network-graph-analysis

name	network-graph-analysis
description	Graph and network analysis — graph construction, centrality measures, community detection, graph ML basics, and visualization. Use when data has relational/network structure.
allowed_agents	["data","experiment"]

Network Graph Analysis

Overview

Graph and network analysis extracts structural information from relational data. Use this skill when your data describes connections between entities — citations, protein interactions, social ties, trade flows, co-occurrence matrices, or any adjacency structure. The key insight: if relationships carry information that tabular rows cannot, model the structure explicitly.

When to Use This Skill

Use this skill when:

• Your data is fundamentally relational (entities linked by edges)
• You need to identify influential nodes, bridging nodes, or tightly-knit communities
• You are building or evaluating a graph neural network (GNN)
• You need to visualize network structure for publication or exploration
• Your dataset is a co-occurrence matrix, adjacency matrix, or edge list

Do not reach for graph methods if rows in your tabular dataset are independent (no meaningful pairwise relationships). First run EDA (EDA skill) to understand data shape, then apply this skill.

---

Graph Construction from Data

From a NumPy/SciPy Adjacency Matrix

import numpy as np
import networkx as nx
import scipy.sparse as sp

# Dense adjacency matrix
A = np.array([[0, 1, 0],
              [1, 0, 1],
              [0, 1, 0]])

G = nx.from_numpy_array(A)                          # undirected
G_directed = nx.from_numpy_array(A, create_using=nx.DiGraph())  # directed

# SciPy sparse matrix (memory-efficient for large graphs)
A_sparse = sp.csr_matrix(A)
G_sparse = nx.from_scipy_sparse_array(A_sparse)

From an Edge List CSV

import pandas as pd

# CSV with columns: source, target (and optionally weight)
edges_df = pd.read_csv('edges.csv')

# Unweighted, undirected
G = nx.from_pandas_edgelist(edges_df, source='source', target='target')

# Weighted, directed
G_weighted = nx.from_pandas_edgelist(
    edges_df,
    source='source',
    target='target',
    edge_attr='weight',
    create_using=nx.DiGraph()
)

From Bipartite Data (e.g., users x items)

from networkx.algorithms import bipartite

# Build bipartite graph
B = nx.Graph()
B.add_nodes_from(['u1', 'u2', 'u3'], bipartite=0)  # user nodes
B.add_nodes_from(['i1', 'i2'], bipartite=1)          # item nodes
B.add_edges_from([('u1','i1'), ('u2','i1'), ('u2','i2'), ('u3','i2')])

# Project onto one set (user-user co-occurrence)
user_nodes = {n for n, d in B.nodes(data=True) if d['bipartite'] == 0}
G_projected = bipartite.weighted_projected_graph(B, user_nodes)

Weighted vs Unweighted / Directed vs Undirected

Choice	When to use
Undirected	Symmetric relationships (co-authorship, friendship)
Directed	Asymmetric relationships (citations, follows, flows)
Unweighted	Only presence/absence of edge matters
Weighted	Edge strength matters (call frequency, correlation strength)

Use nx.DiGraph() for directed, nx.Graph() for undirected. Add weight edge attributes for weighted graphs — most centrality functions respect the weight keyword.

---

Basic Graph Statistics

G = nx.karate_club_graph()  # example

print(f"Nodes:               {G.number_of_nodes()}")
print(f"Edges:               {G.number_of_edges()}")
print(f"Density:             {nx.density(G):.4f}")         # edges / possible edges

degrees = [d for _, d in G.degree()]
print(f"Average degree:      {sum(degrees)/len(degrees):.2f}")
print(f"Max degree:          {max(degrees)}")

# Connected components
components = list(nx.connected_components(G))
print(f"Connected components: {len(components)}")
print(f"Largest component:   {max(len(c) for c in components)} nodes")

# Diameter (only valid for connected graphs)
if nx.is_connected(G):
    print(f"Diameter:            {nx.diameter(G)}")
else:
    # Diameter of largest component
    largest = G.subgraph(max(components, key=len))
    print(f"Diameter (LCC):      {nx.diameter(largest)}")

# Clustering coefficient
()
()

---

Centrality Measures

Degree Centrality — Who Has the Most Connections?

The simplest measure. Node importance = fraction of other nodes it connects to. Use as a baseline; fast to compute.

degree_cent = nx.degree_centrality(G)

# For weighted graphs, use strength (sum of weights)
strength = dict(G.degree(weight='weight'))

# Top 5 nodes by degree centrality
top_degree = sorted(degree_cent.items(), key=lambda x: x[1], reverse=True)[:5]
print("Top nodes by degree centrality:", top_degree)

Betweenness Centrality — Bridges Between Communities

Fraction of shortest paths passing through a node. High betweenness = bottleneck / broker. Expensive: O(VE) for unweighted, slower for weighted. Sample for large graphs.

# Exact
between_cent = nx.betweenness_centrality(G, normalized=True, weight='weight')

# Approximate for large graphs (sample k sources)
between_approx = nx.betweenness_centrality(G, k=200, normalized=True)

top_between = sorted(between_cent.items(), key=lambda x: x[1], reverse=True)[:5]
print("Top nodes by betweenness:", top_between)

Closeness Centrality — Reaches Others Quickly

Average shortest path distance to all other nodes (inverted). High closeness = spreads information fast. Meaningful only in connected graphs (or apply per component).

closeness_cent = nx.closeness_centrality(G)

top_close = sorted(closeness_cent.items(), key=lambda x: x[1], reverse=True)[:5]
print("Top nodes by closeness:", top_close)

Eigenvector Centrality — Connected to Important Nodes

Score proportional to scores of neighbors. Being connected to high-scoring nodes raises your own score. PageRank is a damped variant. Use for undirected or directed networks where prestige propagates.

eigen_cent = nx.eigenvector_centrality(G, max_iter=1000, weight='weight')

# PageRank for directed graphs
pagerank = nx.pagerank(G, alpha=0.85, weight='weight')

top_eigen = sorted(eigen_cent.items(), key=lambda x: x[1], reverse=True)[:5]
print("Top nodes by eigenvector centrality:", top_eigen)

When to Use Each

Measure	Best for
Degree	Quick scan, finding hubs, baseline
Betweenness	Finding brokers, bridges, critical infrastructure nodes
Closeness	Information diffusion, finding broadcast centers
Eigenvector/PageRank	Influence, authority, prestige propagation

Compute all four and compare — they often disagree and the disagreement is informative.

---

Community Detection

Louvain — Fast, Large Graphs (Recommended Default)

Maximizes modularity greedily. Non-deterministic; run multiple times and pick highest modularity. Available via python-louvain (community package) or igraph.

# Using python-louvain
import community as community_louvain

partition = community_louvain.best_partition(G, weight='weight', random_state=42)
modularity = community_louvain.modularity(partition, G)

num_communities = len(set(partition.values()))
print(f"Communities: {num_communities}, Modularity: {modularity:.4f}")

# Assign as node attribute
nx.set_node_attributes(G, partition, 'community')

# Using igraph (faster on very large graphs)
import igraph as ig
G_ig = ig.Graph.from_networkx(G)
louvain_result = G_ig.community_multilevel(weights='weight')
print(f"Modularity (igraph): {louvain_result.modularity:.4f}")

Girvan-Newman — Hierarchical, Small Graphs

Removes edges with highest betweenness iteratively, revealing community hierarchy. Slow (O(E²V)); use only when n < 1000 and you need a dendrogram.

from networkx.algorithms.community import girvan_newman

comp = girvan_newman(G)
# Get partitions at different resolution levels
for communities in itertools.islice(comp, 3):
    print(f"k={len(communities)}: {[len(c) for c in communities]}")

Spectral Clustering — Exactly k Communities

Use when you know k in advance. Builds graph Laplacian, finds k eigenvectors, runs k-means in spectral space.

from sklearn.cluster import SpectralClustering
import numpy as np

A = nx.to_numpy_array(G)
sc = SpectralClustering(
    n_clusters=5,
    affinity='precomputed',
    assign_labels='kmeans',
    random_state=42
)
labels = sc.fit_predict(A)

Label Propagation — Very Fast, Non-Deterministic

Each node inherits the most common label of its neighbors. Converges in O(E). Results vary between runs; use when speed matters more than stability.

from networkx.algorithms.community import label_propagation_communities

communities = list(label_propagation_communities(G))
print(f"Found {len(communities)} communities")

---

Graph ML Basics

When Does Graph Structure Help?

Graph structure adds value beyond tabular features when:

• Nodes are connected in ways that predict the target (social influence, protein proximity)
• Labels propagate through edges (homophily: connected nodes share labels)
• Global topology matters (scale-free vs random network properties)

If nodes have no meaningful connections or edges are arbitrary, stick with tabular models.

GNN Overview: Message Passing

At each layer, each node aggregates feature vectors from its neighbors, then updates its own representation. After k layers, a node sees its k-hop neighborhood.

Key architectures:

• GCN (Graph Convolutional Network): normalized sum aggregation — simple, effective
• GraphSAGE: sample-and-aggregate — scales to large graphs via mini-batching
• GAT (Graph Attention Network): learns attention weights over neighbors — best when neighbor importance varies

Task Types

Task	Input	Output	Example
Node classification	Node features + graph	Label per node	Classify paper topics in citation network
Link prediction	Node features + partial graph	Edge probability	Recommend friends, predict protein interactions
Graph classification	Full graph	Label per graph	Classify molecules, detect fraud subgraphs

Libraries

# PyTorch Geometric (PyG) — most popular, extensive model zoo
from torch_geometric.data import Data
from torch_geometric.nn import GCNConv, GATConv, SAGEConv
import torch

data = Data(
    x=node_features,       # [num_nodes, num_features]
    edge_index=edge_index, # [2, num_edges] — COO format
    y=labels               # [num_nodes] for node classification
)

class GCN(torch.nn.Module):
    def __init__(self, in_channels, hidden, out_channels):
        super().__init__()
        self.conv1 = GCNConv(in_channels, hidden)
        self.conv2 = GCNConv(hidden, out_channels)

    def forward(self, data):
        x, edge_index = data.x, data.edge_index
        x = self.conv1(x, edge_index).relu()
        x = torch.nn.functional.dropout(x, p=0.5, training=self.training)
        x = self.conv2(x, edge_index)
        return x

# DGL — alternative with strong support for heterogeneous graphs
import dgl
g = dgl.from_networkx(G)

---

Visualization

NetworkX + Matplotlib — Small Graphs (<500 nodes)

import matplotlib.pyplot as plt
import networkx as nx

fig, ax = plt.subplots(figsize=(12, 8))

# Layout algorithms
pos = nx.spring_layout(G, seed=42)           # force-directed — general purpose
# pos = nx.kamada_kawai_layout(G)            # minimizes edge crossings — slower but cleaner
# pos = nx.circular_layout(G)               # nodes on a circle — good for small n
# pos = nx.spectral_layout(G)               # based on graph Laplacian eigenvectors

# Color by community
communities = community_louvain.best_partition(G)
colors = [communities[n] for n in G.nodes()]

# Size by degree
sizes = [300 * nx.degree_centrality(G)[n] + 50 for n in G.nodes()]

nx.draw_networkx(
    G, pos,
    node_color=colors,
    node_size=sizes,
    cmap=plt.cm.tab20,
    with_labels=True,
    font_size=8,
    edge_color='gray',
    alpha=0.8,
    ax=ax
)
ax.set_title("Network with community coloring")
plt.tight_layout()
plt.savefig('network.png', dpi=150, bbox_inches='tight')

Pyvis — Interactive HTML Visualization

from pyvis.network import Network

net = Network(height='750px', width='100%', bgcolor='#222222', font_color='white')
net.from_nx(G)

# Customize nodes
for node in net.nodes:
    node['size'] = degree_cent[node['id']] * 50 + 5
    node['title'] = f"Node {node['id']}, degree={G.degree(node['id'])}"

net.set_options("""
{
  "physics": {
    "forceAtlas2Based": {
      "gravitationalConstant": -50,
      "centralGravity": 0.01,
      "springLength": 100
    },
    "solver": "forceAtlas2Based"
  }
}
""")
net.save_graph('network.html')

Gephi — Large Graphs (>5000 nodes)

Export from networkx and open in Gephi for GPU-accelerated layout and interactive exploration:

# Export to GEXF (Gephi's native format)
nx.write_gexf(G, 'network.gexf')

# Or GraphML
nx.write_graphml(G, 'network.graphml')

In Gephi: use ForceAtlas2 layout, color by modularity class (run it under Statistics), size by degree. Export as SVG for publication.

---

Evaluation

Community Detection

from sklearn.metrics import normalized_mutual_info_score, adjusted_rand_score

# When ground truth labels are known
true_labels = [ground_truth[n] for n in G.nodes()]
pred_labels = [partition[n] for n in G.nodes()]

nmi = normalized_mutual_info_score(true_labels, pred_labels)
ari = adjusted_rand_score(true_labels, pred_labels)
print(f"NMI: {nmi:.4f}, ARI: {ari:.4f}")

# When no ground truth — use modularity
modularity = community_louvain.modularity(partition, G)
print(f"Modularity Q: {modularity:.4f}")
# Q > 0.3 is generally considered significant community structure

Link Prediction

from sklearn.metrics import roc_auc_score
import random

# Create train/test split by hiding edges
edges = list(G.edges())
test_edges = random.sample(edges, int(0.2 * len(edges)))
G_train = G.copy()
G_train.remove_edges_from(test_edges)

# Generate negative samples (non-edges)
non_edges = list(nx.non_edges(G))
test_non_edges = random.sample(non_edges, len(test_edges))

# Score with common neighbors as baseline predictor
def score_pairs(G, pairs):
    return [len(list(nx.common_neighbors(G, u, v))) for u, v in pairs]

scores_pos = score_pairs(G_train, test_edges)
scores_neg = score_pairs(G_train, test_non_edges)

y_true = [1]*len(test_edges) + [0]*len(test_non_edges)
y_score = scores_pos + scores_neg
auc = roc_auc_score(y_true, y_score)
print(f"Link prediction AUC (common neighbors): {auc:.4f}")

# networkx also has Jaccard, Adamic-Adar, resource allocation predictors built-in

Node Classification

Use standard classification metrics (accuracy, F1, AUC) from sklearn. Evaluate on a held-out set of nodes — do not leak graph structure from test nodes into training via multi-hop neighborhoods unless you explicitly account for it.

---

Best Practices

1. Inspect the degree distribution first — scale-free (power-law) vs Poisson shapes imply different generative processes
2. Always check for isolated nodes and giant component fraction — disconnected graphs require careful handling
3. Self-loops and multi-edges — decide explicitly: nx.Graph() removes multi-edges, use nx.MultiGraph() to keep them
4. Large graphs (>100k nodes) — switch to igraph or graph-tool; networkx is Python-native and slow at scale
5. Reproducibility — set random_state/seed for any non-deterministic algorithm (Louvain, spring layout)
6. Null models — compare your graph properties against an Erdos-Renyi or configuration model null to validate findings

---

Common Pitfalls

• Treating directed graphs as undirected: PageRank on undirected graphs is meaningless — check nx.is_directed(G)
• Diameter on disconnected graphs: Always check nx.is_connected(G) first
• Betweenness on large graphs without sampling: It will run for hours — use k parameter to approximate
• Comparing modularity across algorithms: Modularity is maximized differently by each algorithm; comparisons are not apples-to-apples
• Forgetting to normalize centrality: Raw degree vs degree centrality (0–1) depend on graph size — always normalize for cross-graph comparison