Co-occurrence Networks with Nestimate

Introduction

Co-occurrence networks analyze binary indicator data—situations where multiple states can be active (1) or inactive (0) simultaneously. Co-occurrence networks differ from transition networks in that the former methods capture contemporaneous relationships (i.e., which states tend to occur together?) whereas the latter capture sequences (i.e., which states occur after one another?)

Nestimate provides two methods for binary data:

Simple Co-occurrence networks (method = "co_occurrence" or "cna") count how often pairs of states are both active at the same time point
Ising networks (method = "ising") use L1-regularized logistic regression to estimate conditional dependencies, producing sparse networks

This vignette demonstrates both methods using the learning_activities dataset—binary indicators of 6 learning activities across 200 students and 30 time points.

Data

The learning_activities dataset contains 6,000 observations (200 students x 30 time points). At each time point, each of 6 learning activities is either active (1) or inactive (0).

library(Nestimate)
data(learning_activities)
head(learning_activities, 10)
#>    student Reading Video Forum Quiz Coding Review
#> 1        1       1     0     0    1      0      0
#> 2        1       1     0     0    1      1      0
#> 3        1       1     0     0    0      1      1
#> 4        1       1     0     0    1      1      1
#> 5        1       1     1     0    0      1      0
#> 6        1       1     1     1    0      1      1
#> 7        1       1     1     1    0      1      1
#> 8        1       1     1     1    0      1      1
#> 9        1       1     1     1    0      1      1
#> 10       1       0     1     1    0      1      1

The 6 activities are:

activities <- c("Reading", "Video", "Forum", "Quiz", "Coding", "Review")
colSums(learning_activities[, activities])
#> Reading   Video   Forum    Quiz  Coding  Review 
#>    2542    2622    2424    2273    2488    2594

Multiple activities can be active simultaneously:

# How many activities are active at each time point?
n_active <- rowSums(learning_activities[, activities])
table(n_active)
#> n_active
#>    1    2    3    4    5    6 
#> 1367 1766 1712  884  254   17

Co-occurrence Network

Co-occurrence networks count how often pairs of states are both active at the same time point. An edge between A and B indicates they frequently co-occur.

Basic Co-occurrence

net_cna <- build_network(learning_activities,
                         method = "co_occurrence",
                         codes = activities,
                         actor = "student")
net_cna
#> Co-occurrence Network [undirected]
#>   Weights: [2681.000, 3290.000]  |  mean: 3047.333
#> 
#>   Weight matrix:
#>           Reading Video Forum Quiz Coding Review
#>   Reading    6100  3169  3211 2891   3138   3113
#>   Video      3169  6262  3183 2903   3278   3290
#>   Forum      3211  3183  5692 2942   2958   3045
#>   Quiz       2891  2903  2942 5379   2681   2725
#>   Coding     3138  3278  2958 2681   5886   3183
#>   Review     3113  3290  3045 2725   3183   6186

Interpretation: Edge weights represent raw co-occurrence counts summed across all students and time points. Higher weights mean those activities frequently happen together.

Normalized Co-occurrence

Raw counts can be misleading—frequent activities will have high co-occurrence simply because they’re common. Normalizing by expected co-occurrence (under independence) reveals associations beyond base rates.

# View the co-occurrence matrix
round(net_cna$weights, 0)
#>         Reading Video Forum Quiz Coding Review
#> Reading    6100  3169  3211 2891   3138   3113
#> Video      3169  6262  3183 2903   3278   3290
#> Forum      3211  3183  5692 2942   2958   3045
#> Quiz       2891  2903  2942 5379   2681   2725
#> Coding     3138  3278  2958 2681   5886   3183
#> Review     3113  3290  3045 2725   3183   6186

The diagonal shows how often each activity occurs (self-co-occurrence = frequency).

Windowed Co-occurrence

You can aggregate across time windows before computing co-occurrence. This captures activities that occur in the same temporal neighborhood, not just the exact same time point:

net_cna_windowed <- build_network(learning_activities,
                                   method = "co_occurrence",
                                   codes = activities,
                                   actor = "student",
                                   window_size = 10)
net_cna_windowed
#> Co-occurrence Network [undirected]
#>   Weights: [9096.000, 11181.000]  |  mean: 10230.667
#> 
#>   Weight matrix:
#>           Reading Video Forum  Quiz Coding Review
#>   Reading   16894 10639 10603  9735  10493  10699
#>   Video     10639 17012 10511  9710  10981  11181
#>   Forum     10603 10511 15090  9777   9992  10353
#>   Quiz       9735  9710  9777 14495   9123   9096
#>   Coding    10493 10981  9992  9123  15956  10567
#>   Review    10699 11181 10353  9096  10567  16696

Larger windows capture broader temporal associations at the cost of temporal precision.

Ising Network

Ising networks use L1-regularized logistic regression to estimate conditional dependencies between binary variables. Each variable is regressed on all others, and the resulting coefficients form the network edges.

Key advantages over simple co-occurrence:

Sparsity: L1 regularization shrinks weak associations to exactly zero
Conditional: Edges represent direct relationships controlling for other variables
Interpretable: Edge weights reflect log-odds of co-activation

Basic Ising Network

# Aggregate to student-level summaries for Ising (requires cross-sectional data)
student_summary <- aggregate(learning_activities[, activities],
                              by = list(student = learning_activities$student),
                              FUN = function(x) as.integer(mean(x) > 0.5))
student_summary <- student_summary[, -1]  # Remove student column

net_ising <- build_network(student_summary,
                           method = "ising",
                           params = list(gamma = 0.25))
net_ising
#> Ising Model Network [undirected]
#>   Sample size: 200
#> 
#>   Weight matrix:
#>           Reading Video Forum Quiz Coding Review
#>   Reading       0     0     0    0      0      0
#>   Video         0     0     0    0      0      0
#>   Forum         0     0     0    0      0      0
#>   Quiz          0     0     0    0      0      0
#>   Coding        0     0     0    0      0      0
#>   Review        0     0     0    0      0      0 
#> 
#>   Gamma: 0.25  |  Rule: AND
#>   Thresholds: [-0.895, -0.532]

Ising Parameters

The gamma parameter controls sparsity via EBIC model selection:

gamma = 0: Less sparse (BIC-like selection)
gamma = 0.25: Moderate sparsity (default)
gamma = 0.5: More sparse

net_ising_sparse <- build_network(student_summary,
                                   method = "ising",
                                   params = list(gamma = 0.5))

# Compare edge counts
cat("Gamma = 0.25:", sum(net_ising$weights != 0), "edges\n")
#> Gamma = 0.25: 0 edges
cat("Gamma = 0.50:", sum(net_ising_sparse$weights != 0), "edges\n")
#> Gamma = 0.50: 0 edges

Symmetrization Rules

Ising estimation produces asymmetric coefficients (A predicting B may differ from B predicting A). The rule parameter controls symmetrization:

"AND" (default): Keep edge only if both directions are non-zero
"OR": Keep edge if either direction is non-zero

net_and <- build_network(student_summary, method = "ising",
                         params = list(gamma = 0.25, rule = "AND"))
net_or <- build_network(student_summary, method = "ising",
                        params = list(gamma = 0.25, rule = "OR"))

cat("AND rule edges:", sum(net_and$weights != 0), "\n")
#> AND rule edges: 0
cat("OR rule edges:", sum(net_or$weights != 0), "\n")
#> OR rule edges: 0

"AND" is more conservative; "OR" retains more edges.