IA03

这是一篇关于IA03的知识分享思维导图，作者给大家梳理归纳了十分详细的知识点，感兴趣的小伙伴可以点击参考。

Knowledge graphe

• Theoretical Overview
  • Introduction
    • Knowledge graph and the open world assumption
      • Closed world assumption
      • Open word assumption
    • ML on knowledges graphs and statistical retional learning
      • Link prediction triple classification
        • Rank problem
        • Information retrieval metrics
          • Informationretrievalmetrics
      • Collective node classification link-based clustering
      • Entity matching
    • Traditional Statistical Relational Learning
      • Logic programming
      • ILP
      • Rule Mining
      • Graphical Models
      • Limitations
    • Graph representation learning
      • Node Representation/Graph Feature based Methods
        • Node Representation/Graph Feature based Methods
      • GNNS
      • Knowledge graph embeddings
        • Automatic, supervised learning of embeddings
        • Projections of entites and relations into a continuous low-dimensional space
  • Anatomy of a KGE model
    • Knowledge graph
    • Scoring function for a triple
      • Factorization-based
        • Rescal
        • Distmult
        • ComplEx
      • Deeper
        • ConvE
        • ConvKB
      • Translation-based
        • RotateE
      • Other
    • Loss function
      • Pairwise margin-based hinge loss
      • Negative log-likelihood cross entropy
      • Binary cross entropy
      • Self adversarial
    • Optimization algorithm
    • Negatives generation strategy
      • Local closed world assumption
      • Training with synthetic negatives
        • Uniform sampling
        • Complete set
        • 1-n scoring
      • Training procedure and optimizer
        • Optimizer
        • Reciprocal triples
      • Model selection
        • Grid search
        • Random search
        • Quasi-random + bayesian
  • Evaluation protocol and metrics
    • The task
      • Link prediction
        • Learning to rank problem
          • Synthetic negatives
          • Local closed world assumption
        • Information retrieval metrics
        • No ground truth negatives in test set required
    • Evaluation metrics
      • Mean rank
      • Mean reciprocal rank
      • Hits@N
  • Advanced KGE topics
    • Calibration
      • Mistrust in model discoveries
      • Poor interpretability in high-stakes scenarios
      • How calibrate KGE
        • With ground truth negatives
        • With synthetic negatives
      • Is effective
        • Better than uncalibrated
        • More trustworthy and interpretable
    • Multimodal knowledge graph
    • Temporal knowledge graph
    • Uncertain knowledge graph
      • UKGE
      • Robustness
  • Open research questions
    • More expressive models
    • Support for multimodality
    • Robustness and interpretability
    • Better benchmarks
    • Beyond link prediction
    • Neuro-symbolic integration
• Applications
  • Industrial
    • Pharmaceutical
      • Time consuming and expensive process
      • Initial step(identification of gene) takes years
      • Highly dependent on the exp of the person
    • HR
      • Evolving at an extremely fast space
      • Roles becoming obsolete, companies are forced to lay off people due to automation
      • KGES function
        • Suggest new tech or tasks for career prog
        • Recommend similar roles
    • Products
      • KGEs can leverage relation between customers and products
      • KGEs can be used for recommend products and group customers
    • Food and beverage
      • Product reformulation
      • Adapting to consumer trends
• Software ecosystem
  • What is out there
    • OpenKE, Ampligraph, pytorch, etc …
  • Libraries comparison
    • Features
      • Models
      • Pre-trained models
      • Other
    • Scalability
    • Sofa reproduced
    • Software developement
      • Documentation
      • Tests
      • Good pratices
      • Code complexity
  • Which library should i use
    • Your task and time
    • Your exp
    • Library sports
    • Consider features