Detailed results of "Exploring the impact of label-level noise on multi-label k-Nearest Neighbor classification"
Description
Detailed experimental results of the eight data assorments studied in the work: Corel5k, bibtex, birds, emotions, genbase, medical, scene, and yeast. Each collection is provided separately in a CSV file, containing with the following columns: - noise: Policy used to induce noise in the assortment. The possible values are: DAAS, PUMN, add, add-sub, sub, and swap. - percen: Percentage of the assortment that may be altered by the "noise" policy. The possible values are: 0, 10, 20, 30, 40, 50, and 60. - prob: Probability of inducing noise in the dataset. The possible values are: 0, 0.1, 0.2, 0.3, 0.4, 0.5, and 0.6. - PG_method: Prototype generation method used to reduce the size of the dataset. Possible values are: ALL, MChen, MRHC, MRSP1, MRSP2, and MRSP3. - Reduction_parameter: Parameters of the Prototype generation method at hand. Possible values are: 1.0, 10.0, 30.0, 50.0, 70.0, and 90.0. - Classifier: Multi-label learning model based on the k-Nearest Neighbor rule used. Possible values are: BRkNNaClassifier, LabelPowerset, and MLkNN. - k: Number of neighbors considered by the classifier at hand. Possible values are: 1.0, 3.0, 5.0, 7.0, 11.0, 15.0, 21.0, and 30.0. - HL: Hamming Loss. - EMR: Exact Match Ratio. - acc: Standard accuracy. - jaccard-m: Micro Jaccard index. - F1-m: Micro F1 score. - jaccard-M: Macro Jaccard index. - F1-M: Macro F1 score. - jaccard-s: Sampled-based Jaccard index. - F1-s: Sampled-based F1 score. - RL: Ranking Loss. - Size: Resulting set size after the reduction process by the Prototype generation method. In addition, a supplementary analysis in terms of the macro-F1 score (document "SupplementaryAnalysis_macro-F1.pdf") is provided to complement the insights observerd in the main manuscript.
Files
Institutions
- Universitat d'AlacantComunitat Valenciana, Alacant