Giter VIP home page Giter VIP logo

gpsp's Introduction

GPSP

This is the code and supplementary materials for the paper appeared in WWW2018, GPSP: Graph Partition and Space Projection based approach for Heterogeneous Network Embedding.

The data and embeddings are aviable at https://drive.google.com/open?id=1PFp1E0O4I2LbitPo4_SV_0VP5hs2Z5gp.

Specification

GPSP-code

  • SubFolder preprocessing : preprocessing steps for the data
  • SubFolder pte : implmentation of pte
  • SubFolder space projection is the space projection step
  • SubFolder node classification : codes for nodes classification
  • SubFolder node clustering : codes for nodes clustering
  • SubFolder visulization : codes for modifying data for the embedding projector

img

  • Algorithm and visulazation images

Development Environment

  • OS: Ubuntu 16.04 LTS
  • Language: Python 3.5.2
  • CPU: Intel® Core™ i7-5820K CPU @ 3.30GHz × 12
  • RAM: 32GB
  • Libraries:
    • numpy 1.13.1
    • pandas 0.21.1
    • NLTK 3.2.5
    • scikit-learn 0.18.1

Algorithm

  • The followings are the summary of GPSP algorithm

alt text

alt text

alt text

Experimental Details

Visulization using Tensorflow Projector (2D t-sne)

  • The learned embeddings are fed into Tensorflow Projector using t-sne.
  • Below is the visulization of GPSP-DeepWalk.
  • Eight groups are Computing Systems, Theoretical Computer Science, Computer Networks & Wireless Communication, Computer Graphics, Human Computer Interaction, Computational Linguistics, Computer Vision & Pattern Recognition, Databases & Information Systems.

alt text

  • GPSPL == GPSP-LINE
  • GPSPD == GPSP-DeepWalk

Multi-label classification results (Micro-F1)

      Method         10%       20%       30%       40%       50%       60%       70%       80%       90%    
LINE-1st 0.7003 0.7069 0.7081 0.7087 0.7087 0.7084 0.7079 0.7087 0.7079
LINE-2nd 0.6436 0.6446 0.6457 0.6462 0.6463 0.6458 0.6456 0.6450 0.6470
LINE-1st+2nd 0.7062 0.7064 0.7067 0.7075 0.7074 0.7077 0.7062 0.7072 0.7075
GPSPL-author 1st 0.6390 0.6420 0.6430 0.6436 0.6439 0.6432 0.6426 0.6448 0.6455
GPSPL-author 2nd 0.6162 0.6179 0.6184 0.6186 0.6181 0.6181 0.6183 0.6199 0.6212
GPSPL-author 1st+2nd 0.6487 0.6509 0.6515 0.6519 0.6522 0.6515 0.6519 0.6534 0.6540
GPSPL-paper 1st 0.7118 0.7148 0.7136 0.7156 0.7167 0.7127 0.7219 0.7206 0.7227
GPSPL-paper 2nd 0.6532 0.6546 0.6553 0.6554 0.6546 0.6540 0.6552 0.6521 0.6565
GPSPL-paper 1st+2nd 0.7235 0.7247 0.7247 0.7252 0.7256 0.7250 0.7262 0.7256 0.7267
PTE 0.7122 0.7125 0.7129 0.7135 0.7133 0.7138 0.7140 0.7135 0.7138
metapath2vec 0.6546 0.6547 0.6549 0.6550 0.6547 0.6551 0.6552 0.6537 0.6529
metapath2vec++ 0.6692 0.6687 0.6681 0.6679 0.6676 0.6678 0.6677 0.6658 0.6651
Deepwalk 0.6992 0.6998 0.7010 0.7008 0.6992 0.6988 0.6986 0.6964 0.6988
GPSPD-author 0.5919 0.5936 0.5950 0.5968 0.5963 0.5993 0.5974 0.5995 0.5980
GPSPD-paper 0.7010 0.7011 0.7016 0.7019 0.7021 0.7020 0.7018 0.7023 0.7020
GPSPD 0.7275 0.7304 0.7318 0.7330 0.7324 0.7328 0.7320 0.7331 0.7318
GPSPL 1st 0.7344 0.7378 0.7397 0.7396 0.7391 0.7401 0.7410 0.7425 0.7388
GPSPL 2nd 0.7121 0.7128 0.7141 0.7130 0.7148 0.7146 0.7137 0.7145 0.7159
GPSPL 1st+2nd 0.7512 0.7540 0.7557 0.7564 0.7564 0.7558 0.7554 0.7574 0.7552

Multi-label classification results (Macro-F1)

Method 10% 20% 30% 40% 50% 60% 70% 80% 90%
LINE-1st 0.6996 0.7050 0.7061 0.7069 0.7067 0.7062 0.7056 0.7063 0.7059
LINE-2nd 0.6389 0.6400 0.6413 0.6417 0.6419 0.6415 0.6409 0.6403 0.6426
LINE-1st+2nd 0.7032 0.7034 0.7036 0.7046 0.7043 0.7049 0.7035 0.7044 0.7036
GPSPL-author 1st 0.6399 0.6427 0.6434 0.6439 0.6438 0.6436 0.6424 0.6451 0.6451
GPSPL-author 2nd 0.6119 0.6136 0.6141 0.6143 0.6140 0.6138 0.6138 0.6162 0.6169
GPSPL-author 1st+2nd 0.6477 0.6498 0.6506 0.6507 0.6508 0.6501 0.6506 0.6529 0.6528
GPSPL-paper 1st 0.7087 0.7112 0.7099 0.7120 0.7130 0.7083 0.7198 0.7177 0.7211
GPSPL-paper 2nd 0.6557 0.6574 0.6580 0.6582 0.6571 0.6570 0.6578 0.6550 0.6591
GPSPL-paper 1st+2nd 0.7212 0.7226 0.7226 0.7230 0.7231 0.7229 0.7243 0.7232 0.7251
PTE 0.7089 0.7093 0.7094 0.7098 0.7101 0.7104 0.7090 0.7099 0.7094
metapath2vec 0.6307 0.6310 0.6313 0.6317 0.6322 0.6325 0.6328 0.6313 0.6301
metapath2vec++ 0.6478 0.6475 0.6473 0.6477 0.6478 0.6474 0.6473 0.6456 0.6445
Deepwalk 0.6964 0.6969 0.6982 0.6981 0.6965 0.6964 0.6963 0.6937 0.6961
GPSPD-author 0.5872 0.5887 0.5912 0.5922 0.5912 0.5977 0.5941 0.5971 0.5944
GPSPD-paper 0.7012 0.7015 0.7018 0.7020 0.7022 0.7021 0.7018 0.7023 0.7016
GPSPD 0.7253 0.7280 0.7290 0.7300 0.7298 0.7302 0.7295 0.7306 0.7289
GPSPL-1st 0.7318 0.7356 0.7369 0.7369 0.7361 0.7374 0.7388 0.7402 0.7364
GPSPL-2nd 0.7111 0.7117 0.7132 0.7119 0.7139 0.7137 0.7130 0.7136 0.7155
GPSPL-1st+2nd 0.7482 0.7513 0.7527 0.7534 0.7534 0.7529 0.7526 0.7544 0.7522

Node clustering results (NMI)

Method(Proximity) LINE PTE GPSPL-author GPSPL-paper GPSPL metapath2v metapath2v++ Deepwalk GPSPD-author GPSPD-paper GPSPD
1-st order(local) 0.3015 NA 0.2609 0.0447 0.1049 NA NA NA NA NA NA
2-nd order (global) 0.2529 0.2634 0.2505 0.2403 0.3118 0.2403 0.2473 0.2873 0.1681 0.3392 0.3555
1st+2nd order 0.2516 NA 0.2607 0.1738 0.1894 NA NA NA NA NA NA

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.