Temporal and Evolving Networks
A static aggregate says that edges occurred during a window. A temporal network also records when, so a valid path must respect event order. Aggregation can create static routes that no message, infection, or resource could have followed in time.
Prerequisites
Complete link prediction. Keep its temporal cutoff discipline and return to the directed event stream rather than the undirected training projection.
Learning goals
- Distinguish event streams, snapshots, and static aggregates.
- Compute a simple earliest-arrival reachability scan.
- Exhibit a static path that is not time-respecting.
- Separate evolving structure from changes in collection, missingness, and coverage.
Analysis contract
- Graph type: the valid directed temporal multigraph; every event retains its identity and UTC timestamp, parallel events remain separate, and weight means strength rather than duration.
- Normalization: daily snapshots report raw event counts; temporal reachability is Boolean and earliest arrival is an absolute timestamp, with no rescaling across windows.
- Disconnected case: if no nondecreasing-time event sequence reaches a target, temporal distance is undefined even when the static aggregate contains a directed path.
- Interpretation limit: time-respecting reachability and snapshot change are evidence, not realized diffusion, influence, future truth, mechanism, or causality.
Event order changes reachability
Starting from Ada before the observation window, the event sequence Ada→Ben, Ben→Cy, then Cy→Dia is time-respecting. Starting from Fox, the only outgoing event to Ada occurs on October 3, after Ada→Ben, Ben→Cy, and Cy→Dia have already happened. The static aggregate forgets this order and incorrectly suggests a usable Fox-to-Dia route.
This scan assumes instantaneous events and allows a later event only when its source has already been reached. Other applications need duration, simultaneous-event rules, waiting-time limits, node availability, or repeated transmission. A snapshot series introduces a separate choice of window boundaries; daily counts below do not make differently sized windows comparable.
Worked example
import csv
from collections import Counter
from datetime import datetime
from pathlib import Path
import networkx as nx
fixture = Path("tests/fixtures/graph-data-science")
with (fixture / "interaction-nodes.csv").open(encoding="utf-8") as handle:
node_ids = [row["node_id"] for row in csv.DictReader(handle)]
with (fixture / "interaction-relationships.csv").open(encoding="utf-8") as handle:
events = [row for row in csv.DictReader(handle) if row["record_status"] == "valid"]
events.sort(key=lambda row: (row["observed_at"], row["relationship_id"]))
aggregate = nx.DiGraph()
aggregate.add_nodes_from(node_ids)
aggregate.add_edges_from((row["source"], row["target"]) for row in events)
def earliest_arrivals(source):
reached = {source: None}
for event in events:
if event["source"] not in reached:
continue
source_time = reached[event["source"]]
event_time = datetime.fromisoformat(event["observed_at"].replace("Z", "+00:00"))
if source_time is None or source_time <= event_time:
reached.setdefault(event["target"], event_time)
return reached
from_fox = earliest_arrivals("fox")
from_ada = earliest_arrivals("ada")
daily = Counter(row["observed_at"][:10] for row in events)
print(f"static_fox_to_dia={nx.has_path(aggregate, 'fox', 'dia')} temporal_fox_to_dia={'dia' in from_fox}")
arrival = from_ada["dia"].isoformat().replace("+00:00", "Z")
print(f"temporal_ada_to_dia={'dia' in from_ada} arrival={arrival}")
print(f"snapshots={sorted(daily.items())}")
Expected output:
static_fox_to_dia=True temporal_fox_to_dia=False
temporal_ada_to_dia=True arrival=2026-10-02T09:00:00Z
snapshots=[('2026-10-01', 4), ('2026-10-02', 4), ('2026-10-03', 2)]
The counterexample shows why temporal order is part of topology for flow-like questions.
Limitations
- The timestamps record observations, not necessarily the exact start, duration, or causal order of real interactions.
- Day boundaries can manufacture apparent change; compare fixed, justified windows and exposure.
- Late arrival, backfill, and collection outages can look like network evolution.
- Computing temporal paths at scale can be expensive, and the appropriate waiting rule is domain-specific.
- Reusing post-cutoff events in earlier features creates time leakage even if the final labels are held out.
Practice
Start the scan from Ben immediately after r03. Decide whether equality at the timestamp boundary is allowed, document that convention, and list the reachable nodes.
Next steps
Continue to knowledge graph fundamentals to separate typed assertions and query semantics from interaction-network structure.
Primary sources
Checked 2026-10-18: P. Holme and J. Saramäki, Temporal networks; R. K. Pan and J. Saramäki, Path lengths, correlations, and centrality in temporal networks; NetworkX, has_path.