Reconstruction-based Unsupervised Feature Selection: An Embedded Approach

Authors: Jundong Li, Jiliang Tang, Huan Liu

IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various types of realworld datasets demonstrate the effectiveness of the proposed framework REFS.
Researcher Affiliation Academia Computer Science and Engineering, Arizona State University, USA Computer Science and Engineering, Michigan State University, USA {jundong.li, huan.liu}@asu.edu, tangjili@msu.edu
Pseudocode Yes Algorithm 1 Reconstruction-based Unsupervised Feature Selection (REFS)
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes We choose 8 benchmark datasets of various types for evaluation, including three image datasets, i.e., object image dataset COIL201 [Nene et al., 1996], face image dataset ORL2, handwritten digit datasets USPS [Hull, 1994]; two text datasets, i.e., RELATHE and BASEHOCK3; two microarray datasets, i.e., Lung [Bhattacharjee et al., 2001] and GLIOMA [Nutt et al., 2003] and one spoken letter recognition dataset Isolet4. Footnotes 1, 2, 3, 4 provide URLs for access.
Dataset Splits No The paper does not provide specific dataset split information (train/validation/test) for reproducing the data partitioning or the clustering experiments. It mentions using K-means for evaluation and varying the number of selected features.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No The paper mentions K-means and Lasso but does not provide specific software dependencies with version numbers.
Experiment Setup Yes For Lap Score, MCFS, UDFS and FSASL, we specify the number of neighborhood size to be 5 to construct the Laplacian matrix on the data instances following previous work. In REFS, we also set the number nearest neighborhood size p to be 5, but the Laplacian matrix is built on the feature vectors instead of on the data instances. ... tune these regularization parameters for all methods by grid search and report the best performance.