spatiotemporal-graph-feature-engineering

Spatiotemporal Graph Feature Engineering

Problem

Predicting outcomes on spatiotemporal graphs (nodes + edges + time) requires specialized features that capture both spatial relationships and temporal patterns. Standard tabular or time-series features often miss critical graph structure.

Context / Trigger Conditions

• Data has nodes, edges, and time steps (e.g., traffic sensors, hydraulic networks)
• Need to predict node-level values over time
• Graph structure influences predictions (neighbors matter)
• Common in: flood modelling, traffic prediction, power grid, social networks

Solution

Core Insight: Graph Features Trump Individual Features

Critical discovery:

Single-node features:
  elevation:     1179 importance
  degree:         345 importance

Neighbor aggregation:
  neigh_elev_mean: 1239 importance  ← MOST IMPORTANT!

Lesson: Aggregating information from neighbors consistently outperforms single-node features in graph-structured prediction tasks.

Feature Categories

1. Graph Structure Features (Static)

Node Degree:

from_counts = pd.concat([
    edges['from_node'],
    edges['to_node']
]).value_counts()

nodes['degree'] = nodes['node_idx'].map(from_counts).fillna(0).astype(int)

Neighbor Statistics (MOST IMPORTANT):

from collections import defaultdict

# Build adjacency list
adj = defaultdict(list)
for _, row in edges.iterrows():
    adj[int(row['from_node'])].append(int(row['to_node']))

# Aggregate neighbor features
elev_map = dict(zip(nodes['node_idx'], nodes['elevation']))

neighbor_features = []
for n in nodes['node_idx']:
    neighbors = adj.get(n, [])
    if neighbors:
        # Multiple aggregation strategies
        neighbor_values = [elev_map.get(x, elev_map[n]) for x in neighbors]

        neighbor_features.append({
            'neigh_elev_mean': np.mean(neighbor_values),
            'neigh_elev_std': np.std(neighbor_values),
            'neigh_elev_min': np.min(neighbor_values),
            'neigh_elev_max': np.max(neighbor_values),
            'neigh_count': len(neighbors),
        })
    else:
        # Isolated nodes
        neighbor_features.append({
            'neigh_elev_mean': elev_map[n],
            'neigh_elev_std': 0,
            'neigh_elev_min': elev_map[n],
            'neigh_elev_max': elev_map[n],
            'neigh_count': 0,
        })