A Spectral-Based Framework for Hypothesis Testing in Populations of Networks

Q: What is the power of the test for weighted networks?

The power of the test for weighted networks is presented in Theorem 6. It involves the assumptions of Theorem 4 and the alternative model F1 = F2. The theorem defines Z'' as an n x n matrix with zero diagonals and calculates Sa and Sb based on Z''. The conditions for the power of the test are satisfied if either (i) 3min(i,k,l)Sa(Z''ik)3 + bn3min(i,k,l)Sb(Z''ik)3 > 0 or (ii) -3max(i,k,l)Sa(Z''ik)3 - bn3max(i,k,l)Sb(Z''ik)3 > 0. In this case, the limit of P(|th| > ua/2) = 1, a > 0. The proof is similar to that of Theorem 3. When S > 2 groups are present, the test considers symmetric binary adjacency matrices A generated from a symmetric link probability matrix Ps. This theorem provides a framework for testing weighted networks and determining the power of the test in different scenarios.

Question

1. What is the proposed spectral-based testing statistic for populations of networks and its asymptotic null distribution?

2. What is the goal of the proposed spectral-based test for binary networks?

3. What is the normalized matrix in binary network testing?

4. What is the alternative test statistic based on estimated link probability matrices?

Accepted Answer

The proposed spectral-based testing statistic for populations of networks is based on the trace of the third power of a centered and scaled adjacency matrix. This statistic is proven to converge to the standard normal distribution as the number of nodes tends to infinity. The asymptotic null distribution of the proposed testing statistic is derived, and an asymptotic power result is also obtained. The statistic is conceptually simple, computationally friendly, and has shown superior performance over existing methods in various simulation studies. It can be applied to both binary and weighted networks in two-sample and multiple-sample frameworks.

Accepted Answer

The goal of the proposed spectral-based test for binary networks is to test whether two samples of networks have the same graph structure or not. This is equivalent to testing the null hypothesis H0: P1 = P2 against the alternative hypothesis H1: P1 = P2. The test utilizes results from random matrix theory to analyze the spectral properties of inhomogeneous networks, which are used heavily in this work. By comparing the link probability matrices P1 and P2, the test determines if the two samples of networks have the same graph structure or not.

Accepted Answer

The normalized matrix, Z, is calculated using elements A1,ij - A2,ij n 1 m 1 P1,ij (1-P1,ij) + 1 m 2 P2,ij (1-P2,ij) if i = j, and B ij if i = j. It is derived from two samples of networks, A (k) 1 m 1 k=1 and A (k) 2 m 2 k=1, sampled from link probability matrices P1 and P2, respectively. The matrix elements are based on the sample average of adjacency matrices in each group, Au, and a diagonal matrix B with i.i.d. random variables determining the sign of each element. The test statistic EQUATION is used to analyze the spectral properties of binary networks.

Accepted Answer

The alternative test statistic based on estimated link probability matrices involves plug-in estimates of P1 and P2, denoted as P1 and P2, respectively. The empirical version of the normalized matrix Z in equation (2.2) can be written as Zij = A1,ij - A2,ij n 1 m 1 P1,ij (1- P1,ij) + 1 m 2 P2,ij (1- P2,ij) if i = j. The resulting test statistic is given by equation (2.7). Under the two-sample framework of binary networks, if max i,j | Pu,ij - P u,ij | = o p (1), and P1 = P2 under the null hypothesis, the scaled test statistic th has an asymptotic distribution of th ~ N(0, 1) as n approaches infinity. The proof of this limiting law can be found in the supplementary material, which relies on rewriting Tr(Z3) as Tr(Z3) + 3Tr(Z2 * H) + 3Tr(Z * H^2) + Tr(H * Z). Each term in the right-hand side of this equality is o(1).

Accepted Answer

The three methods for estimating link probability matrices are SPE-AVG, SPE-SBM, and SPE-MNBS. SPE-AVG is based on simple averages of the adjacency matrices of group u. It requires large sample sizes to achieve good performance. SPE-SBM approximates any graph with a stochastic block model (SBM) and estimates P u,ij by the sample mean of all entries in the submatrix over all A (k) u, restricted to the corresponding block consisting of the communities of i and j. SPE-MNBS is based on modified neighborhood smoothing (MNBS) proposed in Zhao et al. (2019). It performs neighborhood smoothing to the matrix A, resulting in a better bias-variance tradeoff and a smaller estimation error. SPE-MNBS does not put structure conditions on the networks, making it generally applicable.

Accepted Answer

The asymptotic power guarantee refers to the power of a test based on a specific model, where the power is asymptotically one when either of the sets S a or S b is large enough. This guarantee is characterized by the minimum signal difference between P 1 and P 2 required for the theorem to hold. In the given context, Theorem 3 states that if Z '' ij >= 0 and a > O((m u n) -3/2 ), the power is asymptotically one. This means that the test has a high probability of correctly rejecting the null hypothesis when the alternative model is true. The conditions (i) and (ii) in the theorem define the minimum signal difference required for the power guarantee to hold. The proof of this theorem is provided in the supplementary material. Overall, the asymptotic power guarantee ensures that the test has a high probability of correctly detecting the true alternative model.

Accepted Answer

The extension to weighted networks focuses on a more general framework that considers weighted networks. In this framework, two sequences of distributions, F 1 and F 2, are defined on bounded intervals and specified by some parameters. Symmetric weighted adjacency matrices, A (1) and A (2), are used to represent undirected networks without self-loops. The variance of these adjacency matrices, denoted as S u,ij, is calculated for each network. Estimates of S u,ij, denoted as Su,ij, are obtained using various methods. The testing problem in this context is equivalent to testing whether F 1 equals F 2. The test statistic is obtained by replacing the link probability matrix P u,ij and its variance Su,ij in equations (2.1) and (2.7). The estimation methods differ for binary networks and weighted networks. For practical applications, the covariance matrices S 1 and S 2 need to be estimated, ensuring the asymptotic normality of the new test statistic. Two estimates of S u,ij are considered: SPE-AVG and SPE-SBM. The error conditions in Theorem 5 are satisfied as long as K u < m 1/2 u n 1-b, where b is a small positive number. This extension to weighted networks allows for a more comprehensive analysis of weighted networks in research.

Accepted Answer

The power of the test for weighted networks is presented in Theorem 6. It involves the assumptions of Theorem 4 and the alternative model F1 = F2. The theorem defines Z'' as an n x n matrix with zero diagonals and calculates Sa and Sb based on Z''. The conditions for the power of the test are satisfied if either (i) 3min(i,k,l)Sa(Z''ik)3 + bn3min(i,k,l)Sb(Z''ik)3 > 0 or (ii) -3max(i,k,l)Sa(Z''ik)3 - bn3max(i,k,l)Sb(Z''ik)3 > 0. In this case, the limit of P(|th| > ua/2) = 1, a > 0. The proof is similar to that of Theorem 3. When S > 2 groups are present, the test considers symmetric binary adjacency matrices A generated from a symmetric link probability matrix Ps. This theorem provides a framework for testing weighted networks and determining the power of the test in different scenarios.

Accepted Answer

The goal is to test whether there are any differences in the distributions of the S groups, which is equivalent to testing H 0: P 1 = P 2 = ... = P S against H 1: P s are not all equal. This is analogous to one-way ANOVA. It helps in understanding if there are significant variations among the groups being compared. By analyzing the pairwise normalized matrices and calculating the th (s) values, researchers can determine if the null hypothesis can be rejected, indicating significant differences in the distributions. This approach is useful in various fields, including social network analysis, where understanding the differences in network structures can provide valuable insights into the underlying dynamics and relationships within the network.

Accepted Answer

In the simulation studies, five tests (SPE-AVG, SPE-SBM, SPE-MNBS, DFRO, and VEC) were compared. The tests evaluated the performance of plug-in estimators for link probability matrices (AVG, SBM, and MNBS) and the estimated distance between network distributions based on the Frobenius measure. The tests were compared to the test proposed in Ghoshdastidar and von Luxburg (2018) and Chen and Qin (2010). The test proposed in Ginestet et al. (2017) was not included due to its asymptotic nature and decreased power in larger dimensions. The tests were evaluated by estimating the power when the alternative is true and the null rejection rate. The number of nodes (n) and sample sizes (m) were varied, with significance level set at a = 0.05. The results showed that the proposed tests performed well in different scenarios, with the power and null rejection rate being analyzed for 5000 separate samples of networks from the underlying distributions. Additional simulations were conducted for weighted networks, networks from an exponential random graph model, and multiple-sample testing. The results were detailed in the supplementary material.

Accepted Answer

In dense networks, SPE-MNBS and SPE-SBM are close to the nominal level a = 0.05 under H 0 and both achieve good power under H 1. This indicates that SPE-MNBS and SPE-SBM are robust to different network structures and sample sizes, making them suitable for analyzing dense networks. The results show that these plug-in estimates of the link probability matrices yield superior results for two-sample testing compared to DFRO and VEC. Additionally, SPE-AVG also shows significant improvement as the sample size increases, making it a powerful method for dense networks.

Accepted Answer

A graphon structure is a function f that maps a link probability matrix P and nodes number n to a function f : [0, 1]x[0, 1] - [0, 1 ] and a set of independent and identically distributed random variables x i ~ Uniform[0, 1], such that P ij = f (x i , x j ). It is used in hierarchical clustering and link probability estimation. Graphon f is defined by Zhang et al. (2017) as f (u, v) = (u 2 + v 2 )/3 cos{1/(u 2 + v 2 )} + 0.15. It is used to generate probability matrices for simulations in network analysis.

Accepted Answer

The correlated Erdös-Rényi (ER) model, as studied in the third experiment, examines the robustness of four tests to dependency. Two independent ER networks, A1 ~ ER(n, p1) and A2 ~ ER(n, p2), are sampled. The parameter e3 is used to generate correlated edge sets in A1, resulting in m1 networks that are marginally ER(n, p1*e3). Similarly, A2 is generated with parameter e4. Under the null hypothesis, p2 = p1 = 0.9, while p2 = 0.83 under the alternative hypothesis. The results show that DFRO has consistently high power in the alternative setting for the entire range of n, which is only matched for the tests as n increases. This demonstrates the impact of the correlated ER model on the robustness of the tests.

Accepted Answer

In the null setting, where we test networks from the same group, SPE-AVG and SPE-SBM have low rejection rates, indicating that they fail to reject the null hypothesis. However, in the alternative setting, where we test networks from different groups, all three methods correctly reject the null hypothesis with high rejection rates. This suggests that the underlying distributions of the different groups are indeed different. In the in silico experiment, where we subsample within one group, SPE-AVG and SPE-SBM have inflated rejection rates, while WRG has a lower rejection rate. This indicates that the null rejection rates are influenced by the heterogeneity within a class, leading to potential false positives. Overall, the results show that SPE-AVG and SPE-SBM perform better than WRG in distinguishing between different groups, but caution should be exercised due to the potential for inflated rejection rates in certain settings.

Accepted Answer

Binarizing weights can improve the signal-to-noise ratio by reducing the impact of spurious correlations and heterogeneity within classes. In the context of binary tests, weights are binarized based on a threshold, which directly relates to the density of the networks. By setting the adjacency matrix entries to 1 when the absolute values of the weights exceed the threshold and 0 otherwise, the network becomes more sparse, resulting in a higher rejection rate in the null setting. This process helps mitigate the influence of noise and enhances the detection of meaningful differences in the alternative setting. For example, in the COBRE and MB networks, high power is observed for a wide range of threshold values, indicating a better signal-to-noise ratio. However, for the StarPlus networks, different thresholds provide varying balances between signal and noise for different models. Overall, binarizing weights can provide practitioners with a way to understand and optimize the signal-to-noise ratio of their edge weights.

A Spectral-Based Framework for Hypothesis Testing in Populations of Networks

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What is the proposed spectral-based testing statistic for populations of networks and its asymptotic null distribution?

2. What is the goal of the proposed spectral-based test for binary networks?

3. What is the normalized matrix in binary network testing?

4. What is the alternative test statistic based on estimated link probability matrices?

5. What are the three methods for estimating link probability matrices?

6. What is the asymptotic power guarantee?

7. What is the extension to weighted networks in the context of weighted networks?

8. What is the power of the test for weighted networks?

9. What is the goal of testing differences in distributions?

10. How do proposed tests compare in simulation studies?

11. How do SPE-MNBS and SPE-SBM perform in dense networks?

12. What is a graphon structure?

13. How does the correlated Erdös-Rényi model affect robustness tests?

14. How do null rejection rates compare for SPE-AVG, SPE-SBM, and WRG in different settings?

15. How can binarizing weights improve signal-to-noise ratio?

Citations

Network Recovery from Unlabeled Noisy Samples

References

On Spectral Clustering: Analysis and an algorithm

Functional network organization of the human brain

Spectral Analysis of Large Dimensional Random Matrices

A Shrinkage Approach to Large-Scale Covariance Matrix Estimation and Implications for Functional Genomics

Large Networks and Graph Limits

Related Papers (5)

Introduction to Neyman and Pearson (1933) On the Problem of the Most Efficient Tests of Statistical Hypotheses

Significance Testing in CFA

Analysis and importance of statistical power and sample size in empirical scientific research

Statistical Analysis in DEA

A Primer for Statistical Analysis