1  {
 2    "ADBench": {
 3      "paper": {"id": "adbench", "short": "Han et al., NeurIPS 2022"},
 4      "scope": "tabular",
 5      "n_datasets": 57,
 6      "n_algorithms": 30,
 7      "rankings": {
 8        "overall_top_5": ["ECOD", "IForest", "KNN", "COPOD", "HBOS"],
 9        "high_dim_top_3": ["ECOD", "COPOD", "IForest"],
10        "low_dim_top_3": ["KNN", "LOF", "CBLOF"]
11      },
12      "key_finding": "No single algorithm dominates; ensemble of top-5 is robust"
13    },
14    "NLP_ADBench": {
15      "paper": {"id": "nlp_adbench", "short": "Li et al., EMNLP 2025"},
16      "scope": "text",
17      "n_datasets": 8,
18      "n_algorithms": 19,
19      "rankings": {
20        "overall_top_5": ["OpenAI+LUNAR", "OpenAI+LOF", "OpenAI+AE", "MiniLM+KNN", "BERT+LOF"]
21      },
22      "key_finding": "Embedding quality >> detector choice; two-step beats end-to-end"
23    },
24    "TSB_AD": {
25      "paper": {"id": "tsb_ad", "short": "Liu & Paparrizos, NeurIPS 2024"},
26      "scope": "time_series",
27      "n_datasets": 1070,
28      "n_algorithms": 40,
29      "rankings": {
30        "overall_top_5": ["IForest", "LOF", "POLY", "KNN", "KShapeAD"],
31        "subsequence_top_3": ["MatrixProfile", "SAND", "Series2Graph"]
32      },
33      "key_finding": "Classical methods competitive with deep; MatrixProfile strong on subsequence anomalies"
34    },
35    "BOND": {
36      "paper": {"id": "liu2022bond", "short": "Liu et al., NeurIPS 2022"},
37      "scope": "graph",
38      "n_datasets": 14,
39      "n_algorithms": 14,
40      "rankings": {
41        "deep_top_3": ["DOMINANT", "CoLA", "CONAD"],
42        "classical_top_2": ["Radar", "ANOMALOUS"]
43      },
44      "key_finding": "DOMINANT and CoLA are most reliable deep methods; classical MF methods competitive on small graphs"
45    }
46  }