notesum.ai
Published at November 12FM-TS: Flow Matching for Time Series Generation
cs.LG
cs.AI
Released Date: November 12, 2024
Authors: Yang Hu1, Xiao Wang2, Lirong Wu3, Huatian Zhang4, Stan Z. Li3, Sheng Wang2, Tianlong Chen1
Aff.: 1University of North Carolina at Chapel Hill; 2University of Washington; 3Westlake University; 4University of Science and Technology of China

| Dataset | Length | FM-TS | Diffusion-TS | TimeGAN | TimeVAE | Diffwave | DiffTime | Cot-GAN | |
| ETTh | Discriminative (Lower Better) | 64 | 0.010±.004 | 0.106±.048 | 0.227±.078 | 0.171±.142 | 0.254±.074 | 0.150±.003 | 0.296±.348 |
| 128 | 0.040±.012 | 0.144±.060 | 0.188±.074 | 0.154±.087 | 0.274±.047 | 0.176±.015 | 0.451±.080 | ||
| 256 | 0.081±.022 | 0.060±.030 | 0.444±.056 | 0.178±.076 | 0.304±.068 | 0.243±.005 | 0.461±.010 | ||
| Predictive (Lower Better) | 64 | 0.115±.005 | 0.116±.000 | 0.132±.008 | 0.118±.004 | 0.133±.008 | 0.118±.004 | 0.135±.003 | |
| 128 | 0.104±.013 | 0.110±.003 | 0.153±.014 | 0.113±.005 | 0.129±.003 | 0.120±.008 | 0.126±.001 | ||
| 256 | 0.107±.005 | 0.109±.013 | 0.220±.008 | 0.110±.027 | 0.132±.001 | 0.118±.003 | 0.129±.000 | ||
| Context-FID (Lower Better) | 64 | 0.039±.003 | 0.631±.058 | 1.130±.102 | 0.827±.146 | 1.543±.153 | 1.279±.083 | 3.008±.277 | |
| 128 | 0.128±.007 | 0.787±.062 | 1.553±.169 | 1.062±.134 | 2.354±.170 | 2.554±.318 | 2.639±.427 | ||
| 256 | 0.302±.018 | 0.423±.038 | 5.872±.208 | 0.826±.093 | 2.899±.289 | 3.524±.830 | 4.075±.894 | ||
| Correlational (Lower Better) | 64 | 0.027±.015 | 0.082±.005 | 0.483±.019 | 0.067±.006 | 0.186±.008 | 0.094±.010 | 0.271±.007 | |
| 128 | 0.030±.011 | 0.088±.005 | 0.188±.006 | 0.054±.007 | 0.203±.006 | 0.113±.012 | 0.176±.006 | ||
| 256 | 0.025±.008 | 0.064±.007 | 0.522±.013 | 0.046±.007 | 0.199±.003 | 0.135±.006 | 0.222±.010 | ||
| Energy | Discriminative (Lower Better) | 64 | 0.131±.046 | 0.078±.021 | 0.498±.001 | 0.499±.000 | 0.497±.004 | 0.328±.031 | 0.499±.001 |
| 128 | 0.301±.013 | 0.143±.075 | 0.499±.001 | 0.499±.000 | 0.499±.001 | 0.396±.024 | 0.499±.001 | ||
| 256 | 0.404±.070 | 0.290±.123 | 0.499±.000 | 0.499±.000 | 0.499±.000 | 0.437±.095 | 0.498±.004 | ||
| Predictive (Lower Better) | 64 | 0.250±.009 | 0.249±.000 | 0.291±.003 | 0.302±.001 | 0.252±.001 | 0.252±.000 | 0.262±.002 | |
| 128 | 0.249±.001 | 0.247±.001 | 0.303±.002 | 0.318±.000 | 0.252±.000 | 0.251±.000 | 0.269±.002 | ||
| 256 | 0.247±.001 | 0.245±.001 | 0.351±.004 | 0.353±.003 | 0.251±.000 | 0.251±.000 | 0.275±.004 | ||
| Context-FID (Lower Better) | 64 | 0.058±.010 | 0.135±.017 | 1.230±.070 | 2.662±.087 | 2.697±.418 | 0.762±.157 | 1.824±.144 | |
| 128 | 0.100±..002 | 0.087±.019 | 2.535±.372 | 3.125±.106 | 5.552±.528 | 1.344±.131 | 1.822±.271 | ||
| 256 | 0.083±..011 | 0.126±.024 | 5.052±.831 | 3.768±.998 | 5.572±.584 | 4.735±.729 | 2.533±.467 | ||
| Correlational (Lower Better) | 64 | 0.534±.110 | 0.672±.035 | 3.668±.106 | 1.653±.208 | 6.847±.083 | 1.281±.218 | 3.319±.062 | |
| 128 | 0.521±.201 | 0.451±.079 | 4.790±.116 | 1.820±.329 | 6.663±.112 | 1.376±.201 | 3.713±.055 | ||
| 256 | 0.391±.146 | 0.361±.092 | 4.487±.214 | 1.279±.114 | 5.690±.102 | 1.800±.138 | 3.739±.089 | ||