Document Clustering on 20 Newsgroups
73AccuracyAgentic clustering
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Agentic clusteringk setting=Given k, predicted number of clusters (ˆk)=20, total cost ($)=$100 + $222026.05 | 73 | 67 | 55 | |
| Agentic clusteringk setting=Discover k, predicted number of clusters (ˆk)=18, total cost ($)=$100 + $232026.05 | 72 | 68 | 53 | |
| ClusterLLMk setting=Given k, predicted number of clusters (ˆk)=20, total cost ($)=$1.532026.05 | 67 | 64 | 48 | |
| LLM-embedding+k-meansk setting=Given k, predicted number of clusters (ˆk)=20, total cost ($)=$0.392026.05 | 64 | 63 | 49 | |
| SBERT+k-meansk setting=Given k, predicted number of clusters (ˆk)=20, total cost ($)=$02026.05 | 61 | 58 | 42 | |
| BERTopick setting=Given k, predicted number of clusters (ˆk)=20, total cost ($)=$02026.05 | 39 | 53 | 31 | |
| TopicGPTk setting=Discover k, predicted number of clusters (ˆk)=18, total cost ($)=$100 + $662026.05 | 39 | 53 | 28 | |
| k-means on mean poolingFeature representation=mean pooling, Backbone=frozen BERT, K1=32, K2=20, Epochs=152026.04 | 35.1 | 37.7 | 17.8 | |
| LDAk setting=Given k, predicted number of clusters (ˆk)=20, total cost ($)=$02026.05 | 34 | 36 | 19 | |
| Huang & Hek setting=Discover k, predicted number of clusters (ˆk)=377, total cost ($)=$100 + $442026.05 | 29 | 53 | 22 | |
| k-means on [CLS]Feature representation=[CLS], Backbone=frozen BERT, K1=32, K2=20, Epochs=152026.04 | 20 | 21.1 | 6 | |
| BERTopick setting=Discover k, predicted number of clusters (ˆk)=272, total cost ($)=$02026.05 | 20 | 51 | 17 | |
| k-means on CLSBackbone=frozen BERT, epsilon=0.052026.04 | 19.6 | 18.9 | 6.5 | |
| DDCL-AttentionBackbone=frozen BERT, epsilon=0.052026.04 | 17.5 | 15.2 | 3.9 | |
| Hier. DDCL L1+L2ε=0.05, λ=1.5, Backbone=frozen BERT, K1=32, K2=20, Epochs=152026.04 | 13.3 | 9.3 | 1.6 | |
| Hier. DDCL L1+L2ε=0.1, λ=0.5, Backbone=frozen BERT, K1=32, K2=20, Epochs=152026.04 | 11.2 | 7.5 | 0.9 |