Recommendations

One of the most powerful applications of the Learning Graph is resource discovery. When you open the Global library view for any entity (course, campus, profile), the graph traverses multiple paths to find publicly available resources that teach the same or related topics. Scoring is strength-aware so stronger signals contribute more.


How Recommendations Are Scored

How Recommendations Are Scored

When you open the Global library view, ikigize traverses up to six paths through the graph to find relevant resources. Each path has a relevance weight — higher weight means the resource is more likely to match your context.

Direct topic overlapweight 1.00

Resources that directly DEVELOP the exact same topics

Subtopic expansionweight 0.60

Resources in child topics via SUBTOPIC_OF traversal

Essential lateral topicsweight 0.40

Resources in strongly related topics (RELATED_ESSENTIAL_TOPIC)

Optional lateral topicsweight 0.20

Resources in weakly related topics (RELATED_OPTIONAL_TOPIC)

1-hop ancestor siblingsweight 0.30

Resources in sibling subtopics one level up the hierarchy

2-hop ancestor siblingsweight 0.15

Resources in sibling subtopics two levels up the hierarchy

Scores are summed per resource, de-duplicated, and sorted. If no resources score above the threshold, the system falls back to broad domain-level traversal.

Scores are summed per resource, de-duplicated, and sorted. If no resources score above the threshold, the system falls back to broad domain-level traversal. The exact weights may be tuned; the important point is that direct topic overlap counts most, with subtopics and lateral topics adding breadth.


The library search bar combines:

  • Instant client-side text filtering (e.g. ~150 ms) for quick narrowing.
  • Graph-augmented search (e.g. ~400 ms) that finds Topics matching the query in the graph, then returns resources that DEVELOP those topics.

So you get semantic search without extra steps: typing “attention mechanisms” can surface resources mapped to that topic even if the phrase does not appear in the resource title or description.


Collaborative Filtering (Future)

With behavioral edges (ENROLLED_IN, COMPLETED, etc.), the system can add collaborative-filtering style paths — e.g. “learners who completed the same courses as you also engaged with these topics” — and combine them with content-based (topic overlap) scoring for hybrid recommendations.


Next Steps