Research map / Probabilistic and scientific modeling / Molecules and proteins

A Variational Autoencoder for Neural Temporal Point Processes with Dynamic Latent Graphs

AAAI 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Neural point-process models · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VAETPP, a VAE with a sequential latent variable model that learns a dynamic dependency graph over event types across regularly spaced sub-intervals, decoding inter-event times with log-normal mixtures to improve neural temporal point process prediction.

Problem

Neural temporal point processes typically use all past events and learn static dependency graphs, failing to capture time-varying, periodic dependencies between event types for predicting inter-event times and event types.

Approach

Partition the sequence into sub-intervals assumed stationary within each, use a VAE with a sequential latent variable (prior and posterior over per-sub-interval edge types) to infer a dynamic dependency graph, and use it in a GRNN decoder with log-normal mixture inter-event time outputs so non-contributing past events are masked out.

Key results

• VAETPP achieves the lowest inter-event time NLL on all four datasets, e.g. -3.501 on MathOF and -5.952 on NYMVC versus THP at -2.531 and -1.889.
• VAETPP reports the lowest event time RMSE on all datasets, e.g. 0.569 on MathOF versus 0.673 for LogNormMix.
• VAETPP reports the highest event type accuracy on all datasets, e.g. 0.321 on MathOF and 0.301 on NYMVC, against RMTPP at 0.154 and 0.251.

Measured effects reported in the paper

Surprising findings

• The static-graph variant VAETPP (static) underperforms THP and LogNormMix on several datasets in event-type accuracy and NLL, not just the dynamic version, which the authors attribute to the value of dynamic graphs.

Limitations

• The number of sub-intervals K and the regular interval partition are specified a priori rather than learned.
• Evaluated on only four datasets with fixed sub-interval choices (e.g., three-hour and one-day windows).
• The static-graph variant result shows mixed performance, and no statistical significance tests are reported.

Experimental setup

Datasets: MathOF AskUbuntu SuperUser NYMVC

Benchmarks: MathOF AskUbuntu SuperUser NYMVC

Metrics: negative log-likelihood RMSE accuracy

Baselines: Exponential RMTPP FullyNN LogNormMix THP VAETPP (static)

Builds on: Neural relational inference Recurrent marked temporal point process Transformer Hawkes process Log-normal mixture TPP Variational autoencoder

Closest papers by meaning

• C-NTPP: Learning Cluster-Aware Neural Temporal Point Process (AAAI 2023 · similarity 0.89)
• A Variational Point Process Model for Social Event Sequences (AAAI 2020 · similarity 0.88)
• A Variational Auto-Encoder Model for Stochastic Point Processes (CVPR 2019 · similarity 0.88)
• User-Dependent Neural Sequence Models for Continuous-Time Event Data (NeurIPS 2020 · similarity 0.88)
• Learning Temporally Causal Latent Processes from General Temporal Data (ICLR 2022 · similarity 0.87)
• Interacting Diffusion Processes for Event Sequence Forecasting (ICML 2024 · similarity 0.87)