A state space model for long-sequence modeling and outcome prediction of heterogeneous event data with application to sports analytics

Efficiently modeling large-scale heterogeneous long-sequence event data remains a core challenge in deep learning and time-series analysis. Such sequences span hundreds to thousands of time steps, exhibit heterogeneous event types and irregular temporal spacing, and contain pronounced long-range dependencies. Recurrent neural networks struggle with such dependencies due to sequential computation and vanishing gradients, while Transformer architectures incur quadratic computational and memory complexity with sequence length, creating severe efficiency bottlenecks for ultra-long sequences and low-latency inference. We propose a Selective State Space Model (SSM)-based framework for joint outcome classification and continuous regression over heterogeneous event sequences. It employs a unified event embedding layer for discrete and continuous features, a dual-stream spatio-temporal state space module that fuses the two feature streams via cross-stream gating, and a gated context fusion head that dynamically integrates deep hidden representations with real-time context features, maintaining linear computational complexity while enhancing sensitivity to local dynamics. The associativity of the selective scan mechanism enables efficient parallel training via prefix scanning and low-latency recurrent inference. We construct a large-scale benchmark dataset and conduct systematic comparisons against LSTM, GRU, Transformer, Informer, S4, and vanilla Mamba. The proposed model achieves 85.3% accuracy and 0.912 AUC on outcome classification, and 3.61 RMSE on regression, achieving the best predictive performance among the evaluated baselines. Under a controlled inference benchmark on an NVIDIA A100 80 GB GPU, HoopMamba reduces inference latency by approximately 71.9% and peak GPU memory consumption by approximately 53.1% relative to Transformer at a sequence length of T = 1000, while incurring moderate additional computational overhead compared with vanilla Mamba and S4. Ablation studies confirm the critical contributions of the dual-stream module and gated context fusion mechanism. These findings demonstrate a favorable accuracy–efficiency trade-off for heterogeneous long-sequence modeling in basketball play-by-play analytics and suggest potential applicability to other domains involving large-scale heterogeneous event sequences.

Authors

Institutions

Publication Details

Journal
Journal of King Saud University - Computer and Information Sciences
Published
2026-11-19
DOI
https://doi.org/10.1007/s44443-026-01436-5
Primary Topic
Time Series Analysis and Forecasting
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A state space model for long-sequence modeling and outcome prediction of heterogeneous event data with application to sports analytics

魏如松, Yanan Meng, Xiaoqin Zhang
Journal of King Saud University - Computer and Information Sciences
Time Series Analysis and Forecasting
article

A state space model for long-sequence modeling and outcome prediction of heterogeneous event data with application to sports analytics

魏如松, Yanan Meng, Xiaoqin Zhang
article en

Abstract

Efficiently modeling large-scale heterogeneous long-sequence event data remains a core challenge in deep learning and time-series analysis. Such sequences span hundreds to thousands of time steps, exhibit heterogeneous event types and irregular temporal spacing, and contain pronounced long-range dependencies. Recurrent neural networks struggle with such dependencies due to sequential computation and vanishing gradients, while Transformer architectures incur quadratic computational and memory complexity with sequence length, creating severe efficiency bottlenecks for ultra-long sequences and low-latency inference. We propose a Selective State Space Model (SSM)-based framework for joint outcome classification and continuous regression over heterogeneous event sequences. It employs a unified event embedding layer for discrete and continuous features, a dual-stream spatio-temporal state space module that fuses the two feature streams via cross-stream gating, and a gated context fusion head that dynamically integrates deep hidden representations with real-time context features, maintaining linear computational complexity while enhancing sensitivity to local dynamics. The associativity of the selective scan mechanism enables efficient parallel training via prefix scanning and low-latency recurrent inference. We construct a large-scale benchmark dataset and conduct systematic comparisons against LSTM, GRU, Transformer, Informer, S4, and vanilla Mamba. The proposed model achieves 85.3% accuracy and 0.912 AUC on outcome classification, and 3.61 RMSE on regression, achieving the best predictive performance among the evaluated baselines. Under a controlled inference benchmark on an NVIDIA A100 80 GB GPU, HoopMamba reduces inference latency by approximately 71.9% and peak GPU memory consumption by approximately 53.1% relative to Transformer at a sequence length of T = 1000, while incurring moderate additional computational overhead compared with vanilla Mamba and S4. Ablation studies confirm the critical contributions of the dual-stream module and gated context fusion mechanism. These findings demonstrate a favorable accuracy–efficiency trade-off for heterogeneous long-sequence modeling in basketball play-by-play analytics and suggest potential applicability to other domains involving large-scale heterogeneous event sequences.

Journal of King Saud University - Computer and Information SciencesVol. 38(10)
Shanxi Agricultural University (CN), Guangxi Technological College of Machinery and Electricity (CN)
Time Series Analysis and Forecasting
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon