[Paper Review] Trace Complexity of Network Inference

Bruno Abrahão, Flavio Chierichetti+2 more|arXiv (Cornell University)|Aug 13, 2013

Complex Network Analysis Techniques24 references17 citations

TL;DR

This paper introduces the concept of trace complexity to quantify the number of epidemic spread traces needed to infer network topology or properties. It presents efficient algorithms for exact inference on trees and bounded-degree graphs using few traces, and proves information-theoretic lower bounds showing that general graphs require many traces, while degree distribution recovery is feasible with only O(n) traces.

ABSTRACT

The network inference problem consists of reconstructing the edge set of a network given traces representing the chronology of infection times as epidemics spread through the network. This problem is a paradigmatic representative of prediction tasks in machine learning that require deducing a latent structure from observed patterns of activity in a network, which often require an unrealistically large number of resources (e.g., amount of available data, or computational time). A fundamental question is to understand which properties we can predict with a reasonable degree of accuracy with the available resources, and which we cannot. We define the trace complexity as the number of distinct traces required to achieve high fidelity in reconstructing the topology of the unobserved network or, more generally, some of its properties. We give algorithms that are competitive with, while being simpler and more efficient than, existing network inference approaches. Moreover, we prove that our algorithms are nearly optimal, by proving an information-theoretic lower bound on the number of traces that an optimal inference algorithm requires for performing this task in the general case. Given these strong lower bounds, we turn our attention to special cases, such as trees and bounded-degree graphs, and to property recovery tasks, such as reconstructing the degree distribution without inferring the network. We show that these problems require a much smaller (and more realistic) number of traces, making them potentially solvable in practice.

Motivation & Objective

• To define and analyze the trace complexity—the number of epidemic traces required for accurate network inference.
• To develop efficient inference algorithms that minimize the number of required traces while maintaining high accuracy.
• To establish information-theoretic lower bounds on trace complexity for general graphs, showing that exact inference is impractical in the worst case.
• To identify special graph classes—such as trees and bounded-degree graphs—where inference becomes tractable with significantly fewer traces.
• To investigate whether non-topological network properties, like the degree distribution, can be recovered with far fewer traces than full topology reconstruction.

Proposed method

• Proposes a diffusion model where infection spreads along edges with a fixed probability and random incubation times, modeling real-world cascading phenomena.
• Introduces the First-Edge algorithm, which extracts edges from the first two nodes in each trace, and proves it is nearly optimal for general graphs.
• Applies the Berry-Esseen theorem to bound the concentration of sums of exponential random variables, enabling precise analysis of trace likelihoods.
• Uses probabilistic coupling and moment analysis to compare likelihoods of two similar graphs (G₀ and G₁), showing they are indistinguishable with high probability under limited traces.
• Develops a novel algorithm for reconstructing the degree distribution using O(n) traces by aggregating infection time patterns across multiple traces.
• Establishes tight upper and lower bounds on trace complexity via asymptotic analysis, showing Ω(nΔ/log²Δ) traces are necessary and O(nΔ log n) are sufficient for general graphs.

Experimental results

Research questions

• RQ1 What is the minimum number of epidemic traces required to reconstruct the full edge set of a general network with high probability?
• RQ2 Can exact network inference be achieved with significantly fewer traces on special graph classes like trees or bounded-degree graphs?
• RQ3 Is it possible to recover non-topological properties—such as the degree distribution—using far fewer traces than needed for full topology reconstruction?
• RQ4 How much useful information can be extracted from the tail of a trace, beyond the first few infection events?
• RQ5 Are there information-theoretic limits on trace complexity that make exact inference infeasible for general graphs?

Key findings

• For general connected graphs, Ω(nΔ/log²Δ) traces are information-theoretically necessary for exact inference, and O(nΔ log n) traces are sufficient.
• On trees, a simple algorithm reconstructs the full topology using only O(log n) traces, demonstrating a dramatic improvement over general-case bounds.
• For bounded-degree graphs, exact inference is possible with O(poly(Δ) log n) traces under the assumption that epidemics spread throughout the network.
• Reconstructing the degree distribution of a network can be achieved with only O(n) traces, a significant reduction compared to full topology inference.
• The First-Edge algorithm, which uses only the first two nodes in each trace, is nearly optimal for general graphs, as the tail of traces contributes negligible information for exact inference.
• In the worst case, no algorithm can distinguish between two similar graphs (G₀ and G₁) with high probability using fewer than Ω(nΔ/log²Δ) traces, due to the statistical indistinguishability of their trace likelihoods.