Spectrally Transformed Kernel Regression
Authors: Runtian Zhai, Rattana Pukdee, Roger Jin, Maria Florina Balcan, Pradeep Kumar Ravikumar
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We implement STKR-Prop (SP) with inverse Laplacian (Lap), polynomial (poly) s(λ) = λp, and kernel PCA (topd). We run them on several node classification tasks, and compare them to Label-Prop (LP) and KRR with the base kernel (i.e. STKR with s(λ) = λ). Details and full results are deferred to Appendix D, and here we report a portion of the results in Table 1, in which the best and second-best performances for each dataset are marked in red and blue. |
| Researcher Affiliation | Academia | Runtian Zhai, Rattana Pukdee, Roger Jin, Maria-Florina Balcan, Pradeep Ravikumar Carnegie Mellon University {rzhai,rpukdee,rrjin,ninamf,pradeepr}@cs.cmu.edu |
| Pseudocode | Yes | Algorithm 1 STKR-Prop for simple s Input: GK, s(λ), βn, y, γ, ϵ 1: Initialize: ˆα 0 Rn 2: while True do # Compute u = (G ˆ Ks,n + nβn In) ˆα 3: α 1 n+m GK,n+m,n ˆα, v 0 Rn+m 4: for p = q, , 2 do v GKv 5: u G K,n+m,nv + π1GK,n ˆα + nβn ˆα 6: if u y 2 < ϵ y 2 then return ˆα 7: ˆα ˆα γ(u y) |
| Open Source Code | Yes | The code of Section 5 can be found at https://colab.research.google.com/drive/ 1m8OENF2lvx W3BB6CVEu45SGe K9Io Ypd1?usp=sharing. |
| Open Datasets | Yes | We focus on graph node classification tasks, and work with the publicly available datasets in the Py Torch Geometric library (Fey & Lenssen, 2019), among which Cora, Cite Seer and Pub Med are based on Yang et al. (2016); Computers, Photos, CS and Physics are based on Shchur et al. (2018); DBLP and Cora Full are based on Bojchevski & Günnemann (2018). |
| Dataset Splits | Yes | We split a dataset into four sets: train, validation (val), test and other. Among them, train and val contain labeled samples, while test and other contain unlabeled samples. The test performance which we will report later is only evaluated on the test set. The val set is used to select the best model, so it is used in a similar way as the test set as explained below: ... In all our experiments, these four sets are randomly split. |
| Hardware Specification | No | The paper mentions 'a boost in computational power' generally but does not specify any particular hardware (e.g., GPU models, CPU types, or memory) used for their experiments. |
| Software Dependencies | No | The paper mentions using 'Py Torch Geometric library' but does not provide specific version numbers for this or any other software dependencies, which are required for reproducibility. |
| Experiment Setup | Yes | Hyperparameters. Below are the hyperparameters we use in the experiments. Best hyperparameters are selected with the validation split as detailed above. ... Number of iteration T [1, 2, 4, 8, 16, 32] ... η [0.7, 0.8, 0.9, 0.99, 0.999, 0.9999, 0.99999, 0.999999] ... k [1, 2, 4, 6, 8] ... β [103, 102, 101, 100, 10 1, 10 2, 10 3, 10 4, 10 5, 10 6, 10 7, 10 8] ... Number of representation dimension d [32, 64, 128, 256, 512] |