Pairwise-Covariance Linear Discriminant Analysis

Authors: Deguang Kong, Chris Ding

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiment resultsClassification Performance Analysis Table 2 and Fig.4 present the classification performance using different dimension reduction methods.
Researcher Affiliation Academia Deguang Kong and Chris Ding Department of Computer Science & Engineering University of Texas, Arlington, 500 UTA Blvd, TX 76010 doogkong@gmail.com; chqding@uta.edu
Pseudocode Yes Algorithm 1: Computation of r J2(G) (i.e., Eq.11) or r J2(A) (i.e., gradient of Eq.21).
Open Source Code No The paper does not provide concrete access to its source code, nor does it explicitly state that its code is being released.
Open Datasets Yes In Fig.2, we show the results on the widely used iris data1. Iris has 150 data points with K=3 classes. Thus LDA project to K-1=2 dimensions. Fig.2 indicates that pc LDA gives clear discrimination between classes 2 and 3 while standard LDA has strong mixing between classes 2 and 3. 1http://archive.ics.uci.edu/ml/datasets/Iris and Dataset We evaluate the proposed pairwise-covariance LDA using four data sets (see Table 1) for multi-class classification experiments, including one face dataset umist, two digit datasets mnist (Lecun et al. 1998), binalpha, one image scene dataset MSRCv1 (Lee and Grauman 2009)2. Due to space limit, we omit more details of datasets. Table 1 summarizes the datasets. 2http://research.microsoft.com/enus/projects/Object Class Recognition/
Dataset Splits Yes In our experiment, we use 5-round 5-fold cross validation to evaluate the classification performance. Each dataset is evenly partitioned into 5 parts. Only one part is used as testing and the other 4 parts are used for training. We report the average results for 5 rounds.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions software components like 'k(k=3) nearest neighbor classifier' and 'RBF kernel' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes We set q = 1 for Eq.(10), Eq.(21) in our experiments. The parameter β is set to be {0.1, 0.5, 1}. To make a fair comparison, we project all original data to (C-1) dimension, and k(k=3) nearest neighbor classifier is used for classification purpose. For kernel LDA, we use RBF kernel to construct the pairwise similarity Wij = e γkxi xjk2, where bandwidth γ is searched in the grid {10 4, 10 3, , 103, 104}.