Reproduce “Edit-Based Flow Matching for Temporal Point Processes” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

David Lüdke, Marten Lienen, Marcel Kollovieh, Stephan Günnemann

EditPP generates temporal point-process sequences by transforming noise through insertions, deletions, and local substitutions, rather than predicting events one at a time. Its continuous-time flow model uses constrained, order-preserving edits. Across evaluated datasets, unconditional samples require an average of 199.65 edit operations, compared with 234.52 for PSDIFF; this is not a measure of runtime, and quality results are mixed.

Key insights

  • Unifying discrete Edit Flow concepts with continuous-time temporal point processes through a continuous-time Markov Chain enables a flexible and efficient generative framework for event sequences.
  • The auxiliary alignment space is crucial for making the learning of complex continuous-time edit rates tractable, allowing the model to effectively match learned rates to ground-truth dynamics.
  • Incorporating a 'substitution' edit operation significantly reduces the total number of operations required for sequence transformation and improves generative efficiency compared to only using insertions and deletions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.