NFC: A Multi-Dimensional Root Cause Localization Dataset

NFC is a benchmark dataset for root cause localization on multi-dimensional data, built from anonymized NFC (Near-Field Communication) transaction records. Each dimension combination carries categorical attributes, and anomalies manifest as unexpected concentrations of records under specific attribute combinations. The task: given a set of anomalous records, identify the attribute combination (e.g., B=273 & D=42) that best explains them.

This dataset was used to evaluate Prism (PRecision-driven Iterative Search for Multi-dimensional root cause localization), a root cause localization algorithm that formulates the problem as a precision–recall joint optimization over attribute-selection weights.

Dataset Structure

File	Size	Description
df.csv	~146 MiB	Main data table: 5,135,932 records × 15 categorical dimensions
label.json	~6 MB	100 labeled test cases with ground-truth root causes

df.csv

• 15 anonymized categorical dimensions, named A through O.
• Values are integer-encoded; missing values are left empty (a dedicated missing-value category).
• 0-based row positions serve as record indices (the default RangeIndex after pd.read_csv).

Example:

A,B,C,D,E,F,G,H,I,J,K,L,M,N,O
3,3,,,3,3,3,,3,3,3,,,3,3
3,3,,,3,3,3,357,3,3,3,,,3,3

label.json

A JSON array of 100 test cases. Each case contains:

• label — the ground-truth root cause, a list of attribute combinations of the form {"column": "value"} (values are the integer codes from df.csv):

  [{"B": "273"}]
  

• anomalyIndex — row indices into df.csv of the records flagged as anomalous in this case (typically a few hundred indices).

Task Definition

For each test case, given anomalyIndex, a method must predict one or more root causes as attribute combinations. Predictions are compared against the ground-truth label:

• Precision — fraction of predicted combinations that are true root causes
• Recall — fraction of true root causes that are predicted
• F1 — harmonic mean

Aggregating TP/FP/FN across all 100 cases yields micro-averaged Precision/Recall/F1.

Quick Start

import json
import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_csv(
    hf_hub_download("Ysqq/NFC", "df.csv", repo_type="dataset")
)

with open(
    hf_hub_download("Ysqq/NFC", "label.json", repo_type="dataset")
) as f:
    cases = json.load(f)

case = cases[0]
anomaly_rows = df.iloc[case["anomalyIndex"]]
print(f"Ground-truth root cause: {case['label']}")

A reference implementation (the Prism algorithm with a statistical noise-filtering variant, PrismNFC) is available in the accompanying code repository.

Citation

If you use this dataset in your research, please cite the associated Prism paper (citation to be added upon publication).