7 AI for Social Good
What is AI4SG?
- Definition: AI research that can deliver societal benefits now or in the near future.
Typical workflow: from research to deployment
-
Problem → Data → Model/Decision method → Deployment → Evaluation (often RCT) → Iteration
-
Deployment is highly non-trivial (real-time systems + constraints + stakeholders).
Example domain map (from lecture)
- Environmental conservation, food rescue, mental/behavioral health, etc.
Example Projects & Key Ideas
Food Rescue Platforms (412 Food Rescue / Food Rescue Hero)
Goal: notify the “right” volunteers to claim a rescue (food donation pickup), improving operational efficiency.
Baseline vs ML: core metrics
-
Hit Ratio / HR@k: % of rescues claimed by the k notified volunteers.
-
RCT deployment results (example table):
-
Control: Hit Ratio 0.468, Claim Ratio 0.807
-
ML: Hit Ratio 0.651 (p=0.001), Claim Ratio 0.882 (p=0.047)
-
Why “pure recommender” caused problems
-
Caveat: ML model tends to over-notify a small set of frequent “super volunteers”.
-
Homework framing: the key issue that motivates online planning is notification fatigue.
Fix: ML + Online Planning (notification budget)
-
Add constraint: each volunteer receives at most L notifications per day; plan with projected future rescues.
-
With L = 5, avoids over-concentration; HR@k reported as 0.645 (better than current practice).
oaicite:8
Mental Health: Simulated Patient for CBT Training (PATIENT-Ψ)
Goal: train mental health professionals via a consistent simulated patient.
-
System idea: LLM + a detailed “Cognitive Model” (history, core beliefs, emotions, etc.) to create a consistent patient.
-
Lecture reference: PATIENT-Ψ (EMNLP 2024).
Behavioral Health: PeerCoPilot (resource recommendation for peer providers)
Goal: help peer providers give reliable, verifiable resources; reduce cognitive burden.
-
Purpose statement: improve efficiency, reduce cognitive burden, improve service quality.
-
Approach: LLM + RAG-based resource recommendation + benefit eligibility checker (homework wording).
-
User study results shown in lecture:
-
PeerCoPilot: Contact provided 100%, Bad link 0%
-
Baseline: Contact 56%, Bad link 11%
-
-
Willingness to use: all peer providers & service users willing to use PeerCoPilot; deployed now.