Network Layout Algorithm with Covariate Smoothing

TL;DR

The paper tackles robust visualization of networks with observation errors by leveraging nodal covariates to inform edge probabilities. It introduces a model-based dyadic probability estimate $\hat{B}$ from covariates $X$ and observed adjacency $A$, and embeds this into a modified Fruchterman-Reingold energy $Q_2$ with a smoothing parameter $\gamma \in [0,1]$, allowing a continuum between observed edges and covariate-based probabilities. A tuning metric $\psi_{\gamma}$ based on standardized cross-terms $m_{\gamma}$ and edge-length changes $e_{\gamma}$ guides selection of $\gamma$, with validation on simulated networks (SBM and continuous covariates) and a real Add Health data application. Results show increased clustering and layout robustness when covariates strongly predict connections, and provide practical guidance on when to apply covariate smoothing versus standard FR, along with reproducibility resources.

Abstract

Network science explores intricate connections among objects, employed in diverse domains like social interactions, fraud detection, and disease spread. Visualization of networks facilitates conceptualizing research questions and forming scientific hypotheses. Networks, as mathematical high-dimensional objects, require dimensionality reduction for (planar) visualization. Visualizing empirical networks present additional challenges. They often contain false positive (spurious) and false negative (missing) edges. Traditional visualization methods don't account for errors in observation, potentially biasing interpretations. Moreover, contemporary network data includes rich nodal attributes. However, traditional methods neglect these attributes when computing node locations. Our visualization approach aims to leverage nodal attribute richness to compensate for network data limitations. We employ a statistical model estimating the probability of edge connections between nodes based on their covariates. We enhance the Fruchterman-Reingold algorithm to incorporate estimated dyad connection probabilities, allowing practitioners to balance reliance on observed versus estimated edges. We explore optimal smoothing levels, offering a natural way to include relevant nodal information in layouts. Results demonstrate the effectiveness of our method in achieving robust network visualization, providing insights for improved analysis.

Figures (6)

• Figure 1: Each two-row block corresponds to one of the three scenarios where nodal covariates have two categories (top), have five categories (middle), or are continuous (bottom). Each column represents odds: 1:1 (left), 1.5:1, 3.5:1 (right). For each scenario and odds combination, the visualization generated by our method is shown on top and the corresponding Fruchterman and Reingold (FR) visualization is shown on bottom. Each graph has 100 nodes, the selected $\gamma$ value for our method is presented above each graph.
• Figure 2: Choosing an appropriate value for the smoothing parameter $\gamma$. We utilize our selection metric described in Section \ref{['subsec:tuning']} a total of 100 times per graph size, type of nodal covariate, and odds. We select $\gamma$ as the average among the 100 samples and present the 95% confidence interval of the mean. We stratify our plots by graph size and data type. The number of nodes is a) 20, b) 50, c) 100, and d) 200 across the panels. The 5 Groups error bar is on the true odds while the other data types are offset.
• Figure 3: Impact of missing edges on graph layout discrepancy. We plot the average Procrustes value between the nodal coordinates of a graph with a fixed percentage of missing edges and the nodal coordinates of a graph with no missing edges. The nodal coordinates are considered minimally different if the value of the Procrustes distance is 0 and maximally different if the value is 1. The grey lines represents the 100 realizations contributing to the average (shown in black). Results are stratified by algorithm, our algorithm (left) and the Fruchterman-Reingold algorithm (right), and by odds: 1:1 (top), 1.5:1 (middle), and 3.5:1 (bottom). All plots represent a graph with a 100 nodes and one continuous nodal covariate sampled from the uniform distribution. The value of $\gamma$ is selected separately for each point in the graph. In general, our algorithm outperforms the FR algorithm when there are increased odds, whereas the outcome is reversed when the odds are even.
• Figure 4: Visualization of graphs with missing edges. We plot a graph with varying levels of missingness with our algorithm (top) and the FR algorithm (bottom). Proportion of missing edges ranges across the panels: a) 0%, b) 22.5%, c) 0.45%, d) 67.5%, and e) 90%. All plots represent a graph with a 100 nodes and one continuous nodal covariate sampled from the uniform distribution with odds set at 3.5:1 and $\gamma$ is selected individually for each plot. In this high odds setting, the robustness of our algorithm appears to outperform the FR algorithm.
• Figure 5: Visualization of network data from the Add Health study using our method with the full linkage model (top row) and the FR algorithm (bottom row). All pairwise linkage probabilities were estimated using logistic regression, and all covariates are 1 if two nodes share the category and 0 if not. We considered four covariates: a) $\tilde{sex}$, b) $\tilde{race}$, c) $\tilde{grade}$, and d) $\tilde{school}$. The coefficient estimates and standard errors are as follows: $\tilde{sex}$ 0.14 (0.062), $\tilde{race}$ 0.11 (0.079), $\tilde{grade}$ 1.99 (0.070), and $\tilde{school}$ 1.92 (0.163). We chose $\gamma$ = 0.672 using the procedure described in the paper. We colored the nodes in each panel by the values of the corresponding categorical covariate.