Research map / Time series / Causal and generative time series
Non-autoregressive Conditional Diffusion Models for Time Series Prediction
Lifeng Shen, James Kwok
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add future mixup to conditioning network (ablation) | TimeDiff · Caiso, Electricity, Exchange, ETTh1 (univariate) | MSE: ETTh1 0.182 -> 0.086 without AR; Caiso 0.170 -> 0.160 with AR only | Improved |
| add AR initialization to conditioning network (ablation) | TimeDiff · Caiso, Electricity, Exchange, ETTh1 (univariate) | MSE: e.g., ETTh1 0.182 -> 0.162 without future mixup; Electricity 0.340 -> 0.295 | Improved |
| soft-mixup (uniform mask) vs hard-mixup or segment-mixup (ablation) | future mixup · Caiso, Electricity, Exchange, ETTh1 | MSE: soft-mixup best overall; hard-mixup at tau=0.5 gives Caiso 0.118 vs 0.122 | Mixed |
| predict data x0 instead of noise epsilon (ablation) | TimeDiff · Caiso, Electricity, Exchange, ETTh1 | MSE: ETTh1 0.077 -> 0.066; Electricity 0.317 -> 0.232 | Improved |
| add future mixup and AR initialization to CSDI (ablation) | CSDI · ETTh1 and ETTm1 (univariate) | MSE: ETTh1 0.083 -> 0.075; ETTm1 0.050 -> 0.043 | Improved |
| add future mixup and AR initialization to SSSD (ablation) | SSSD · ETTh1 and ETTm1 (univariate) | MSE: ETTh1 0.097 -> 0.071; ETTm1 0.049 -> 0.040 | Improved |
| TimeDiff vs existing diffusion models (TimeGrad, CSDI, SSSD, D3VAE) | TimeDiff · univariate and multivariate benchmarks | MSE: TimeDiff best average rank 2.7 vs 9.4-11.6 for diffusion baselines | Improved |
| TimeDiff vs DLinear and FiLM (strong non-diffusion baselines) | TimeDiff · univariate and multivariate benchmarks | average rank: TimeDiff avg rank 2.7 (univariate) vs DLinear 4.8; multivariate 1.7 vs FiLM 3.2 | Mixed |
| TimeDiff inference with DPM-Solver (<20 denoising steps) | TimeDiff · ETTh1 inference time | inference time (ms): H=720: TimeDiff 34.6 vs TimeGrad 6724.1, CSDI 513.1, SSSD 2516.9 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 4 on most shared results.
- CSDI — CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation (NeurIPS 2021) · better on 4 of 4 shared results
- Depts — DEPTS: Deep Expansion Learning for Periodic Time Series Forecasting (ICLR 2022) · better on 0 of 1 shared result
- Informer — Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (AAAI 2021) · better on 1 of 1 shared result
- Autoformer — Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (NeurIPS 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation (ICML 2025 · similarity 0.90)
- Non-stationary Diffusion For Probabilistic Time Series Forecasting (ICML 2025 · similarity 0.89)
- Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models (ICLR 2025 · similarity 0.89)
- ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space (NeurIPS 2026 · similarity 0.89)
- Retrieval-Augmented Diffusion Models for Time Series Forecasting (NeurIPS 2024 · similarity 0.89)
- Auto-Regressive Moving Diffusion Models for Time Series Forecasting (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
