Interpretable Predictive Modeling for Climate Variables with Weighted Lasso
Authors: Sijie He, Xinyan Li, Vidyashankar Sivakumar, Arindam Banerjee1385-1392
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We establish finite sample estimation error bounds for weighted Lasso, and illustrate its superior empirical performance and interpretability over complex models such as deep neural networks (Deep nets) and gradient boosted trees (GBT).We perform extensive comparison of the weighted Lasso with baselines on data from 3 different Earth System Models (ESMs). |
| Researcher Affiliation | Academia | 1Depertment of Computer Science & Engineering University of Minnesota, Twin Cities Minneapolis, MN 55455 hexxx893@umn.edu, lixx1166@umn.edu, sivak017@umn.edu, banerjee@cs.umn.edu |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions using third-party packages like xgboost and Keras but does not provide a specific statement or link for the source code of the weighted Lasso model or other custom implementations described in the paper. |
| Open Datasets | Yes | We use data from the historical runs of 3 ESMs (see Table 1) included as part of the core set of experiments in CMIP5 (Taylor, Stouffer, and Meehl 2012). and Table 1: Description of the Earth System Models used in the experiments. Model name Origin Reference CMCCCESM Centro Euro-Mediterraneo per I Cambiamenti Climatici (Italy) (Fogli et al. 2009) INM-CM4 Institute for Numerical Mathematics (Russia) (Volodin, Dianskii, and Gusev 2010) MIROC5r1i1p1 Atmosphere and Ocean Research Institute... (Watanabe et al. 2010) |
| Dataset Splits | Yes | We divide the data into 10 training sets by applying a moving window of 100 years with a stride of 5 years. The 10 years subsequent to the end of the training set are used for testing. Specifically, in each training set we select the first 80 years to train the model and use the next 20 years as a validation set. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory, or cloud instance types) used for running the experiments. |
| Software Dependencies | No | GBTs are implemented in Python using xgboost package (Chen and Guestrin 2016). All networks are implemented in Python using Keras package (Chollet 2015). No version numbers for Python, xgboost, or Keras are provided. |
| Experiment Setup | Yes | The hyperparameters for weighted Lasso (regularization parameter), Lasso (regularization parameter), PCR (number of principal components for regression) and GBT (learning rate and maximum depth of tree) are selected by validation set. For GBT, we fix the number of trees to 100, and perform a grid-search to find the optimal learning rate and maximum tree depth. For all models the optimal value of learning rate on the validation set varies between 0.05 and 0.07 and the optimal maximum tree depth is found to be 3. For Deep nets we experiment with various combinations of: (a) the number of hidden layers, (b) the number of hidden units in each layer, (c) different mini-batch size when training using the Adam optimization algorithm (Kingma and Ba 2014), and (d) ℓ1, ℓ2 and no regularization. The maximum number of epochs for training is set as 150. We also use early-stopping by examining validation set error. In almost all cases, an 8 hidden layer Deep nets with ℓ1 regularization on the weights gave the best performance on the validation set. We report results with mini batch size set to 32. |