Loading...

Representation Learning for Dynamic Graphs

Loghmani, Erfan | 2021

449 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 53837 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Fazli, Mohammad Amin
  7. Abstract:
  8. Representation learning methods on graphs have enabled using machine learning methods on graphs' discrete structure by transferring them to a continuous domain. As graphs' structures are not always static and may evolve through time, dynamic representation learning methods have recently gained scholars' attention. Several methods have been proposed to enable the model to update the embeddings graph changes, or new interactions happen between nodes. These online methods could significantly reduce the learning time by refreshing the model as the changes occur, so we don't need to retrain the model with the complete graph information. Moreover, by using the temporal information of interactions, the model could learn richer patterns that may benefit the model's accuracy. T-batching is a technique used for training representation learning models on dynamic graphs. This technique reduces training time using the batching idea while preserving conditions that are vital in dynamic graph modeling. This research indicates a problem with the training loss function used with t-batching in a previous study. By mathematically analyzing the loss function, we show its downsides and suggest two other loss functions that do not suffer from the original function's problems. Then, we study the effect of loss function on the model's accuracy and optimization by designing several experiments on real and generated dynamic graphs. These experiments show that alternative loss functions could improve some evaluation metrics on the edge prediction task from %1.8 to %27.4 depending on some of the graph's structural properties
  9. Keywords:
  10. Representation Learning ; Loss Function ; Dynamic Graphs ; Deep Networks ; Deep Learning

 Digital Object List

 Bookmark

No TOC