Hello SHINE

Signed Heterogeneous Information Network Embedding

speaker : Kevin Xu

Jump on the bandwagon?

  • Follow the crowd

  • What do people say

  • Who am I

    • What should I do?

  • Nature Language Processing
  • Sentiment Analyse
  • Avoid to speak the irony

    • Google NLP

    source : https://cloud.google.com/natural-language/

    Heterogeneous Information

    In order to increse the prediction accuracy

    Sentiment network

    Positive or negative sentiment link for someone.

    e.g., Triangle love

    Social network

  • Node : user (pages, image, post...)

  • Edge : Be friends (like, comment, share...)

  • Field : info about node

  • We are in the same boat

    • source : https://developers.facebook.com/docs/graph-api?locale=zh_TW

    Profile network

  • Bipartite graph

  • Recommend by user's profile

    • source : https://www.researchgate.net/profile/Nemi_Rathore

    Target

    Use SHINE to predict the sentiment link from the user to celebrities.

  • Heterogeneous info
  • Cold start scenario

    • Autoencoder

  • Encoder

  • Decoder

  • Error backpropagation

  • Dimension reduction

    • source : http://curiousily.com/

    Sentiment extraction

    Extract user's sentiment toward celebrities by SO scores.

  • Lexicon ([hate], negative)
  • Normalized to [-1, 1]
  • Validate the effect

    • SHINE network

  • One-hot encoding

  • 4-layer hidden (Inputs : 1000 units)

  • Aggregation function : Concatenation (100 units)

  • Similarity : inner product

  • Optimizer : AdaGrad

    • source : https://arxiv.org/abs/1712.00732

    Loss function

  • Reconstruction

  • Penalty for loss

    • source : https://arxiv.org/abs/1712.00732

    Hyper-parameters

  • Cross-validation

  • Trade-off

    • source : https://arxiv.org/abs/1712.00732

    Datasets

  • 3 bilion tweets on Weibo
  • Define the celebrities
  • Weibo-STC (20% for balanced test set)
  • Wiki-RfA (validate)

    • Baselines

  • LINE
  • Node2vec
  • SDNE
  • FxG
  • LIBFM

    • Performance

    Link prediction

  • Accuracy

  • Micro-Fi (unbalanced)

    • source : https://arxiv.org/abs/1712.00732

    Performance

    Node recommendation

  • Precision@K

  • Recall@K (sensitive)

    • source : https://arxiv.org/abs/1712.00732

    Furthermore

  • Over-fitting

  • Application

    • source : https://towardsdatascience.com/@ardendertat

    Thanks