Social-Space User Latent Factors

• Social-space user latent factors are abstract, unobserved variables that reveal hidden social network structures, including community memberships and local similarities.
• The GLSSBM integrates latent space modeling with stochastic blockmodels to capture both discrete community assignments and continuous positional differences.
• Bayesian inference using MCMC enables robust estimation of latent factors, providing insights into group dynamics and individual activity within complex networks.

Social-space user latent factors are abstract, unobserved variables that characterize the hidden structure governing social network interactions. These factors underlie observed patterns such as friendship formation, clustering, homophily, and community structure. The modeling of such latent factors has progressed from classical latent space models in Euclidean settings to recent developments in generalized frameworks—most notably the generalized latent space stochastic blockmodel (GLSSBM)—to enable more nuanced representations of social systems and more robust inference of relational patterns.

1. Generalized Latent Space Stochastic Blockmodel (GLSSBM): Structure and Formulation

The GLSSBM combines the principles of latent space models and stochastic blockmodels to model network data. Each actor (or node) is associated with:

• A discrete block membership variable, γᵢ ∈ {1, …, K}, indicating class or “community” assignment.
• A continuous latent position Zᵢ ∈ ℝᵈ representing their location in an unobserved latent space.

The probability of an edge Yᵢⱼ between two nodes depends on both their block memberships and latent positions:

• For nodes within the same block (γᵢ = γⱼ = k):

$$\text{logit}~P(Y_{ij} = 1 \mid \gamma_i = \gamma_j = k, Z_i, Z_j) = \beta_k - \|Z_i - Z_j\|$$

where β_k is a block-specific intercept.

• For nodes in different blocks (γᵢ = k, γⱼ = l, k ≠ l), the log-odds are governed by block-to-block parameters and individual sender/receiver effects:

$$\text{logit}~P(Y_{ij} = 1 \mid \gamma_i = k, \gamma_j = l, s_i, r_j) = \eta_{kl} + s_i + r_j$$

where η_{kl} encodes block interactions and s_i, r_j represent individual-level effects.

This formulation enables the model to simultaneously encode transitivity, reciprocity, clustering, assortativity, and disassortative mixing.

2. Inference Methodology and Model Selection

Inference in the GLSSBM is performed within a fully Bayesian framework.

• Priors:
  • Latent positions: Zᵢ ∼ MVN(0, (σ^z)^2 I_d)
  • Block-specific parameters: Independent normal priors for βk, η{kl}
  • Sender/receiver effects: Independent normal priors
• Estimation:
  • Markov Chain Monte Carlo (MCMC) sampling is used, updating block assignments γᵢ, latent positions Zᵢ, and associated parameters. To facilitate sampling, data augmentation methods are used, such as Polya–Gamma latent variables, which yield closed-form updates for certain conditional distributions.
  • The likelihood is the product over all unordered dyads, accounting for block membership and latent space distances.
• Model Selection:
  • Competing models of varying number of classes (K) and latent space dimension (d) are compared using the Watanabe–Akaike Information Criterion (WAIC), balancing fit and complexity.

This approach provides a principled mechanism for uncovering latent structure, as well as for quantifying uncertainty in class assignments and latent position estimates.

3. Interpretation of Latent Factors: Community and Position

Within GLSSBM, latent user factors are dual:

• Block Membership (γᵢ): Encodes discrete community or party affinity, capturing the equivalence and clustering typical in political or organizational social systems.
• Latent Position (Zᵢ): Quantifies continuous variation—possibly ideological proximity, political centrality, or social similarity—within (and across) blocks.

Sender and receiver random effects further account for heterogeneity in node activity, such as propensity to follow or be followed, beyond what is explained by position or community.

This structure allows for nuanced explanations of social phenomena: nodes with similar block membership but separated in latent space may share broad attributes but differ in finer behavioral or ideological dimensions. Conversely, block-to-block parameters (η_{kl}) can highlight alliances, antagonism, or asymmetric relationships between groups.

4. Empirical Application: Irish Parliament Twitter Network

Applying the GLSSBM to the Twitter follower network among Irish parliamentarians (members of Dáil Éireann), the analysis revealed:

• Predominance of political party membership as the main driver of following relationships—as captured by recovered blocks corresponding to Fine Gael, Fianna Fáil, Labour, and other smaller or independent groups.
• Within-block modeling effectively reproduced patterns of high within-party connectivity (strong βk, low latent distances), while between-block interaction parameters η{kl} quantified the extent of inter-party following.
• Sender and receiver effects captured variations in individual activity both within and between parties.
• Strong correspondence was observed between model-inferred connection probabilities and similarity in voting behavior, indicating that latent social factors from the network are related to offline political alignment.

This application demonstrates the capacity of GLSSBM to recover and quantify both discrete and continuous user latent factors aligned with substantive political constructs.

5. Broader Implications for Social-Space User Latent Factors

The GLSSBM’s integration of block modeling with latent spatial positioning offers several advantages for understanding and interpreting user latent factors in social contexts:

• Clustering and Detection of Community Structure: Formal recovery of block structure is highly interpretable and aligns with known group divisions (parties, communities, organizations).
• Continuous Similarity and Local Structure: Latent positions uncover within-group differentiation and local proximity, revealing substructures and gradients invisible to block models alone.
• Multi-Level Effects: Sender/receiver components allow for individual idiosyncrasies, enabling more accurate representation and prediction of ties.
• Transferability: The modeling framework can be adapted to diverse social networks (beyond politics), such as organizational structures, online communities, or collaboration networks.

The approach is particularly well-suited to contexts where group membership and local similarity jointly drive network formation, and where finer gradations within communities are substantively meaningful.

6. Significance for Political and Social Science

For political and social scientists, the GLSSBM enables:

• Joint analysis of online behavioral patterns (such as social media following) and offline phenomena (such as legislative voting).
• Assessment of coalition dynamics, group cohesion, party discipline, and roles of independents.
• Quantitative comparison of observed and predicted ties, given latent factors, to test substantive theories of network formation and influence.

Model outputs—including posterior probabilities of edge existence, block assignments, and latent positions—provide interpretable metrics for exploratory and confirmatory research in the analysis of complex social-space user factors.

---

In summary, the generalized latent space stochastic blockmodel represents a significant methodological advance for the modeling of social-space user latent factors, seamlessly integrating discrete community structure with continuous latent similarity. Its Bayesian inference framework, statistical interpretability, and applicability to empirical political networks illustrate its utility for uncovering the multi-level structure underlying social interactions and behaviors (Ng et al., 2018).