Research map / Time series / Long-term forecasting
AutoMixer for Improved Multivariate Time-Series Forecasting on Business and IT Observability Data
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 AutoMixer, which pretrains an RNN autoencoder to compress channels of BizITObs multivariate time series and then finetunes it end-to-end with TSMixer for Biz-KPI forecasting.
Problem
Forecasting business KPIs from multivariate BizITObs data, where noisy inter-channel interactions between Biz-KPIs and IT events degrade existing multivariate forecasters.
Approach
Use an autoencoder pretrained to compress the channel dimension (not hidden features) to prune noisy cross-channel correlations, then plug its encoder/decoder around a TSMixer backbone and finetune end-to-end as a foundation-model workflow.
Key results
- AutoMixer reduces Biz-KPI forecasting MSE by about 11% on average versus the best SOTA (TSMixer variants) and 15% versus the best baseline, with roughly 1-2X PCC gains over baselines.
- Pretraining the autoencoder reduces MSE by 2.3% on average compared to random initialization.
- On downstream tasks, AutoMixer improves IT event forecasting MSE by 3.2% and IT event classification accuracy about 2X over TSMixer-CC (Application and Service).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AE pretraining instead of randomly initialized AE weights (ablation) | AutoMixer GRU CC · Application | MSE: 0.0956 -> 0.0937 | Helped |
| Use AE pretraining instead of randomly initialized AE weights (ablation) | AutoMixer GRU CC · SO | MSE: 0.5551 -> 0.5697 | Hurt |
| Use AE pretraining instead of randomly initialized AE weights (ablation) | AutoMixer GRU CC · Service | MSE: 0.1618 -> 0.1494 | Helped |
| Use AE pretraining instead of randomly initialized AE weights (ablation) | AutoMixer GRU CC · L2C | MSE: 0.079 -> 0.0772 | Helped |
| Vary channel compression ratio (ablation) | AutoMixer GRU CC · Application | MSE: 0.0935; 0.1074; 0.0937; 0.0731 | Mixed |
| Vary channel compression ratio (ablation) | AutoMixer GRU CC · SO | MSE: 0.5697; 0.5715; 0.5708; 0.5598 | Mixed |
| Vary channel compression ratio (ablation) | AutoMixer GRU CC · Service | MSE: 0.1677; 0.1463; 0.1494; 0.1485 | Mixed |
| Vary channel compression ratio (ablation) | AutoMixer GRU CC · L2C | MSE: 0.0819; 0.0772 | Mixed |
Surprising findings
- TSMixer's strong trend capture limits PCC gains, so AutoMixer improves MSE notably but not PCC against SOTAs.
- Pretraining gives only a small MSE gain (2.3%), and on the SO dataset the non-pretrained variant was better.
Limitations
- Gains over TSMixer on PCC are negligible; improvements are mostly in MSE.
- Pretraining is per-dataset rather than cross-dataset, limiting the foundation-model claim.
- Best channel compression ratio and variant must be chosen per dataset on validation data.
Experimental setup
Datasets: Service (Stan's Robot Shop, synthesized) Application (synthesized) L2C (Lead-to-Cash, synthesized) SO (Sales Order, synthesized)
Benchmarks: Service Application L2C SO
Metrics: MSE PCC accuracy
Baselines: TSMixer TSMixer-CC PatchTST FEDFormer DLinear AGCRN TFT XGBoost VAR Recurrent GRU DecoderMLP
Builds on: TSMixer PatchTST DLinear FEDFormer
Closest papers by meaning
- SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion (NeurIPS 2024 · similarity 0.83)
- TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · similarity 0.83)
- Time-IMM: A Dataset and Benchmark for Irregular Multimodal Multivariate Time Series (NeurIPS 2025 · similarity 0.83)
- HDMixer: Hierarchical Dependency with Extendable Patch for Multivariate Time Series Forecasting (AAAI 2024 · similarity 0.83)
- WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.83)
- xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories (NeurIPS 2025 · similarity 0.83)
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.