TFN

Official implementation of TFN: "Twin Fuzzy Networks with Interpolation Consistency Regularization for Weakly-supervised Anomaly Detection" on IEEE Transactions on Fuzzy Systems

Dependencies

MATLAB R2022b

Data Preparation

• Get datasets mentioned in the paper:
  • Download ".zip" file in this share link, and unzip it to the folder "./Datasets"
• Generate your anomaly detection datasets:
  • All datasets mentioned in the paper are generated using the python code of DevNet proposed in the SIGKDD paper Deep Anomaly Detection with Deviation Networks
  • To run your own datasets with weakly supervision, name your generated train/test datasets in the following format:
    • Train dataset: "{}_weakly_train_{}_{}.mat".format(origin_name, contamination_rate, num_known_anomalies)
    • Test dataset: "{}_weakly_test_{}_{}.mat".format(origin_name, contamination_rate, num_known_anomalies)

Experiments

• To fetch the results in the paper:
  • Run "main.m".

Useful Information

• Results:
  • The "log_{n}.txt" files that keep the display output for experiments are stored in "./Logs" directory.
  • The corresponding tables "log_Result_{n}.csv" that store results for that experiments are stored in "./Logs" directory.
• Parameter Settings (in "main.m" file):
  1. EXP: Set the random seed, and the number of runs for each dataset
  2. datasets: List all the origin names of the datasets under weak supervision
  3. WSAD: Set the hyperparameters for the weak supervision scenarios. known_outlier is the number of known anoamlies. contamination is the contamination rate.
  4. PCA: pca determine whether PCA is used. min_dim is the minimum dimension for the usage of PCA. threshold is the explained variation ratio.
  5. PTRAIN: In method, fixed uses the pre-set number of rules fix_rule for these datasets, and search will search for ideal number of rules resulting non-empty clusters between min_fule and max_rule.
  6. AUG:
     • num_train is the number of data pairs (hyperparameter $M$ in the paper).
     • c_values is the list of $[C_{a,a}, C_{u,a}, C_{u,u}]$.
     • E_test is the number of train data sampled in the test phase (hyperparameter $E$ in the paper).
  7. MIXUP:
     • type determine whether ICR is used, chosen between "No" and "ICR".
     • M is the number of virtual training pairs in the ICR process.
     • gamma is the weight of the ICR loss.
  8. REGU: lambda is the weight of the $l_2$ regularization term.
  9. TRAIN: cluster choose the cluster method used in TFN, choosen between "p_fcm" and "k-means".

Full Paper

The full paper can be found at this link.

Citation

@ARTICLE{cao2024twin,
  author={Cao, Zhi and Shi, Ye and Chang, Yu-Cheng and Yao, Xin and Lin, Chin-Teng},
  journal={IEEE Transactions on Fuzzy Systems}, 
  title={Twin Fuzzy Networks With Interpolation Consistency Regularization for Weakly-Supervised Anomaly Detection}, 
  year={2024},
  volume={32},
  number={9},
  pages={5086-5097},
  keywords={Anomaly detection;Training;Prototypes;Uncertainty;Interpolation;Optimization;Knowledge engineering;Data uncertainty;fuzzy c-means clustering;interpolation consistency regularization;twin fuzzy networks;weakly-supervised anomaly detection},
  doi={10.1109/TFUZZ.2024.3412435}}