{"id":"W4211256381","doi":"10.1109/access.2022.3151048","title":"MLCM: Multi-Label Confusion Matrix","year":2022,"lang":"en","type":"article","venue":"IEEE Access","topic":"Text and Document Classification Technologies","field":"Computer Science","cited_by":409,"is_retracted":false,"has_abstract":true,"ca_institutions":"Toronto Metropolitan University; Vector Institute; McMaster University","funders":"Ministère de la Défense Nationale","keywords":"Computer science; Confusion; Confusion matrix; Artificial intelligence; Classifier (UML); Class (philosophy); Ambiguity; Machine learning; Programming language","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005264724,0.003701651,0.001887485,0.008003796,0.001862052,0.004771274,0.005331772,0.003175834,0.03626817],"category_scores_gemma":[0.02744752,0.00111726,0.002018406,0.004376899,0.001346187,0.006020645,0.005036145,0.004282397,0.02732536],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002237642,"about_ca_system_score_gemma":0.004852526,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.007191714,"about_ca_topic_score_gemma":0.008582709,"domain_scores_codex":[0.9889236,0.003254606,0.00096604,0.001922103,0.004395531,0.0005382273],"domain_scores_gemma":[0.9861109,0.005012678,0.001227253,0.002500374,0.004664953,0.0004837859],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0007385141,0.0003433214,0.001644938,0.001581808,0.0001941073,0.0004220669,0.0003862878,0.03430916,0.011061,0.03098334,0.2615465,0.6567889],"study_design_scores_gemma":[0.0001903997,0.0003516891,0.002654561,0.0004550941,0.00008576733,0.0008977198,0.0003039681,0.6371062,0.04776569,0.1204492,0.1892549,0.0004847844],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.002471395,0.0007961969,0.9293271,0.0008850894,0.0005481102,0.0007746046,0.01029852,0.04942818,0.00547084],"genre_scores_gemma":[0.06476128,0.0006537208,0.8915112,0.001008067,0.0004546756,0.002335115,0.024719,0.003436018,0.0111208],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.03626817,"threshold_uncertainty_score":0.1213291,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.08270770067299012,"score_gpt":0.3675705466220109,"score_spread":0.2848628459490208,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}