Informfully

Appearance

User Behavior Knowledge Graph ​

To give researchers insight into individual participants' reading behavior beyond raw numbers, the researcher admin dashboard includes a knowledge-graph visualization. For each participant, the graph maps their reading history as a network of topics and named entities (persons, organizations, locations).

img/knowledge_graph_assets/knowledge_graph.png

INFO

The dropdown in the top-right of the screenshot reads "Demo (sample data)" — it's not clear from this screenshot alone whether that's the only data source currently wired up, or one option among others (e.g. a real per-experiment/participant selector). Confirm before documenting this as working against live participant data.

Graph Structure (as described in the report) ​

  • Topics are shown as red circles, named entities as blue circles (darker shades distinguish persons from organizations/locations).
  • Node size reflects how many articles featuring that topic/entity the participant has read.
  • Edge thickness between two nodes reflects how often they co-occur across those articles.

Wikidata Enrichment ​

Named entities are resolved to Wikidata Q-IDs, and pairs of entities are checked for shared properties (e.g. shared country, occupation, political party, industry, employer); matching pairs get an additional dashed edge annotated with the relation label. Location entities are further linked to topic nodes when Wikidata identifies the location as being within a country/region matching a topic in the graph.

Reading Diversity Metrics ​

Alongside the graph, a metrics bar summarizes a participant's reading diversity for the current week across three measures:

  • Reading diversity — Shannon entropy over the distribution of articles read per topic, normalized by log k (k = number of distinct topics read), giving a score from 0 (all reading concentrated on one topic) to 1 (perfectly even across topics).
  • Top-topic share — the fraction of articles read belonging to the single most-read topic; lower means more evenly spread reading.
  • Connection inequality — the Gini coefficient applied to the distribution of edge weights (co-occurrence counts); 0 means all topic pairs co-occur equally, 1 means all co-occurrence is concentrated on a single pair.

Researchers can export these metrics for all participants and all weeks as a CSV file.

Status ​

This is described in the report as an initial/preliminary implementation, with further extensions expected. Treat it as such: verify against the actual code before relying on any detail on this page for development work.