Adaptive Learning Engines

Overview

ImpactMatrix runs five continuously operating machine learning engines that make the platform increasingly accurate and relevant over time. Unlike static tools that require manual recalibration, these engines learn from observed outcomes — updating their models automatically as new data arrives.

The result: the longer you use the platform, the better its recommendations become — for your organization and for the community of users as a whole.

The Five Engines

1. Scoring Weight Learner

What it optimizes: Which organizational characteristics best predict consortium success

How it works: After each consortium cycle completes, the engine correlates the entry-point characteristics of member entities (size, sector, geographic context, prior ESGETC scores) with actual outcomes (SDG impact achieved, collaboration quality, goal completion rates). It uses Bayesian weight optimization to adjust how it scores similar entities in the future.

Effect on you: Entity recommendations and peer benchmarking become more accurate over time, especially within your sector and geography.

2. Adaptive Stakeholder Predictor

What it optimizes: Who is most likely to accept an invitation to participate in a consortium or survey

How it works: After each invitation batch, the engine records actual acceptance and response rates, updating its predictions using exponential smoothing (an online learning algorithm that prioritizes recent behavior). It builds profiles by organization type, geography, time of year, and stakeholder role.

Effect on you: Invitation lists are pre-ranked by likelihood to engage, saving coordination time and improving your consortium composition quality.

3. Adaptive Threshold Controller

What it optimizes: Alert thresholds that balance actionability vs. noise

How it works: The engine monitors how users respond to alerts — which ones trigger action versus which are dismissed or ignored. It adjusts thresholds dimension-by-dimension to reduce false positives while maintaining sensitivity to genuine issues.

Effect on you: Fewer irrelevant alerts. When you receive a notification, it is more likely to require attention.

4. SDG Impact Learner

What it optimizes: Which Organic Whole Use (OWU) intervention methods actually achieve measurable SDG impact

How it works: As organizations complete action cycles and report outcomes, the engine runs empirical validation comparing stated methods against measured impact across SDGs. Where certain approaches consistently outperform others (validated through A/B comparisons within peer groups), those methods are recommended more often.

Effect on you: Action recommendations are grounded in evidence from organizations similar to yours, not just theoretical best practices.

5. Narrative Multi-Armed Bandit

What it optimizes: Which framing, language, and narratives drive the highest stakeholder engagement

How it works: The engine uses Thompson Sampling — a Bayesian multi-armed bandit algorithm — to test different cultural narratives, framings, and message structures across the platform’s stakeholder communication features. It balances exploration (trying new narratives) with exploitation (using what works).

Effect on you: Survey invitations, consortium outreach messages, and engagement communications are automatically refined to maximize response rates in your community context.

How the Engines Are Coordinated

The five engines are orchestrated by a central Learning Controller that:

Collects outcome data from every user interaction involving prediction or recommendation
Routes outcomes to the appropriate engine
Schedules regular update cycles (most engines update daily; the Bandit updates in real-time)
Ensures updates don’t degrade previously learned patterns (catastrophic forgetting prevention)

The data flow:

User Action → Outcome Collected → Engine Updates → Platform Adapts

Privacy & Aggregation

Learning engines operate on aggregated, anonymized outcome data. No individual user’s data is used to update models without consent. Organization-level outcomes are contributed to aggregate learning only when:

The organization has opted into the platform’s collaborative improvement program (enabled by default; opt-out available at Settings → Privacy → Learning Contribution)
The outcome data passes anonymization checks

Individual organizations can view their contribution to the aggregate learning pool at Analytics → Learning Contribution.

Transparency & Explainability

Every recommendation powered by a learning engine includes an explanation card:

Which engine generated the recommendation
What data signals drove it
Confidence score (0–100%)
How many similar organizations the model was trained on

You can always override AI recommendations. Overrides are logged and feed back into the learning loop.