MISSION: Ultra Large-Scale Feature Selection using Count-Sketches

Authors: Amirali Aghazadeh, Ryan Spring, Daniel Lejeune, Gautam Dasarathy, Anshumali Shrivastava, baraniuk

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We designed a set of simulations to evaluate MISSION in a controlled setting. All experiments were performed on a single machine, 2x Intel Xeon E5-2660 v4 processors (28 cores / 56 threads) with 512 GB of memory. The code1 for training and running our randomized-hashing approach is available online.
Researcher Affiliation Academia 1Department of Electrical Engineering, Stanford University, Stanford, California 2Department of Computer Science, Rice University, Houston, Texas 3Department of Electrical and Computer Engineering, Rice University, Houston, Texas.
Pseudocode Yes Algorithm 1 MISSION
Open Source Code Yes The code1 for training and running our randomized-hashing approach is available online. 1https://github.com/rdspring1/MISSION
Open Datasets Yes Datasets: We used four datasets in the experiments: 1) KDD2012, 2) RCV1, 3) Webspam Trigram, 4) DNA2. The statistics of these datasets are summarized in Table 2. 2http://projects.cbio.mines-paristech.fr/largescalemetagenomics/ 3https://www.kaggle.com/c/criteo-display-ad-challenge
Dataset Splits No The paper provides 'Train Size' and 'Test Size' for the datasets but does not explicitly mention a 'validation' split or describe its configuration.
Hardware Specification Yes All experiments were performed on a single machine, 2x Intel Xeon E5-2660 v4 processors (28 cores / 56 threads) with 512 GB of memory.
Software Dependencies No The paper states 'The code1 for training and running our randomized-hashing approach is available online.' but does not specify particular software dependencies with version numbers (e.g., Python, PyTorch, etc.).
Experiment Setup Yes For all methods, we used the logistic loss for binary classification and the cross-entropy loss for multi-class classification. For all the experiments, the Count-Sketch data structure used 3 hash functions, and the model weights were divided equally among the hash arrays. All the methods were trained for a single epoch with a learning rate of 0.5.