Fi-sLSTM-Mixer: Multivariate Time Series Forecasting with Fuzzy-Conditioned Scalar Memories

TMLR Paper9041 Authors

Abstract: Multivariate time series forecasting in real-world deployments must contend with noisy, uncertain, and shifting data conditions that expose a structural weakness in state-of-the-art recurrent architectures: their gates rely on deterministic pre-activations and cannot adapt to input reliability. We introduce Fi-sLSTM-Mixer, a Fuzzy-integrated sLSTM which augments xLSTM-Mixer with a fuzzy relevance value rt derived from an ITTTFL inference system and injected into the forget and output gate pre-activations via a zero-initialized projection $W_r$ . The normalized input gate is provably invariant to rt by construction, producing a clean separation between data-driven variate attention and reliability-driven memory modulation. A twelve-test mechanistic protocol across 12 benchmarks confirms that the model learns consistent, domain-interpretable routing policies, statistically significant on every dataset, without sacrificing the backbone’s representational capacity. Empirically, Fi-sLSTM-Mixer achieves 41 wins across 90 metric slots against five state-of-the-art baselines, with the largest gains on volatile industrial and high-dimensional streams where reliability signals matter most, at a cost of under 600 additional parameters and negligible training overhead.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Jacek_Cyranka1

Submission Number: 9041