Convex Calibrated Surrogates for Hierarchical Classification

Authors: Harish Ramaswamy, Ambuj Tewari, Shivani Agarwal

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on a number of benchmark datasets show that the resulting algorithm, which we term Ov A-Cascade, gives improved performance over other state-of-the-art hierarchical classification algorithms.
Researcher Affiliation Academia Harish G. Ramaswamy HARISH GURUP@CSA.IISC.ERNET.IN Indian Institute of Science, Bangalore, INDIA Ambuj Tewari TEWARIA@UMICH.EDU University of Michigan, Ann Arbor, USA Shivani Agarwal SHIVANI@CSA.IISC.ERNET.IN Indian Institute of Science, Bangalore, INDIA
Pseudocode Yes Algorithm 1 OVA-Cascade Training
Open Source Code No The paper does not provide any specific repository link, explicit code release statement, or mention of code in supplementary materials for the methodology described.
Open Datasets Yes CLEF (Dimitrovski et al., 2011) Medical X-ray images organized according to a hierarchy. IPC 3 Patents organized according to the International Patent Classification Hierarchy. LSHTC-small, DMOZ-2010 and DMOZ-2012 4 Web-pages, from the LSHTC (Large-Scale Hierarchical Text Classification) challenges 2010-12, organized according to a hierarchy. 3http://www.wipo.int/classifications/ipc/en/support/ 4http://lshtc.iit.demokritos.gr/node/3
Dataset Splits Yes Table 1. Dataset Statistics Dataset #Train #Validation #Test #Labels #Leaf-Labels Depth #Features CLEF 9,000 1,000 1,006 97 63 3 89 LSHTC-small 4,463 1,860 1,858 2,388 1,139 5 51,033 IPC 35,000 11,324 28,926 553 451 3 541,869 DMOZ-2010 80,000 13,805 34,905 17,222 12,294 5 381,580 DMOZ-2012 250,000 50,000 83,408 13,347 11,947 5 348,548
Hardware Specification No The paper mentions running algorithms 'on a 4-core CPU' but does not provide specific details such as the CPU model, GPU, memory, or other detailed computer specifications.
Software Dependencies No The paper states 'We use LIBLINEAR (Fan et al., 2008) for the SVM-train subroutine' but does not specify the version number of this or any other software dependency.
Experiment Setup No The paper mentions that 'The regularization parameter C is chosen via a separate validation set. The thresholds τj for j [h] are also chosen via a coarse grid search using the validation set.' but does not provide the specific hyperparameter values for C or τj.