Zijian Zeng

Zijian Zeng


Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

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

Representative Publications

View All Publications
4 March 2025 in Sustainability
Verified
Journal Article
Open access
PDF file
YZYong Zhang
Yong Zhang
WTWee Hoe Tan
Wee Hoe Tan
ZZZijian Zeng
Zijian Zeng
This paper introduces a novel hybrid forecasting model for tourism demand that combines Bidirectional Long Short-Term Memory (BiLSTM) and Transformer networks, addressing the challenge of capturing both short-term fluctuations and long-term trends in complex tourism data. Unlike traditional models, such as ARIMA, which often struggle with nonlinear patterns, our hybrid approach leverages the sequential learning capabilities of BiLSTM and the self-attention mechanism of the Transformer to effectively model intricate temporal dependencies. Our experiments on Thailand’s domestic tourism data showed that the hybrid model outperformed traditional methods and standalone deep learning models, where it achieved a 12% reduction in the RMSE, a 15% reduction in the MAE, and a 10% increase in the R2. This improved accuracy offers significant practical benefits for sustainable tourism, enabling policymakers and tourism managers to optimize resource allocation, anticipate peak season demand, and develop strategies to mitigate over-tourism. The model’s robustness and adaptability make it a valuable tool for data-driven decision-making in the tourism sector.

28 May 2025 in Journal of Circuits, Systems and Computers
Verified
Journal Article
Closed access
Xianwei LiXianwei Li
Xianwei Li
Bo WeiBo Wei
Bo Wei
ZZZijian Zeng
Zijian Zeng
XYXiaoying Yang
Xiaoying Yang
GCGuolong Chen
Guolong Chen

18 August 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
ZZZijian Zeng
Zijian Zeng
HLHoude Liu
Houde Liu
KRKuru Ratnavelu
Kuru Ratnavelu
For a real correlation matrix C, the self-conjugate specializa- tion of Chollet’s conjecture asserts per(C ◦ C) ≤ per(C)2. We determine the sharp constant in order five under the rank-two constraint: per(C ◦ C) ≤ 137 1440 per(C)2, C ∈ R5×5, C ⪰ 0, diag C = 1, rank C ≤ 2. Equality holds exactly when the five Gram lines form a regular pentagon in RP1, up to permutation and sign switching. The proof reduces the two permanents to elementary symmetric functions of five unimodular numbers, solves the remaining two-phase minimization, and ends with explicit sum-of-squares and positive-semidefinite Gram certificates. We then study the rank-three stratum. The matrix J2 ⊕ ( 1 −1/2 −1/2 −1/2 1 −1/2 −1/2 −1/2 1 ) has self-Chollet ratio 13/48. We prove that this value is sharp in two infinite structured rank-three families and that the displayed configura- tion is a strict local maximum in the full rank-three configuration space. We also prove Chollet’s inequality whenever |cij | ≤ 27/50 for i̸ = j, with the sharper estimate R(C) ≤ 621/800 at coherence 1/2. Finally, without a rank assumption, Chollet’s inequality is strict away from the identity whenever λmin(C) ≥ 3/4. The proposed global rank-three value 13/48 remains open. All finite algebraic certificates are supplied in exact arithmetic.

8 December 2023
Verified
Conference Paper
Open access
BYBoshen Yuan
Boshen Yuan
TLTian Li
Tian Li
YLYucheng Lin
Yucheng Lin
ZZZijian Zeng
Zijian Zeng
News topic classification is a crucial aspect of natural language processing, aiding in effective information organization and accessibility. This paper explores the unique challenges posed by classifying financial news, highlighting the intricacies of language, market dynamics, and the necessity for accurate categorization. Leveraging large-scale language models, exemplified by Baichuan2-7b, proves advantageous for news topic classification due to their ability to capture nuances, facilitate transfer learning, and handle extensive datasets efficiently. The study introduces an innovative approach involving the fine-tuning of Baichuan2-7b, coupled with the use of QLora for orchestrating the fine-tuning process. Experimental results on Twitter financial news datasets showcase the state-of-the-art performance achieved by our proposed model.

16 January 2026 in IEEE Transactions on Intelligent Transportation Systems
Verified
Journal Article
Closed access
HYHuiming Yang
Huiming Yang
WLWenzhuo Liu
Wenzhuo Liu
YQYicheng Qiao
Yicheng Qiao
LYLei Yang
Lei Yang
XZXianzhu Zeng
Xianzhu Zeng
KWKunfeng Wang
Kunfeng Wang
The sparse cross-modality detector offers more advantages than its counterpart, the Bird’s-Eye-View (BEV) detector, particularly in terms of adaptability for downstream tasks and computational cost savings. However, existing sparse detectors overlook the quality of token representation, leaving it with a sub-optimal foreground quality and limited performance. In this paper, we identify that the geometric structure preserved and the class distribution are the key to improving the performance of the sparse detector, and propose a Sparse Selector (SS). The core module of SS is Ray-Aware Supervision (RAS), which preserves rich geometric information during the training stage, and Class-Balanced Supervision, which adaptively reweights the salience of class semantics, ensuring that tokens associated with small objects are retained during token sampling. Thereby, outperforming other sparse multi-modal detectors in the representation of tokens. Additionally, we design Ray Positional Encoding (Ray PE) to address the distribution differences between the LiDAR modality and the image. Finally, we integrate the aforementioned module into an end-to-end sparse multi-modality detector, dubbed CrossRay3D. Experiments show that, on the challenging nuScenes benchmark, CrossRay3D achieves state-of-the-art performance with 72.4% mAP and 74.7% NDS, while running $1.84\times $ faster than other leading methods. Moreover, CrossRay3D demonstrates strong robustness even in scenarios where LiDAR or camera data are partially or entirely missing. The code is available on https://github.com/xuehaipiaoxiang/CrossRay3D

3 September 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
ZZZijian Zeng
Zijian Zeng
We derive exact restrictions inside the full minimal threecover branch of the current structure theory for a smallest empty-set-free height-five counterexample to Frankl's union-closed sets conjecture. If q0 is the size of the fibre with empty private support and q2 counts the excess members in the three two-private-coordinate fibres, then q0 + q2 ≥ 9. The proof combines deletion-block injections, exact source-capacity saturation, cross-fibre complement intersection, and height-three generation. Its last case uses a new near-extremal lemma: an intersectionclosed height-three family generated by five nonempty sets and having at least fifteen members is the empty set, five pairwise-disjoint generators, and nine or ten pairwise unions. The lemma is proved by a ground-independent reduction to B5. Twenty fail-closed exact checks, implemented twice at every proof-essential finite reduction, audit the boundary-eight case split and the Boolean classification. The result is strictly conditional: it does not eliminate the full three-cover branch, the four-cover branches, or the general height-five case. At the first residual size r = 10, for t = 27, the two independent bounds collapse the fibre parameters to the unique triple (q0, u, q2) = (4, 0, 6).