This repo implements the embedding models in the 2017 NIPS paper "Context Selection for Embedding Models".
Embedding models aim to find vector representations of items that encode item relations. The approach is to predict the count/presence of one item given its context items. However, some context items in the context set do not reflect item relations and should not be considered by the embedding model. Context selection is designed down-weight unrelated context items when learning embeddings. This new model use hidden indicators to decide which context item should be used in the embedding model. In the learning stage when hidden indicators are approximately marginalized out, unrelated context items contribute less to embedding learning. Please see the details in the paper.
python demo.py
Note: this repo does not contain any data -- it only use some random data to show how to use the code. The code requires numpy, scipy, and tensorflow.
If you have any questions, please contact the Li-Ping Liu (liping.liulp at gmail).
If you have used the code in your work, please cite:
@inproceedings{context-select17, title = {Context Selection for Embedding Models}, author = {Li-Ping Liu, Francisco J.R. Ruiz, Susan C. Athey, and David M. Blei}, booktitle ={Accepted by Advances in Neural Information Processing Systems 30}, pages = {}, year = {2017}, }