SiameseXML: Siamese Networks meet Extreme Classifiers with 100M Labels
Authors: Kunal Dahiya, Ananye Agarwal, Deepak Saini, Gururaj K, Jian Jiao, Amit Singh, Sumeet Agarwal, Purushottam Kar, Manik Varma
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6. Experiments, Tables 1 and 2 present results on benchmark datasets for shortand full-text XML tasks where Siamese XML could be 2 13% more accurate than methods which make use of label meta-data, viz. ECLARE, Gala XC, DECAF and X-Transformer in propensity scored precision. |
| Researcher Affiliation | Collaboration | 1Indian Institute of Technology Delhi 2Microsoft Research 3Microsoft 4Indian Institute of Technology Kanpur. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code for Siamese XML is available at https: //github.com/Extreme-classification/siamesexml |
| Open Datasets | Yes | The datasets and provenance details thereof can be found on the Extreme Classification Repository (Bhatia et al., 2016). URL http: //manikvarma.org/downloads/XC/XMLRepository.html. |
| Dataset Splits | Yes | where α [0, 1] is a hyper-parameter set via validation. |
| Hardware Specification | Yes | All training times are reported on a 24-core Intel Xeon 2.6 GHz machine with a single Nvidia V100 GPU. However, Q2BP datasets were afforded multiple GPUs. |
| Software Dependencies | No | The paper mentions the 'Adam optimizer' but does not specify any software libraries or dependencies with version numbers. |
| Experiment Setup | Yes | For each label l in a batch, a single positive data point was sampled uniformly at random and κ = 1 5 hard negative data points were selected..., A fixed value of β = 0.7 was used and not tuned., Module IV ... minimized using mini-batches over data points and the Adam optimizer (Kingma & Ba, 2015), where Pi = {l | yil = 1} is the set of positive labels for data point i and ˆ Ni is the set of κ |Pi| hardest negatives from the label shortlist... (κ = 2 5). |