Zhixuan Jia

Dr. Zhixuan Jia

orcid

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
4 September 2025 in Entropy
Verified
Journal Article
Open access
PDF file
ZJZhixuan Jia
Zhixuan Jia
CZChengcheng Zhang
Chengcheng Zhang
Time series models are widely used to examine temporal dynamics and uncover patterns across diverse fields. A commonly employed approach for modeling such data is the (Vector) Autoregressive (AR/VAR) model, in which each variable is represented as a linear combination of its own and others’ lagged values. However, the traditional (V)AR framework relies on the key assumption of stationarity, that autoregressive coefficients remain constant over time, which is often violated in practice, especially in systems affected by structural breaks, seasonal fluctuations, or evolving causal mechanisms. To overcome this limitation, Time-Varying (Vector) Autoregressive (TV-AR/TV-VAR) models have been developed, enabling model parameters to evolve over time and thus better capturing non-stationary behavior. Conventional approaches to estimating such models, including generalized additive modeling and kernel smoothing techniques, often require strong assumptions about basis functions, which can restrict their flexibility and applicability. To address these challenges, we introduce a novel framework that leverages physics-informed neural networks (PINN) to model TV-AR/TV-VAR processes. The proposed method extends the PINN framework to time series analysis by reducing reliance on explicitly defined physical structures, thereby broadening its applicability. Its effectiveness is validated through simulations on synthetic data and an empirical study of real-world health-related time series.

23 October 2024 in Mathematics
Verified
Journal Article
Open access
PDF file
ZJZhixuan Jia
Zhixuan Jia
WTWenfang Tian
Wenfang Tian
WLWang Li
Wang Li
KSKai Song
Kai Song
FWFuxin Wang
Fuxin Wang
CRCongjing Ran
Congjing Ran
The patent serves as a vital component of scientific text, and over time, escalating competition has generated a substantial demand for patent analysis encompassing areas such as company strategy and legal services, necessitating fast, accurate, and easily applicable similarity estimators. At present, conducting natural language processing(NLP) on patent content, including titles, abstracts, etc., can serve as an effective method for estimating similarity. However, the traditional NLP approach has some disadvantages, such as the requirement for a huge amount of labeled data and poor explanation of deep-learning-based model internals, exacerbated by the high compression of patent content. On the other hand, most knowledge-based deep learning models require a vast amount of additional analysis results as training variables in similarity estimation, which are limited due to human participation in the analysis part. Thus, in this research, addressing these challenges, we introduce a novel estimator to enhance the transparency of similarity estimation. This approach integrates a patent’s content with international patent classification (IPC), leveraging bidirectional encoder representations from transformers (BERT), and non-negative matrix factorization (NMF). By integrating these techniques, we aim to improve knowledge discovery transparency in NLP across various IPC dimensions and incorporate more background knowledge into context similarity estimation. The experimental results demonstrate that our model is reliable, explainable, highly accurate, and practically usable.

9 July 2025 in Mathematics
Verified
Journal Article
Open access
PDF file
ZJZhixuan Jia
Zhixuan Jia
WLWang Li
Wang Li
YJYunlong Jiang
Yunlong Jiang
XLXingshen Liu
Xingshen Liu
Time series data are fundamental for analyzing temporal dynamics and patterns, enabling researchers and practitioners to model, forecast, and support decision-making across a wide range of domains, such as finance, climate science, environmental studies, and signal processing. In the context of high-dimensional time series, the Vector Autoregressive model (VAR) is widely used, wherein each variable is modeled as a linear combination of lagged values of all variables in the system. However, the traditional VAR framework relies on the assumption of stationarity, which states that the autoregressive coefficients remain constant over time. Unfortunately, this assumption often fails in practice, especially in systems subject to structural breaks or evolving temporal dynamics. The Time-Varying Vector Autoregressive (TV-VAR) model has been developed to address this limitation, allowing model parameters to vary over time and thereby offering greater flexibility in capturing non-stationary behavior. In this study, we propose an enhanced modeling approach for the TV-VAR framework by incorporating minimization solvers in generalized additive models and one-sided kernel smoothing techniques. The effectiveness of the proposed methodology is assessed using simulations based on non-homogeneous Markov chains, accompanied by a detailed discussion of its advantages and limitations. Finally, we illustrate the practical utility of our approach using an application to real-world financial data.

26 November 2025 in Fractal and Fractional
Verified
Journal Article
Open access
PDF file
ZJZhixuan Jia
Zhixuan Jia
NRNan Rao
Nan Rao
Time-series modeling plays a crucial role in analyzing and forecasting financial volatility. Classical approaches, such as the Autoregressive (AR) and Fractional Autoregressive (FAR) models, capture short-term linear dependencies and long-range correlations, respectively, but their reliance on fixed structures and stationarity assumptions limits their adaptability to evolving market dynamics. To overcome these limitations, this study introduces a Fractional Time-Varying Autoregressive (FTVAR) framework that allows model parameters to evolve smoothly over time, integrating long-memory effects with nonstationary behavior. The FTVAR process is examined through two complementary methods: a Generalized Additive Model (GAM) for interpretable estimation of time-varying coefficients, and a Physics-Informed Neural Network (PINN) that embeds system dynamics to enhance forecasting under complex conditions. Simulation studies demonstrate that the FTVAR model consistently outperforms conventional AR approaches, offering superior forecasting accuracy, robustness to nonstationarity, and a more comprehensive representation of evolving volatility structures. Empirical analyses on the S&P 500 and VIX indices further confirm the effectiveness and practical relevance of the proposed framework.

25 June 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
LXLiu Xingshen
Liu Xingshen
FWFuxin Wang
Fuxin Wang
FWFan Wang
Fan Wang
ZJZhixuan Jia
Zhixuan Jia
HWHaowei Wang
Haowei Wang
CRCongjing Ran
Congjing Ran
Bias handling in expert-augmented LLMs is not merely a matter of routing tasks to experts, it requires coordinating how biased cues are detected, how their social meanings are attributed, and how interventions are generated. However, existing parameter-efficient fine-tuning and mixture-of-experts approaches typically organize specialization around task labels, domains, or input-level routing signals, rather than explicitly modeling how evidence of bias should inform its interpretation and subsequent rewriting. This paper proposes CoPAMoE, a cognitive process-aware PEFT-MoE framework for bias understanding and mitigation. CoPAMoE organizes bias-related subtasks into three functional stages, Cue Perception, Attribution Reasoning, and Intervention Control. It combines stage-partitioned LoRA experts, shared experts, a task--stage routing prior, and an Adjacent Stage Gated Coordination (ASGC) module. The resulting architecture explicitly separates stage-specific adaptation from task-general bias signals and supports controlled information transfer between adjacent stages. Experiments on a unified benchmark built from CrowS-Pairs, StereoSet, and WinoBias show that CoPAMoE improves average performance over representative PEFT and MoE baselines on two backbone LLMs. Ablation, routing, representation, oracle, and efficiency analyses further suggest that the gains are associated with stage-aware specialization and cross-stage dependency transfer.

24 June 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
LXLiu Xingshen
Liu Xingshen
FWFuxin Wang
Fuxin Wang
FWFan Wang
Fan Wang
ZJZhixuan Jia
Zhixuan Jia
HWHaowei Wang
Haowei Wang
CRCongjing Ran
Congjing Ran
Bias handling in expert-augmented LLMs is not merely a matter of routing tasks to experts, it requires coordinating how biased cues are detected, how their social meanings are attributed, and how interventions are generated. However, existing parameter-efficient fine-tuning and mixture-of-experts approaches typically organize specialization around task labels, domains, or input-level routing signals, rather than explicitly modeling how evidence of bias should inform its interpretation and subsequent rewriting. This paper proposes CoPAMoE, a cognitive process-aware PEFT-MoE framework for bias understanding and mitigation. CoPAMoE organizes bias-related subtasks into three functional stages, Cue Perception, Attribution Reasoning, and Intervention Control. It combines stage-partitioned LoRA experts, shared experts, a task--stage routing prior, and an Adjacent Stage Gated Coordination (ASGC) module. The resulting architecture explicitly separates stage-specific adaptation from task-general bias signals and supports controlled information transfer between adjacent stages. Experiments on a unified benchmark built from CrowS-Pairs, StereoSet, and WinoBias show that CoPAMoE improves average performance over representative PEFT and MoE baselines on two backbone LLMs. Ablation, routing, representation, oracle, and efficiency analyses further suggest that the gains are associated with stage-aware specialization and cross-stage dependency transfer.