Skip to contents

Introduction

Clusters represent subgroups within the data that share similar patterns. Such patterns may reflect similar temporal dynamics (when we are analyzing sequence data, for example) or relationships between variables (as is the case in psychological networks). Units within the same cluster are more similar to each other, while units in different clusters differ more substantially. In this vignette, we demonstrate how to perform clustering on sequence data using Nestimate.

Data

To illustrate clustering, we will use the human_cat dataset, which contains 10,796 coded human interactions from 429 human-AI pair from programming sessions across 34 projects classified into 9 behavioral categories. Each row represents a single interaction event with a timestamp, session identifier, and category label.

library(Nestimate)
data("human_cat")
head(human_cat)
#>        id   project   session_id                timestamp session_date actor
#> 5094 3439 Project_7 0086cabebd15 2026-03-05T11:32:52.057Z   2026-03-05 Human
#> 5095 3439 Project_7 0086cabebd15 2026-03-05T11:32:52.057Z   2026-03-05 Human
#> 5096 3439 Project_7 0086cabebd15 2026-03-05T11:32:52.057Z   2026-03-05 Human
#> 5097 3440 Project_7 0086cabebd15 2026-03-05T11:32:52.068Z   2026-03-05 Human
#> 5100 3442 Project_7 0086cabebd15 2026-03-05T11:39:19.098Z   2026-03-05 Human
#> 5103 3444 Project_7 0086cabebd15 2026-03-05T11:41:55.500Z   2026-03-05 Human
#>               code  category    superclass
#> 5094       Context   Specify     Directive
#> 5095        Direct   Command     Directive
#> 5096 Specification   Specify     Directive
#> 5097     Interrupt Interrupt Metacognitive
#> 5100  Verification    Verify    Evaluative
#> 5103 Specification   Specify     Directive

We can build a transition network using this dataset using build_network. We need to determine the actor (session_id), the action (session_id), and the time (timestamp). We will use the overall network object as the starting point to find subgroups since it structures the raw data into the appropriate units of analysis to perform clustering.

net <- build_network(human_cat, 
                     method = "tna",
                     action = "category", 
                     actor = "session_id",
                     time = "timestamp")
#> Metadata aggregated per session: ties resolved by first occurrence in 'code' (628 sessions), 'superclass' (158 sessions)

Dissimilarity-based Clustering

Dissimilarity-based clustering groups units of analysis (in our case, sessions, since that is what we provided as actor) by directly comparing their observed sequences. In our case, each session is represented by its sequence of actions, and similarity between sessions is defined using a distance metric that quantifies how different two sequences are.

To implement this method using Nestimate, we can use the cluster_data() function, which takes either raw sequence data or a network object such as the net object that we estimated (which also contains the original sequences in $data):

clust <- cluster_data(net, k = 3)

clust
#> Sequence Clustering
#>   Method:        pam 
#>   Dissimilarity: hamming  
#>   Clusters:      3 
#>   Silhouette:    0.1305 
#>   Cluster sizes: 302, 753, 380 
#>   Medoids:       1429, 80, 128

The default clustering mechanism uses Hamming distance (number of positions where sequences differ) with PAM (Partitioning Around Medoids).

The result contains the cluster assignments (which cluster each session belongs to), the cluster sizes, and a silhouette score that reflects the quality of the clustering (higher values indicate better separation between clusters), among other useful information.

# Cluster assignments (first 20 sessions)
head(clust$assignments, 20)
#>  [1] 1 2 3 1 2 2 2 1 2 2 3 2 3 2 2 3 2 2 2 3

# Cluster sizes
clust$sizes
#>   1   2   3 
#> 302 753 380

# Silhouette score (clustering quality: higher is better)
clust$silhouette
#> [1] 0.1305163

Visualizing Clusters

The silhouette plot shows how well each sequence fits its assigned cluster. Values near 1 indicate good fit; values near 0 suggest the sequence is between clusters; negative values indicate possible misclassification.

plot(clust, type = "silhouette")

Silhouette plot showing cluster quality The MDS (multidimensional scaling) plot projects the distance matrix to 2D, showing cluster separation.

plot(clust, type = "mds")

MDS plot showing cluster separation

Distance Metrics

A distance metric defines how (dis)similarity between sequences is measured. In other words, it quantifies how different two sequences are from each other. Nestimate currently supports 9 distance metrics for comparing sequences:

Metric Description Best for
hamming Positions where sequences differ Equal-length sequences
lv Levenshtein (edit distance) Variable-length, insertions/deletions
osa Optimal string alignment Edit distance + transpositions
dl Damerau-Levenshtein Full edit + adjacent transpositions
lcs Longest common subsequence Preserving order, ignoring gaps
qgram Q-gram frequency difference Pattern-based similarity
cosine Cosine of q-gram vectors Normalized pattern similarity
jaccard Jaccard index of q-grams Set-based pattern overlap
jw Jaro-Winkler Short strings, typo detection

Different metrics may produce different clustering results. You need to choose this based on your research question:

  • Hamming: compares sequences position by position (best for aligned sequences of equal length).
  • Edit distances (lv, osa, dl): allow insertions and deletions (best when sequences may be shifted or vary in length).
  • LCS: captures shared subsequences (best when overall patterns matter more than exact alignment).

We can specify which distance metric we want to use through the dissimilarity argument:

# Levenshtein distance (allows insertions/deletions)
clust_lv <- cluster_data(human_wide, k = 3, dissimilarity = "lv")
clust_lv$silhouette
#> [1] 0.24561

# Longest common subsequence
clust_lcs <- cluster_data(human_wide, k = 3, dissimilarity = "lcs")
clust_lcs$silhouette
#> [1] 0.0339432

Some distance metrics may take additional arguments. For example, the Hamming distance accepts temporal weighting to emphasize earlier or later positions:

# Emphasize earlier positions (higher lambda = faster decay)
clust_weighted <- cluster_data(net, 
                               k = 3,
                               dissimilarity = "hamming",
                               weighted = TRUE,
                               lambda = 0.5)
clust_weighted$silhouette
#> [1] 0.2650696

Clustering Methods

By default, Nestimate uses PAM (Partitioning Around Medoids) to form clusters, which assigns each sequence to the cluster represented by the most central sequence (the medoid). Besides PAM, Nestimate supports hierarchical clustering methods, which build clusters step by step by progressively merging similar units into a tree-like structure (a dendrogram):

  • ward.D2 (“Ward’s Method, Squared Distances”): Minimizes the increase in within-cluster variance using squared distances. Typically produces compact, well-separated clusters.
  • ward.D (“Ward’s Method”): An alternative implementation of Ward’s approach using a different distance formulation. Similar behavior, but results may vary slightly.
  • complete (“Complete Linkage”): Defines the distance between clusters as the maximum distance between their members. Produces tight, compact clusters.
  • average (“Average Linkage”): Uses the average distance between all pairs of points across clusters. Provides a balance between compactness and flexibility.
  • single (“Single Linkage”): Uses the minimum distance between points in two clusters. Can capture chain-like structures but may lead to - loosely connected clusters.
  • mcquitty (“McQuitty’s Method” / “WPGMA”): A weighted version of average linkage that gives equal weight to clusters regardless of size.
  • centroid (“Centroid Linkage”): Defines cluster distance based on the distance between cluster centroids (means). Can produce intuitive groupings but may introduce inconsistencies in the hierarchy.

To use any of these methods instead of PAM, we need to provide the method argument to cluster_data.

# Ward's method (minimizes within-cluster variance)
clust_ward <- cluster_data(net, k = 3, method = "ward.D2")
clust_ward$silhouette
#> [1] 0.533638

# Complete linkage
clust_complete <- cluster_data(net, k = 3, method = "complete")
clust_complete$silhouette
#> [1] 0.9149815

Choosing k (Number of Clusters)

To choose the right clustering solution and method, we need to compare the silhouette scores across different k values and clustering methods (and also distance metrics if we want):

methods <- c("pam", "ward.D2", "ward.D", "complete",
             "average", "single", "mcquitty", "median", "centroid")

silhouettes <- lapply(methods, function(m) {
  sapply(2:6, function(k) {
    cluster_data(net, k = k, method = m, seed = 42)$silhouette
  })
})

names(silhouettes) <- methods

silhouettes
#> $pam
#> [1]  0.1758288  0.1305163  0.1683305 -0.1164963  0.1586104
#> 
#> $ward.D2
#> [1] 0.8455357 0.5336380 0.5353121 0.5354583 0.4860168
#> 
#> $ward.D
#> [1] 0.5520351 0.5028652 0.1537042 0.1595718 0.1746830
#> 
#> $complete
#> [1] 0.9315712 0.9149815 0.8960074 0.8761943 0.8685024
#> 
#> $average
#> [1] 0.9315712 0.9149815 0.8960074 0.8823133 0.8685024
#> 
#> $single
#> [1] 0.9315712 0.9149815 0.8960074 0.8823133 0.8685024
#> 
#> $mcquitty
#> [1] 0.9315712 0.9149815 0.8960074 0.8823133 0.8685024
#> 
#> $median
#> [1] 0.9315712 0.9149815 0.8960074 0.8823133 0.8685024
#> 
#> $centroid
#> [1] 0.9315712 0.9149815 0.8960074 0.8823133 0.8685024
methods <- names(silhouettes)
colors <- rainbow(length(methods))

plot(2:6, silhouettes[[1]], type = "b", pch = 19, col = colors[1],
     xlab = "Number of clusters (k)",
     ylab = "Average silhouette width",
     ylim = c(0, 1),
     main = "Choosing k")

for (i in 2:length(methods)) {
  lines(2:6, silhouettes[[i]], type = "b", pch = 19, col = colors[i])
}

legend("topright", legend = methods, col = colors, lty = 1, pch = 19)

Silhouette scores across different k values

Higher silhouette scores indicate better-defined clusters. Look for an “elbow” or maximum. In our case, the best-performing cluster method is centroid, for k = 2. However, if we inspect the results of using this method, the cluster sizes are really unbalanced, since one cluster only contains one sequence.

clust_centroid_2 <- cluster_data(net, k = 2, method = "centroid", seed = 42)
summary(clust_centroid_2)
#> Sequence Clustering Summary
#>   Method:        centroid 
#>   Dissimilarity: hamming 
#>   Silhouette:    0.9316 
#> 
#> Per-cluster statistics:
#>  cluster size mean_within_dist
#>        1 1434         10.64515
#>        2    1          0.00000

A more balanced option seems to be using ward.D2, also with 2 clusters, which yields a reasonable silhouette width (0.5-0.7).

clust <- cluster_data(net, k = 2, method = "ward.D2", seed = 42)

summary(clust)
#> Sequence Clustering Summary
#>   Method:        ward.D2 
#>   Dissimilarity: hamming 
#>   Silhouette:    0.8455 
#> 
#> Per-cluster statistics:
#>  cluster size mean_within_dist
#>        1 1411         8.972509
#>        2   24        72.601449

Mixture Markov Models

Instead of clustering sequences based on how similar they are to one another, we can cluster them together based on their transition dynamics. Mixture Markov models (MMM) fit separate Markov models, and sequences are assigned to the cluster whose transition structure best matches their observed behavior.

To implement MMM, we can use the build_mmm provided by Nestimate, and we pass the sequence data or network estimated and the number of clusters (k, by default 2)

mmm_default <- build_mmm(net)

We can inspect the results using summary and obtain the cluster assignment from the results using mmm_default$assignments.

summary(mmm_default)
#> Mixed Markov Model
#>   k = 2 | 1435 sequences | 9 states
#>   LL = -20355.7 | BIC = 41881.8 | ICL = 42095.0
#> 
#>   Cluster  Size  Mix%%   AvePP
#>   ------------------------------
#>         1   340  26.1%  0.919
#>         2  1095  73.9%  0.943
#> 
#>   Overall AvePP = 0.937 | Entropy = 0.222 | Class.Err = 0.0%
#> 
#> --- Cluster 1 (26.1%, n=340) ---
#>           Command Correct Frustrate Inquire Interrupt Refine Request Specify
#> Command     0.066   0.020     0.017   0.006     0.009  0.015   0.382   0.434
#> Correct     0.058   0.071     0.096   0.044     0.150  0.127   0.060   0.290
#> Frustrate   0.098   0.098     0.160   0.074     0.003  0.222   0.180   0.117
#> Inquire     0.162   0.165     0.203   0.202     0.015  0.054   0.103   0.084
#> Interrupt   0.203   0.095     0.103   0.070     0.199  0.082   0.022   0.112
#> Refine      0.042   0.093     0.055   0.074     0.054  0.084   0.143   0.443
#> Request     0.032   0.004     0.032   0.000     0.031  0.010   0.010   0.870
#> Specify     0.463   0.012     0.013   0.014     0.352  0.017   0.028   0.073
#> Verify      0.215   0.027     0.189   0.154     0.014  0.175   0.093   0.057
#>           Verify
#> Command    0.052
#> Correct    0.103
#> Frustrate  0.049
#> Inquire    0.011
#> Interrupt  0.114
#> Refine     0.013
#> Request    0.011
#> Specify    0.028
#> Verify     0.076
#> 
#> --- Cluster 2 (73.9%, n=1095) ---
#>           Command Correct Frustrate Inquire Interrupt Refine Request Specify
#> Command     0.255   0.114     0.064   0.077     0.040  0.047   0.090   0.264
#> Correct     0.088   0.096     0.145   0.052     0.037  0.114   0.124   0.291
#> Frustrate   0.098   0.122     0.175   0.069     0.050  0.166   0.101   0.132
#> Inquire     0.193   0.135     0.090   0.187     0.085  0.067   0.084   0.110
#> Interrupt   0.266   0.096     0.097   0.143     0.035  0.093   0.094   0.155
#> Refine      0.055   0.073     0.071   0.040     0.030  0.087   0.151   0.474
#> Request     0.102   0.021     0.047   0.082     0.039  0.040   0.043   0.590
#> Specify     0.187   0.079     0.104   0.089     0.097  0.102   0.076   0.225
#> Verify      0.203   0.091     0.162   0.116     0.041  0.094   0.115   0.091
#>           Verify
#> Command    0.051
#> Correct    0.055
#> Frustrate  0.086
#> Inquire    0.050
#> Interrupt  0.021
#> Refine     0.019
#> Request    0.037
#> Specify    0.041
#> Verify     0.086
head(mmm_default$assignments,10)
#>  [1] 1 2 2 2 2 2 2 1 2 2

To decide which number of clusters best represents the structure of our data, we can use the compare_mmm function and provide again the net object and a range of k to run a full comparison.

mmm_comparison <- compare_mmm(net, k = 2:8)
mmm_comparison
#> MMM Model Comparison
#> 
#>  k log_likelihood AIC      BIC      ICL      AvePP     Entropy   converged
#>  2 -20355.74      41033.49 41881.79 42094.96 0.9370297 0.2221720  TRUE    
#>  3 -20152.03      40788.06 42063.14 42763.96 0.8034701 0.3962539  TRUE    
#>  4 -20026.82      40699.65 42401.51 43276.76 0.7659423 0.4252817 FALSE    
#>  5 -19926.77      40661.55 42790.19 43964.70 0.7085776 0.4411911  TRUE    
#>  6 -19813.15      40596.30 43151.73 44335.26 0.7071495 0.4148854  TRUE    
#>  7 -19752.65      40637.30 43619.50 44950.25 0.6838534 0.4236204 FALSE    
#>  8 -19660.85      40615.69 44024.68 45379.74 0.6746012 0.4241401 FALSE    
#>  best   
#>  <-- BIC
#>         
#>         
#>         
#>         
#>         
#> 

The results show that k = 2 is indeed the best clustering solution.

Building Networks per Cluster

Once sequences are clustered, we can create separate networks by cluster. We need to pass the clustering result to build_network and use the group argument to indicate that we want to group by cluster assignment.

cluster_net <- build_network(clust)

We may also compare which transition probabilities differ significantly among clusters using permutation testing:

comparison <- permutation_test(cluster_net)