Categorical Feature Compression via Submodular Optimization

Authors: Mohammadhossein Bateni, Lin Chen, Hossein Esfandiari, Thomas Fu, Vahab Mirrokni, Afshin Rostamizadeh

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 3, we present our empirical evaluation of optimizing the mutual information objective as well as an end-to-end learning task. All the experiments are performed using the Criteo click prediction dataset (Criteo Labs, 2014)
Researcher Affiliation Collaboration 1Google, New York, NY, USA 2Department of Electrical Engineering, Yale University, New Haven, CT, USA.
Pseudocode Yes Algorithm 1 Data structure to compute s F(S)
Open Source Code No The paper does not state that the authors' code for the described methodology is open source or provide a link to it.
Open Datasets Yes All the experiments are performed using the Criteo click prediction dataset (Criteo Labs, 2014)
Dataset Splits Yes All the experiments are performed using the Criteo click prediction dataset (Criteo Labs, 2014), which consists of 37 million instances for training and 4.4 million held-out points. Note, we use the labeled training file from this challenge and chronologically partitioned it into train/hold-out sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running its experiments, only mentioning general terms like 'large-scale machine learning tasks'.
Software Dependencies No The paper mentions 'TensorFlow' and 'Adam' as tools used but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup No The paper mentions obtaining vocabularies of certain sizes (e.g., 10K to 160K) and that feature values appearing in at least 100 instances were used. However, it lacks specific hyperparameter values (e.g., learning rate, batch size) or detailed optimizer settings needed to reproduce the experimental setup.