---
description: "Link predictions: concept pairs that never co-occurred yet probably belong together — worked out from the shape of your own network, with no AI and no cost."
---
<!--
  Markdown representation of https://docs.jotaid.com/en/guide/predictions
  Generated by scripts/gen-markdown-variants.mjs — do not edit by hand.
  Served from the canonical URL under `Accept: text/markdown`.
-->

# Link Predictions

The **PREDICTIONS** section of a node's detail view answers one precise question:

> Which two concepts have **never appeared in the same note**, yet look — from the shape of the whole network — like they should be connected?

This layer is not AI: no network calls, no API key, works offline.

## The Intuition: Rarer Shared Concepts Carry More Signal

Suppose A and B never co-occur, but both relate to C. How much C counts depends on how common C is:

- C is something like "design thinking", appearing everywhere → **almost no information**, since it touches everything
- C appears in only three notes → **strong signal**, because two things landing on the same obscure concept is unlikely to be coincidence

So a shared concept you lean on constantly counts for little, while a rare one you reached from two different directions counts for a lot.

## Reading a Prediction

Each one carries three things:

| Element | Meaning |
|---------|---------|
| The candidate node | The concept proposed as a connection |
| **Confidence** | **Strong**, **Moderate** or **Weak** — how much the shared concepts add up to |
| Shared neighbours | The intermediate concepts behind the judgement, **ordered rarest first** — the top one weighed most |

That last column is what makes it explainable: you need not trust a rating, you can read "so A and B both relate to these three concepts" and judge for yourself.

> **Only pairs that never co-occurred**
>
> Pairs that have co-occurred never appear here — those are **facts**, and they belong to the [matrix](https://docs.jotaid.com/en/guide/matrix). Prediction deals strictly with **what has not happened yet**.

## Where It Falls Short

- **Too few notes** produces almost nothing: not enough shared concepts to work from
- **A universal node** (one you linked from every single note) dilutes the signal — though a concept that common already counts for very little
- Nodes with zero references are skipped

## Versus Semantic Relations

Both hunt for connections you have not made, on entirely different evidence:

| | Based on | Finds |
|---|---------|-------|
| **Link prediction** | Your **link structure** | Concepts worded differently but sitting in similar positions in the network |
| [Semantic relations](https://docs.jotaid.com/en/guide/semantic-relations) | The **meaning** of the text | Notes saying the same thing that you never connected |

One reads structure, the other meaning. Both are free, and they corroborate each other — a pair flagged by both is worth connecting first.

## Next Steps

- [Semantic Relations](https://docs.jotaid.com/en/guide/semantic-relations) — Relatedness by meaning
- [Co-occurrence & Matrix](https://docs.jotaid.com/en/guide/matrix) — Relationships that already exist
- [Node Types](https://docs.jotaid.com/en/guide/node-types) — Badges read from the same network