Affinity Weighted Embedding
Authors: Jason Weston, Ron Weiss, Hector Yee
ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We give results on a variety of datasets showing the usefulness of these various choices for music annotation, image annotation and You Tube video recommendation. |
| Researcher Affiliation | Industry | Jason Weston JWESTON@GOOGLE.COM Ron Weiss RONW@GOOGLE.COM Google Inc, New York, NY, USA Hector Yee HYEE@GOOGLE.COM Google Inc, San Bruno, CA, USA. |
| Pseudocode | Yes | Algorithm 1 Affinity Weighted Embedding SGD training. |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code for the described methodology or a link to a code repository. |
| Open Datasets | Yes | We conducted experiments on three different tasks: (i) Magnatagatune (annotating music with text tags); (ii) Image Net (annotation images with labels); and (iii) You Tube video recommendation (recommending videos for a given user). The paper cites the sources for Magnatagatune and ImageNet: (Law et al., 2009) and (Deng et al., 2009). |
| Dataset Splits | Yes | There are 16,289 data examples used for training and validation, 6498 examples used for test, and 160 possible tags. (Magnatagatune), We used the Fall 2011 version, which contains about 10M images, from which we kept 10% for validation, 10% for test, and the remaining 80% for training. (Image Net), We set aside 0.5M examples for validation, and 1M for test. (You Tube) |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory specifications) used to run its experiments. It mentions using 'a Map Reduce framework' for efficiency but no concrete hardware. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific solver versions). |
| Experiment Setup | Yes | We optimized the other hyperparameters of WSABIE (learning rate and margin) and then used the same hyperparameters for AWE. We used the latent k-NN affinity function (cf. eq. (9)), with k = 20, and optimized with the iterative training method, using only 3 steps... and We used an embedding dimension of 100. and We used an embedding dimension of 128 for both. and We took the best performing existing WSABIE system (although for memory and speed reasons we fixed the embedding dimension to be m = 64) and used the same parameters for AWE. |