Stochastic Multiresolution Persistent Homology Kernel
Authors: Xiaojin Zhu, Ara Vartanian, Manish Bansal, Duy Nguyen, Luke Brandl
IJCAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate SMURPH s potential for clustering and classification on several applications, including eye disease classification and human activity recognition. 3 Experiments We present three applications in clustering and classification to demonstrate the potential of the SMURPH kernel. |
| Researcher Affiliation | Academia | Xiaojin Zhu, Ara Vartanian, Manish Bansal, Duy Nguyen, Luke Brandl Department of Computer Sciences, University of Wisconsin Madison 1210 W. Dayton Street, Madison, Wisconsin, USA 53706 |
| Pseudocode | Yes | Algorithm 1 SMURPH Kernel |
| Open Source Code | No | The paper references third-party software like Javaplex, Perseus, TDA, Dionysus, GUDHI, kernlab, and SVMlight, but does not provide a statement or link to its own open-source code for the SMURPH kernel implementation described. |
| Open Datasets | Yes | Our sample consists of 67 retinal images, collected as part of the STARE [Hoover and Goldbaum, 2013] project. The daily and sports activities data set [Altun et al., 2010] contains sensor data of several everyday activities |
| Dataset Splits | Yes | We tune all parameters (regularization C for SVM and bandwidth σ for RBF kernel) using an inner cross validation (CV) inside the outer CV training portion. We feed the kernels to an SVM (kernlab [Karatzoglou et al., 2004]) for classification and report 5-fold CV accuracy. We measure 8-fold CV error, where in each fold we use 7 people (total of 35 activities) as the training data, and leave all 5 activities from one person out as test data. |
| Hardware Specification | No | The paper mentions 'a typical desktop computer' when discussing computation time for existing TDA software in Figure 1, but does not provide specific hardware details (e.g., CPU, GPU models, or memory) for the experiments conducted using SMURPH. |
| Software Dependencies | No | The paper mentions 'MATLAB s bwmorph', 'kernlab', and 'SV M light' but does not specify their version numbers or other software dependencies with versions. |
| Experiment Setup | Yes | We compute SMURPH kernel matrix using a radius of r = 0.1, m = 20 centers per point cloud, s = 1 samples per center, and a budget of b = 350 points per sample. We choose a resolution level of r = 8 and m = 50, s = 1, b = 300. We tune all parameters (regularization C for SVM and bandwidth σ for RBF kernel) using an inner cross validation (CV) inside the outer CV training portion. The tuning grid is C 2 {2 7, 2 6, . . . , 27} and σ 2 {100, 101, . . . , 103}. We use two resolution levels: r1 = 125 and r2 = 25 which correspond to 10-second and 2-second intervals. We set the parameters m = 10, s = 1, b = 100. We compute the 40 40 SMURPH kernel matrix with weights w1 = w2 = 1 in (7). On a parameter grid C 2 {10 5, 10 4, . . . , 105}, this inner CV selects C = 100. |