Code Monkey home page Code Monkey logo

experiments's Introduction

Experiments on more baselines

We add BGRL, MASKGAE, AFGRL and COSTA as baseline and report their performance. Also, we train the previous "OOM" result in batched manner to show the performance.

Experiments on clustering

clustering Cora CiteSeer PubMed Photo Computer CS Physics Mean
BGRL 58.5 45.09 41.3 64.56 58.15 83.28 76.91 61.11
MaskGAE 60.35 44.58 43.41 64.96 59.51 82.89 76.78 61.78
AFGRL 60.87 42.41 30.82 63.45 52.41 82 75.22 58.17
COSTA 54.7 40.66 36.79 54.76 40.8 77.95 67.42 53.30
GRACE 53.48 41.17 44.85 63.14 55.8 66.4 65.34 55.74
DGI 57.13 44.6 36.83 64.1 54.64 68.79 62.87 55.57
CCA-SSG 58.12 31.54 38.92 65.79 55.29 80.53 75.62 57.97
MVGRL 56.37 44.17 37.85 63.75 55.24 69.32 67.95 56.38
GraphMAE 60.27 44.55 40.48 56.4 56.04 80.3 73.54 58.80
SeeGera 59.81 45.5 41.27 58 56.42 76.6 77.32 59.27
GCLFormer 61.15 45.62 39.44 65.87 57.32 88.71 77.6 62.24

Experiments on node classification

node Cora CiteSeer PubMed Photo Computer CS Physics Mean
BRGL 84.27 73.04 84.47 92.89 89.72 93.97 95.64 87.71
MaskGAE 84.52 73.32 85.26 92.74 89.41 93.46 95.82 87.79
AFGRL 84.46 72.45 84.92 93.14 89.74 93.16 95.32 87.60
COSTA 84.73 72.05 84.31 92.17 88.41 92.73 95.41 87.12
GRACE 84.08 71.76 86.03 77.79 77.98 89.65 94.82 83.16
DGI 83.86 70.25 86.05 92.21 88.03 90.36 94.71 86.50
CCA-SSG 83.71 70.45 86.07 92.76 79.26 93.18 95.7 85.88
MVGRL 83.5 72.3 80.1 91.74 87.52 92.11 95.28 86.08
GraphMAE 84.46 72.81 85.8 88.81 81.87 92.69 95.53 86.00
SeeGera 84.34 71.72 85.68 92.33 88.27 93.79 95.79 87.42
GCLFormer 84.74 72.97 85.32 93.71 87.47 94.87 95.98 87.87

Experiments on link prediction, the ROC is reported

The result of CAN is omited because CAN takes the feature of nodes as sparse matrix with only "0" and "1", so it fails on some datasets like PubMed.

ROC Cora CiteSeer PubMed Photo Computer CS Physics Mean
BGRL 95.23 94.71 97.02 95.57 96.42 95.06 92.02 95.15
MaskGAE 94.82 92.78 92.79 96.93 96.57 96.69 97.15 95.39
AFGRL 95.29 94.69 96.54 82.75 75.99 94.81 94.52 90.66
COSTA 88.39 86.14 94.37 79.06 82.62 88.55 87.26 86.63
DGI 95.48 96.75 97.37 83.97 84.91 84.57 92.79 90.83
MVGRL 93.33 88.66 95.89 69.58 92.37 91.45 87.92 88.46
GRACE 80.94 83.21 97.11 85.24 83.36 87.67 84.57 86.01
GCA 81.46 84.81 94.2 70.02 89.92 84.35 85.75 84.36
CCA-SSG 95.77 94.29 98.09 97.97 86.6 83.34 98.6 93.52
CAN 93.67 94.56 - 97 96.03 - - -
SIG-VAE 94.1 92.88 85.89 94.98 91.14 95.26 96.47 92.96
GraphMAE 93.79 94.06 92.75 70.87 67.68 95.58 95.95 87.24
SEEGERA 95.97 92.31 84.88 98.02 98.77 97.87 97.24 95.01
GCLFormer 96.91 97.53 89.52 92.4 95.97 98.88 99.19 95.77

Experiments on link prediction, the AP is reported

The result of CAN is omited because CAN takes the feature of nodes as sparse matrix with only "0" and "1", so it fails on some datasets like PubMed.

AP Cora CiteSeer PubMed Photo Computer CS Physics Mean
BGRL 92.15 90.24 94.61 92.56 95.06 91.53 89.47 92.23
MaskGAE 94.22 92.15 92.71 92.74 95.37 96.8 97.18 94.45
AFGRL 96.34 95.77 96.28 80.87 72.98 94.97 95.32 90.36
COSTA 79.98 87.38 93.19 78.16 82.99 90.34 91.28 86.19
MVGRL 92.95 89.37 95.53 63.43 91.73 89.14 86.47 86.95
GRACE 82.57 81.95 96.77 83.17 83.39 86.84 83.39 85.44
GCA 80.87 81.93 93.31 65.17 89.5 83.24 82.86 82.41
CCA-SSG 96.33 95.33 97.99 97.71 86.97 76.31 98.57 92.74
CAN 94.49 95.49 - 96.68 95.96 - - -
SIG-VAE 94.79 94.21 85.02 94.53 91.23 94.93 96.28 93.00
GraphMAE 94.01 94.78 92.26 67.96 61.05 94.6 95.22 85.70
SEEGERA 96.65 93.79 83.96 97.69 98.81 97.12 98.12 95.16
GCLFormer 97.14 97.19 89.27 90.52 95.97 98.07 99.4 95.37

In general, Contrastive based methods are more good at node level task and generative based methods excels at link prediction tasks because generative based methods try to reconstruct the topology while contrastive learning focus on the similarity between positive samples. But GCLFormer utalize the attention calculation, so to some extent, we also try to reconstruct the topology, so GCLFormer also performs well on link prediction tasks.

experiments's People

Contributors

somebodyhh1 avatar

Watchers

avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.