TraMineR工具：预测新离散状态序列的类别归属

Got it, let's walk through how to assign new discrete state sequences to existing clusters using TraMineR, building on your precomputed dissimilarity matrix D and clustering results. Here's a practical, step-by-step breakdown with R code examples:

1. Align New Sequences with Training Data Structure #

First, you need to make sure your new sequences match the training data's alphabet, state labels, and encoding exactly—otherwise your distance calculations will be meaningless. Use seqdef to standardize them:

# Assume your preprocessed training sequence object is named train.seq
new_seq_raw <- # Your new sequence data (e.g., a data frame/vector of state strings)
new.seq <- seqdef(
  data = new_seq_raw,
  alphabet = alphabet(train.seq),  # Reuse the training set's state alphabet
  labels = labels(train.seq),      # Match state labels from training
  states = states(train.seq),      # Keep state encoding consistent
  missing = NA                     # Handle missing values the same way as training
)

2. Calculate Distances from New Sequences to Training Data #

Compute the dissimilarity between each new sequence and every sequence in your training set. Critical: Use the exact same dissimilarity method and parameters you used to create matrix D (e.g., Optimal Matching with your cost matrix, LCS, Hamming distance):

# Example: If you used Optimal Matching (OM) with a substitution cost matrix 'cost.mat'
new_distances <- seqdist(
  seq1 = new.seq,
  seq2 = train.seq,
  method = "OM",  # Match your original method (e.g., "LCS", "HAM")
  sm = cost.mat   # Reuse your original substitution cost matrix (if applicable)
)

# new_distances will be a 1 x N matrix (N = number of training sequences)

3. Predict Cluster Membership #

The method depends on the type of clustering you ran on your training data:

Case 1: K-Means Clustering #

If you used seqclust(method = "kmeans"), your result object includes cluster centroids. Calculate the distance from the new sequence to each centroid, then assign it to the closest cluster:

# Assume your k-means clustering result is stored in km_clust
km_clust <- seqclust(train.seq, method = "kmeans", diss = D, k = 3)

# Calculate distance from new sequence to each centroid
centroid_distances <- seqdist(
  seq1 = new.seq,
  seq2 = km_clust$centers,
  method = "OM",
  sm = cost.mat
)

# Assign to the cluster with the smallest distance
predicted_cluster <- which.min(centroid_distances)

Case 2: Hierarchical Clustering #

For hierarchical clustering (e.g., agnes or diana paired with seqdist), a simple k-Nearest Neighbors (k-NN) vote works best:

1. Find the k closest training sequences to your new sequence
2. Assign the most common cluster label among those neighbors

# Assume your training cluster labels are stored in train_clusters (from cutree())
train_clusters <- cutree(hierarchical_clust_obj, k = 3)

# Choose k (e.g., 5 nearest neighbors—adjust based on your data size)
k <- 5

# Get indices of the k closest training sequences
nearest_indices <- order(new_distances)[1:k]

# Pull their cluster labels
nearest_clusters <- train_clusters[nearest_indices]

# Vote for the most frequent cluster
predicted_cluster <- names(sort(table(nearest_clusters), decreasing = TRUE))[1]

Quick Pro Tips #

• Always verify state encoding consistency (e.g., if "S" = Single in training, don’t use "Single" in new data).
• For large training sets, k-NN with a small k (3-7) is efficient and avoids recomputing the full dissimilarity matrix.
• If you want to avoid k-NN for hierarchical clustering, you could append the new sequence to training data and re-run clustering—but this is computationally heavy for big datasets.

内容的提问来源于stack exchange，提问作者Maxim.K