priority

• Criminal sentencing → process priority

  • AI judges threaten reciprocity and dignity

2. Procedures

• Well-being is not a passive pleasure
  • It requires active engagement.
  • Participation in Democracy = civic dignity

• AI should strengthen, not weaken, democratic participation

3. Participation

Social Choice Theory in a Nutshell

How do we aggregate individual preferences, judgments, or welfare into coherent collective decisions?

The Challenge

Rational Individual Preferences

• Voter 1: x > y > z
• Voter 2: y > z > x
• Voter 3: z > x > y

Pairwise majorities:

• x beats y,
• y beats z,
• z beats x 

Condorcet's Paradox (1785)

Cycle: x>y>z>x

This violates transitivity, a fundamental principle of rational preferences.

• If you try to turn individual preference rankings into
• one “society-wide” ranking in a fair way,
• you can’t satisfy a small set of very reasonable fairness conditions at the same time—
• unless you accept a dictator.

Arrow's Impossibility Theorem

1. The rule must work for any possible ranking (Universality).
2. If everyone prefers A>B, society must rank A>B (Pareto).
3. How society ranks A and B should only depend on citizens' choices concerning A vs B, and not depend on irrelevant third options C (IIA).
4. No single person should always determine social ranking (No Dictator).

Four reasonable conditions (Axioms)

• A, B, C
• Each person has a ranking (A>B>C)
• a social welfare function outputs a single collective ranking

​Why it matters for AI  preference aggregation

1. Condorcet (Pairwise Majority)

• Compare every pair of models head-to-head

• Winner: beats all others in pairwise votes

2. Borda Count

• Rank models, assign points (1st place = k points, 2nd = k-1, etc.)

• Winner: highest total points

3. Plurality (Highest Average)

• Model with highest average score wins

4. Approval Voting

• Users "approve" models scoring above threshold (e.g., >70)

• Winner: most approvals

5. Utilitarian (Total Welfare)

• Maximize sum of all scores

• Winner: highest total satisfaction

5 Aggregation Methods We'll Explore

It's your turn!

github.com/carinahausladen/konstanz-ai-behavior-2026

Research Idea:
X Community Notes

https://x.com/i/communitynotes/download-data

• Users write notes on posts
• Crowd rates notes as helpful / not helpful

How it works

• Platforms shift from expert fact-checking to crowdsourcing
• Community Notes aims at scalable, democratic moderation
• Core assumption: diversity → consensus → better moderation

Motivation

• 69% of posts show classification disagreement
• Only 11.5% of notes reach rating consensus
• → Dissensus is the dominant outcome

Consensus or Dissensus?

• Is dissensus caused only by polarization — or by the aggregation rule itself?

New RQ

• Recompute outcomes using: Condorcet, Borda, Approval, Utilitarian
• Compare with X’s algorithm on:
  • Consensus rate
  • Elite influence
  • Do welfare analysis similar to Kirk et al. (2025)
    • define clusters (we do not have sociodemographics, so we need to come up with other variables to cluster)

Possible Extension

Preference Elicitation

How Voting Rules Impact Legitimacy

Carina I. Hausladen, Regula Hänggli-Fricker, Dirk Helbing, Renato Kunz, Junling Wang, Evangelos Pournaras

Guest Lecture

* slides are on our GitHub in folder 01_23

3. RL in Social Dilemmas

January 23

January 30

Plan for today

First Session

• 45'
  • ​5' Recap on Social Dilemmas
  • 5' Play a PGG 
  • 5' Learning in Social Dilemmas
  • 10' RL and QL
  • 20' JN
    
    —5' break—
     
• 40'
  • Paper: IRL for Social Dilemmas

Second Session

• 75'
  • Guest Lecture: 
    • Damian Dailisan
    • Javier Argota Sánchez–Vaquerizo

Social Dilemmas

What’s one everyday situation that feels like the exact same game as another totally different situation?

Christoph Kuzmics

What is a Social Dilemma?

Most real problems reduce to a few abstract games:

• Prisoner’s Dilemma
• Stag Hunt
• Public Goods Game (PGG)

Core games that capture social dilemmas

• Two suspects are arrested for a joint crime and interrogated separately.
• Each can cooperate with the other (stay silent) or
  defect (confess against the other).
• The police offer a deal: betray your partner for a lighter sentence, but if both betray, both get punished harder.
• Payoffs are years in prison 

Prisoner's Dilemma

• A pair of players has the choice of hunting a stag or a hare.
• If a player chooses to hunt a stag, he can only succeed with the cooperation of the other participant.
• Players can get a hare by themselves, but a hare is worth less than a stag.

Stag Hunt

Data from our in-class experiment

• Cooperation collapsed over time - Both games started with moderate contributions (~€150-200 total) but declined as players learned they could free-ride on others' generosity
• Free-riding pays off - Player 5 contributed only €54 total but earned €848 in payoffs, while Player 6 contributed €1,000 and lost €115 - rational self-interest beats altruism
• End-game effect is real - Round 10 showed dramatic drops in contributions (Game 2 crashed from €200+ to €40) when players knew there was no future punishment
• Trust erodes when exploited - Players who initially cooperated (Players 1, 3, 4) reduced contributions after seeing Player 6's consistent €100 wasn't reciprocated by others

The Public Goods Game

Learning

Why Learning Matters

Best-Response Learning

Agents form beliefs about others’ behavior.

Reinforcement Learning

Agents repeat actions that paid off.

Evolutionary Learning

Strategies spread based on success.

Three Main Learning Approaches

                          Evolutionary         Reinforcement            Best Response

🧬 Evolutionary Learning in the PGG

• Some people are “cooperators” 

• Some are “free-riders” 

• Groups with more cooperators do better overall

• Those strategies spread over time

• Key idea: successful behaviors reproduce

Result:

If structure exists (small groups, reputation, punishment types):

➡️ cooperative strategies survive and dominate

If not:

➡️ free-riders take over

Key idea: successful behaviors reproduce

🎯 Reinforcement Learning in the PGG

Think habit-building.

What an agent does:

• Try contributing $5 one round

• See payoff

• Try contributing $0 another round

• Compare outcomes

If cooperation leads to higher long-run rewards:

➡️ agent gradually contributes more

If free-riding pays more:

➡️ agent learns to free-ride

No thinking about others — just reward history

🧠 Best-Response Learning in the PGG

• Observes others’ past contributions

• Estimates what others will contribute next round

• Chooses best response

Examples:

• “Others usually contribute high → I should contribute to sustain it”

• “Others are free-riding → I should free-ride too”

Uses beliefs about others’ behavior

Reinforcement Learning

• No strategy reasoning.
• No understanding others.
• Just: “Did this pay off last time?”

1. Try actions
2. Observe payoffs
3. Repeat what works |
   Avoid what doesn’t

Agents

Reinforcement learning =
learning by experience

If over many rounds:

• Mutual cooperation gives highest long-term payoff
• Free-riding triggers others to stop contributing
   

Then RL agents learn: contributing more is better in the long run

Cooperation emerges

Contribute $5 → gets $12

Contribute $0 → gets $14

Contribute $10 → gets $16
 

The agent slowly updates:

• Actions with higher rewards become more likely
• Low rewards become less likely

An agent experiments

• Q(contribute 0) = expected reward of free-riding
• Q(contribute 5) = expected reward of moderate cooperation
• Q(contribute 10) = expected reward of full cooperation

Example

Basic RL: Try actions → see reward → repeat what works

Q-learning just adds memory:
It keeps a score (Q-value) for each action.

From Basic RL to Q-Learning

Intuition

• Q(0) = 11
• Q(5) = 13
• Q(10) = 16
   

The agent will increasingly choose to contribute 10 because it learned it pays most over time.

So if

Q(action) ≈ average long-term payoff
if
I choose that action

Q-Learning

Q-Learning

\( Q_{new}(s,a) = (1- \alpha) Q_{old}(s,a) + \alpha \left( r + \gamma \max Q_{old}(s', a') - Q_{old}(s,a) \right) \)

Q-Learning

\( Q_{new}(s,a) = (1- \alpha) Q_{old}(s,a) + \alpha \left( r + \gamma \max Q_{old}(s', a') - Q_{old}(s,a) \right) \)

reward

discount

learning rate

What is the best payoff I think I can get next round if I act optimally?

Q-Learning

\( Q_{new}(s,a) = (1- \alpha) Q_{old}(s,a) + \alpha \left( r + \gamma \max Q_{old}(s', a') - Q_{old}(s,a) \right) \)

reward

discount

learning rate

subtract what I used to think this action was worth

It's your turn!

github.com/carinahausladen/konstanz-ai-behavior-2026

4. Patterns in Multidimensional Timeseries

January 30

Identifying Latent Intentions

via

Inverse Reinforcement Learning

in

Repeated Public Good Games

Carina I Hausladen, Marcel H Schubert, Christoph Engel

MAX PLANCK INSTITUTE

FOR RESEARCH ON COLLECTIVE GOODS

Standard Cooperation Games

• Dal Bó & Fréchette (2018)
  • 103 sessions, 1,734 subjects, 116,644 observations
• Bigoni et al. (2024)
  • 27 sessions, 598 subjects, 60,334 observations
• Embrey et al. (2018)
  • 27 sessions, 552 subjects, 65,720 observations
• Mazur et al. (2025)

Research Idea

Guest Lecture

Slides can be found here: https://polybox.ethz.ch/index.php/s/4xa6gEAqt93Fx57

Course Evaluation

https://evasys.uni-konstanz.de/evasys/online.php?pswd=HAEV8

Goals

• Abstract submission deadline: March 3, 2026
• 2 page abstract 

How I approach initial drafts

• Tip 1: Develop Code & Writing Together
  • Don’t separate “analysis first, writing later”
  • After each analysis step, immediately write your insights into your paper
• Tip 2: Capture Ideas the Moment They Appear
  • Paste new insights or references directly into your draft
  • Place them roughly where they belong — refine later
  • This prevents losing strong ideas you’ll forget otherwise

• Tip 3: Keep Code Structured & Results Visible
  • Organize code into functions
  • Prefer scripts over notebooks for projects (notebooks encourage exploratory structure that often does not match your final workflow, leading to duplicated effort)
  • Figures
    • Insert figures into the draft immediately
    • Write captions right away — don’t postpone
• Tip 4: Talk Through Your Results Early & Often
  • Explain your work to others — it builds clarity and confidence
  • Turn your project into slides to refine your narrative
    • Slides often reveal better explanations than text alone

How I approach refining drafts

•  Tip 5: Let Your Narrative Evolve

  • Revisit and rewrite your introduction often

  • Feedback from peers and presentations reveals what actually resonates

  • Adjust the storyline and “red thread” of the introduction often/daily

• Tip 6: Lower the Cost of Cutting

  • Keep a `trash.tex` (or notes file) for removed paragraphs

  • Move unused text there instead of deleting it

  • This makes restructuring psychologically easier

  • I almost never use this material again

• Tip 7: Rewrite from Different Styles

  • After drafting your scientific version, ask an LLM to rewrite it in Economist/any newspaper you like- style

  • Notice how ideas become clearer and more tangible

  • Borrow stylistic tools that improve flow and accessibility

• Tip 8: Use Grammarly

  • Use Grammarly (even free is helpful — works in Overleaf)

  • Catch small errors early instead of fixing everything at the end

  • Clean writing improves thinking

• Tip 9: Work on Your Paper Every Day

  • Set a daily writing goal (even small)

  • Avoid skipping days — momentum matters

  • Keeping ideas in short-term memory makes progress faster and easier

• Tip 10: Listen to Your Paper

  • Use text-to-speech tools (e.g., Speechify) to hear your draft

  • Hearing reveals awkward phrasing and logic gaps

  • It gives a completely new perspective on clarity

• Tip 11: Use the Writing Center Strategically

  • Go at least once before submission

  • Content may change little — but clarity and confidence improve a lot

  • Writing becomes more enjoyable and effective

• Tip 12: Keep Your Draft “Submission-Ready” at All Times

  • No spelling errors

  • Clean citations

  • Figures always with captions

  • Avoid messy drafts --> **Psychological effect:**

    • A clean document feels professional → boosts motivation → improves quality.

Appendix

• Dutch Childcare Benefits Scandal

• What happened

  • ~26,000–35,000 families wrongly accused of childcare-benefit fraud

  • Parents forced to repay tens of thousands of euros

    • Many families fell into severe poverty;

    • children were removed from some families as a downstream consequence

• Where the bias came from

  • Fraud risk-scoring system used nationality/dual nationality as risk indicators

  • Zero-tolerance rule:

    • any suspected irregularity ⇒ 100% benefit clawback

    • Minor administrative errors treated as intentional fraud

  • Caseworkers did not independently evaluate cases.
    

    They treated the system’s risk flags as ground truth, not as advice.

• 3. Where Bias Enters the LLM Lifecycle

• 2. Social Biases

• Topics

• No advanced math or ML required

  • Focus on intuition, discussion, and conceptual understanding.

• Choose what interests you

  • You can catch up on background knowledge as needed.

  • Work in groups to support and complement each other’s skills.

• Recommended:

  • Interest in machine learning, social science, or AI ethics

  • Basic probability and statistics

  • Introductory Python programming

• Prerequisites

• 1. Measuring Bias in AI

• Where Bias in AI Appears

  • Hiring

  • Predictive policing

  • Ad targeting

• Sources of Bias

  • Human bias & feedback loops

  • Sample imbalance / unreliable data

  • Model & deployment effects

• ​Fairness Criteria

•  

• Bias and Embeddings

  • ​Word embeddings encode stereotypes

  • Embedding geometry

• Causality

  • Simpson’s Paradox

  • Causal inference

• ​Case Study

• Carina I. Hausladen, Manuel Knott, Colin F. Camerer, Pietro Perona

• Social perception of faces in a vision-language model

• Preference elicitation
  • Ordinal vs cardinal preferences
  • Methods of elicitation
• From individual to collective choice
  • Fairness and proportionality principles
  • Key properties: monotonicity ...
• Committee elections

• Participatory budgeting (PB)
  • PB as generalization of committee elections
  • Aggregation methods for PB: proportional and cost-aware
• Human-centered LLMs
  • Learning from human preferences (RLHF)
  • Pluralistic alignment

• 2. Social Choice and

• LLM Alignment

• Guest Lecture

• 3. Clustering Multidimensional Time Series

• Behavioral data as multidimensional time series
• Distance Metrics
  • Local
    • e.g. Euclidean Distance
  • Global
    • Dynamic Time Warping (DTW)

• Clustering Methods
  • Hierarchical clustering:
  • PAM (Partitioning Around Medoids)
  • DBSCAN/HDBSCAN: density-based
• Evaluation & Validation
  • Internal indices 
  • External validation 

• 4. Modeling Social Dilemmas

• Social Dilemma Games
  • Prisoner’s Dilemma, Stag Hunt, Public Goods Game.
  • Emergent dynamics.
• Reinforcement Learning
  • Agents learn from rewards and punishments over time.
• Markov Decision Processes
  • Sequential decision-making under uncertainty.

• Q-Learning
  • learning state–action values through trial and error
  • latest literature on social dilemmas
• Inverse Reinforcement Learning
  • Infer the hidden reward function.
  • Useful in social science: recover fairness concerns, reciprocity, etc.

• Identifying Latent Intentions
• via
• Inverse Reinforcement Learning
• in
• Repeated Public Good Games

• Carina I Hausladen, Marcel H Schubert, Christoph Engel

• MAX PLANCK INSTITUTE
  
• FOR RESEARCH ON COLLECTIVE GOODS