Multiple-Instance Learning from Distributions
Authors: Gary Doran, Soumya Ray
JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform an extensive empirical evaluation that supports the theoretical predictions entailed by the new framework. The proposed theoretical framework leads to a better understanding of the relationship between the MI and standard supervised learning settings, and it provides new methods for learning from MI data that are more accurate, more efficient, and have better understood theoretical properties than existing MI-specific algorithms.Our evaluation uses 55 data sets from a wide variety of domains, and supports both our theoretical results as well as the assumptions made by our generative model. |
| Researcher Affiliation | Academia | Gary Doran EMAIL Soumya Ray EMAIL Department of Electrical Engineering and Computer Science Case Western Reserve University 10900 Euclid Ave, Glennan 320 Cleveland, OH 44106, USA |
| Pseudocode | No | The paper contains numerous theorems, lemmas, and proofs, but no clearly labeled 'Pseudocode' or 'Algorithm' sections, nor any structured code-like blocks. |
| Open Source Code | No | We use the authors original MATLAB code, found at http://lamda.nju.edu.cn/code_KISVM.ashx, for the key instance SVM (KI-SVM) approach (Liu et al., 2012). |
| Open Datasets | Yes | To evaluate our hypothesis that a supervised SVM can perform well with respect to AUC for learning instanceand bag-labeling functions, we use a total of 55 real-world data sets across a variety of problem domains, including 3D-QSAR (Dietterich et al., 1997), CBIR (Andrews et al., 2003; Maron and Ratan, 1998; Rahmani et al., 2005), text categorization (Andrews et al., 2003; Settles et al., 2008), and audio classification (Briggs et al., 2012). |
| Dataset Splits | Yes | We evaluate algorithms using 10-fold stratified cross-validation, with 5-fold inner-validation used to select parameters using random search (Bergstra and Bengio, 2012). |
| Hardware Specification | No | This work made use of the High Performance Computing Resource in the Core Facility for Advanced Research Computing at Case Western Reserve University. |
| Software Dependencies | No | The experiments used for this work were implemented in Python using Num Py (Ascher et al., 2001) and Sci Py (Jones et al., 2001) for general matrix computations and the CVXOPT library (Dahl and Vandenberghe, 2009) for solving quadratic programs (QPs). |
| Experiment Setup | Yes | Parameter selection is performed with respect to bag-level labels (since instance-level labels are unavailable at training time, even during cross-validation). We use the radial basis function (RBF) kernel with all algorithms, with scale parameter γ [10 6, 101], and regularization loss trade-off parameter C [10 2, 105]. The L2 norm is used for regularization in all algorithms. |