MambaTech: A Hybrid Mamba–Transformer Architecture with Adaptive Gating for Science and Technology Innovation Time Series Analysis and Pattern Recognition

Quantitative forecasting and lifecycle pattern recognition of science and technology (S&T) innovation activities are essential for evidence-based technology management, yet modeling long-range dependencies in multivariate S&T time series while capturing cross-variate interactions remains a fundamental challenge. Existing Transformer-based forecasters suffer from quadratic complexity that limits effective lookback windows, while state space models typically process each variate independently, ignoring inter-variate coupling among technology fields. We propose MambaTech, a problem-specific hybrid architecture for patent innovation time series that explicitly decomposes intra-domain temporal dynamics from inter-domain technology interactions. The architecture fuses a Mamba-based selective state space branch for O(n)-complexity temporal modeling with a Transformer cross-attention branch for inter-variate association modeling, unified through a learnable domain-conditioned adaptive gating mechanism. An auxiliary classification head regularizes the shared representation by performing lifecycle stage recognition across technology domains. Evaluated on patent filing time series spanning eight CPC sections over 180 months (Harvard USPTO Patent Dataset, 4.5M records) and a 632-variate subclass setting, with additional cross-domain assessment on public benchmarks (Electricity, Traffic, Weather), MambaTech achieves one-step MAPEs of 3.71% (coarse) and 6.24% (fine), reducing errors by 31.4% over Transformer and 52.2% over ARIMA baselines (averaged over five random seeds). At a 12-step horizon, the degradation rate is only +0.48%/step versus +0.89% for the best competing method. The gating fusion yields a 12.8% improvement over naïve concatenation, and the lifecycle classification head attains a macro-F1 of 0.915 across 632 technology domains, with statistically significant improvements under paired bootstrap and Diebold–Mariano tests. Ablation studies confirm the complementary branch contributions, and the learned cross-attention weights reveal technology association patterns consistent with known S&T innovation spillover theory and are validated against external co-classification networks. The originality of MambaTech lies in its task-specific hybridization and analytical framework for patent innovation time series, rather than in proposing a fundamentally new sequence modeling primitive.