Clustering-oriented representation learning with Attractive-Repulsive loss

Code for my AAAI 2019 paper in the Network Interpretability for Deep Learning Workshop.

This repository includes the following:

• The split-up AGNews dataset into 8 topics (see directory agnews);
• Implementations of Gaussian-COREL and Cosine-COREL in high- and low-level ways for ease of integration.

COREL Implementations

Here, you can the different ways for implementing COREL models, depending on your use-cases.

• Direction 0 (the high-level API): pass your model (which does NOT have an output layer) into the constructor for a CORELWrapper class, such that you can simply feed forward any input to get predictions, then using the function get_loss_function() to get exactly the correct loss that you will need.

• Direction 1 (the low-level API): use the loss functions, prediction functions, and attractive-repulsive helpers directly as you see fit.

See example.py for a simple, straightforward example of how to do option 0 (recommended).

---

If you have any questions, please feel free to email me at kiankd@gmail.com.

Best, Kian