Skip to content

Latest commit

 

History

History
43 lines (33 loc) · 2.06 KB

README.md

File metadata and controls

43 lines (33 loc) · 2.06 KB
Raw

cdm-icml

Code that Accompanies "Discovering Context Effects from Raw Choice Data"

The following code uses the open source machine learning framework pytorch (https://pytorch.org/) to implement the CDM as described in "Discovering Context Effects from Raw Choice Data" by Arjun Seshadri, Alexander Peysakhovich and Johan Ugander in ICML 2019 (http://proceedings.mlr.press/v97/seshadri19a.html).

Below is code for a sample run to train the CDM on SFShop, a dataset explored in the paper (as run from the code directory):

import cdm_pytorch as cp
val_loss, tr_loss, gv, train_ds, val_ds, model, opt = cp.default_run(dataset='SFshop',
                                                                     batch_size=None,
                                                                     epochs=500, 
                                                                     embedding_dim=3, 
                                                                     lr=5e-2, 
                                                                     seed=0,
                                                                     wd=0)
print(f'Val Loss: {val_loss.item()}, Tr Loss: {tr_loss.item()}, final_grad: {gv.item()}')

The above function is a wrapper function to use the CDM. The inputs used are described below, but see in-line documentation for more details.

dataset - name of dataset to run CDM on. Default options are 'SFwork', 'SFshop', or 'syn_nature_triplet' to run the datasets in the paper.

batch_size - hyperparameter for training, number of points to be used for each gradient step (None means full batch)

epochs - number of epochs to perform optimization. If batch_size is set to None, this is just the number of gradient steps taken

embedding_dim - dimension of CDM (r in the paper)

lr - learning rate for Adam optimizer (the default optimizer)

seed - random seed that sets dataset splits + initialization for reproducibility (None means no seed)

wd - weight decay parameter for the model (similar to l2 regularization, but faster - see pytorch docs for more info)