Yixin Ji

Yixin Ji

Auto-generated profile, based on publicly available information

Yixin Ji

College of Artificial Intelligence and the Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing University of Aeronautics and Astronautics, Nanjing, China

Research Focus & Metrics

Academic Career Length

7Years (2020-Present)

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

Representative Publications

View All Publications
1 January 2020 in Computer Modeling in Engineering & Sciences
Verified
Journal Article
Open access
ZJZhuqing Jiao
Zhuqing Jiao
YJYixin Ji
Yixin Ji
TJTingxuan Jiao
Tingxuan Jiao
Shuihua WangShuihua Wang
Shuihua Wang
Currently, functional connectomes constructed from neuroimaging data have emerged as a powerful tool in identifying brain disorders. If one brain disease just manifests as some cognitive dysfunction, it means that the disease may affect some local connectivity in the brain functional network. That is, there are functional abnormalities in the sub-network. Therefore, it is crucial to accurately identify them in pathological diagnosis. To solve these problems, we proposed a sub-network extraction method based on graph regularization nonnegative matrix factorization (GNMF). The dynamic functional networks of normal subjects and early mild cognitive impairment (eMCI) subjects were vectorized and the functional connection vectors (FCV) were assembled to aggregation matrices. Then GNMF was applied to factorize the aggregation matrix to get the base matrix, in which the column vectors were restored to a common sub-network and a distinctive sub-network, and visualization and statistical analysis were conducted on the two sub-networks, respectively. Experimental results demonstrated that, compared with other matrix factorization methods, the proposed method can more obviously reflect the similarity between the common subnetwork of eMCI subjects and normal subjects, as well as the difference between the distinctive sub-network of eMCI subjects and normal subjects, Therefore, the high-dimensional features in brain functional networks can be best represented locally in the lowdimensional space, which provides a new idea for studying brain functional connectomes.

23 April 2025 in IEEE Transactions on Consumer Electronics
Verified
Journal Article
Closed access
SLShengrong Li
Shengrong Li
QZQi Zhu
Qi Zhu
LSLiang Sun
Liang Sun
KMKai Ma
Kai Ma
YJYixin Ji
Yixin Ji
DZDaoqiang Zhang
Daoqiang Zhang
Embedding multi-modal brain network analysis technology into consumer electronics, such as smart wearables, helps enable early intelligent diagnosis of brain diseases. Recent studies confirm that the functional-structural coupling in certain regions is more tightly correlated than in others. However, existing multi-modal methods often directly fuse functional brain networks (FBN) and structural brain networks (SBN), ignoring the regional heterogeneity between them. Additionally, identity information encoded in brain networks may interfere with disease diagnosis. In this paper, we develop a multi-modal brain network fusion method with regional heterogeneity constraints, and design a feature decoupling module to alleviate disease-irrelevant information. Specifically, we first divide FBN and SBN into multiple subnetworks, and introduce penalty weights to reduce the communication cost between cross-modal brain regions within the subnetwork while increasing the cost between different subnetworks. Then, under the regional heterogeneity constraints, we adopt optimal transport to simulate the transfer of brain region hubness from FBN to SBN, thereby effectively integrating the complex cross-modal interactions. Furthermore, we design a feature decoupling module to suppress ineffective features and enhance the discrimination between modality-specific features and multi-modal features. Experimental results show that the proposed method has promising performance and can identify multi-modal biomarkers for brain disease diagnosis.

10 July 2024
Verified
Conference Paper
Open access
CWChen Wei
Chen Wei
YJYixin Ji
Yixin Ji
ZCZeyuan Chen
Zeyuan Chen
JXJia Xu
Jia Xu
ZLZhongyi Liu
Zhongyi Liu
The search system aims to help users quickly find items according to queries they enter, which includes the retrieval and ranking modules. Traditional retrieval is a multi-stage process, including indexing and sorting, which cannot be optimized end-to-end. With the real data about mini-apps in the Alipay search, we find that many complex queries fail to display the relevant mini-apps, seriously threatening users' search experience. To address the challenges, we propose a Large Language Model-based Generative Retrieval (LLMGR) approach for retrieving mini-app candidates. The information of the mini-apps is encoded into the large model, and the title of the mini-app is directly generated. Through the online A/B test in Alipay search, LLMGR as a supplementary source has statistically significant improvements in the Click-Through Rate (CTR) of the search system compared to traditional methods. In this paper, we have deployed a novel retrieval method for the Alipay search system and demonstrated that generative retrieval methods based on LLM can improve the performance of search system, particularly for complex queries, which have an average increase of 0.2% in CTR.

21 October 2024
Verified
Conference Paper
Open access
HLHong Liu
Hong Liu
SGSaisai Gong
Saisai Gong
YJYixin Ji
Yixin Ji
KWKaixin Wu
Kaixin Wu
JXJia Xu
Jia Xu
JGJinjie Gu
Jinjie Gu
Relevance modeling plays a crucial role in e-commerce search engines, striving to identify the utmost pertinent items corresponding to a given search query. With the rapid advancement of pre-trained large language models (LLMs), recent endeavors have leveraged the capabilities of LLMs in relevance modeling, resulting in enhanced performance. This is usually done through the process of fine-tuning LLMs on specifically annotated datasets to determine the relevance between queries and items. However, there are two limitations when LLMs are naively employed for relevance modeling through fine-tuning and inference. First, it is not inherently efficient for performing nuanced tasks beyond simple yes or no answers, such as assessing search relevance. It may therefore tend to be overconfident and struggle to distinguish fine-grained degrees of relevance (e.g., strong relevance, weak relevance, irrelevance) used in search engines. Second, it exhibits significant performance degradation when confronted with data distribution shift in real-world scenarios. In this paper, we propose a novel Distribution-Aware Robust Learning framework (DaRL) for relevance modeling in Alipay Search. Specifically, we design an effective loss function to enhance the discriminability of LLM-based relevance modeling across various fine-grained degrees of query-item relevance. To improve the generalizability of LLM-based relevance modeling, we first propose the Distribution-Aware Sample Augmentation (DASA) module. This module utilizes out-of-distribution (OOD) detection techniques to actively select appropriate samples that are not well covered by the original training set for model fine-tuning. Furthermore, we adopt a multi-stage fine-tuning strategy to simultaneously improve in-distribution (ID) and OOD performance, bridging the performance gap between them. DaRL has been deployed online to serve the Alipay's insurance product search. Both offline experiments on real-world industry data and online A/B testing show that DaRL effectively improves the performance of relevance modeling.

13 March 2025 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
SLShengrong Li
Shengrong Li
QZQi Zhu
Qi Zhu
DGDonghai Guan
Donghai Guan
BSBo Shen
Bo Shen
LZLi Zhang
Li Zhang
DZDaoqiang Zhang
Daoqiang Zhang
Functional magnetic resonance imaging is a non-invasive neuroimaging technique capable of recording dynamic brain activity, which can be used to construct dynamic functional brain network (DFBN). DFBN adeptly captures intricate spatio-temporal characteristics of brain connectivities, and this proficiency is increasingly recognized in the field of brain disease diagnosis. However, the existing DFBN analysis methods predominantly focus on capturing the topological information between brain regions within individual time windows, often neglecting the critical long-interval temporal interactions that extend across these windows. Moreover, most of them extract temporal and spatial features independently, which hinders the effective mining of the complex spatio-temporal topological coupling in DFBNs. To tackle these problems, we propose a long-interval interactive graph convolutions method that collaboratively captures dynamic spatio-temporal topological features from DFBNs. Specifically, we first capture the brain networks’ long-interval communication patterns through the Hawkes process, which comprehensively exploits the impact of brain connectivity topological associations across different time windows. Then, we develop an interactive graph convolution framework to collaboratively extract spatio-temporal features in DFBNs, which overcomes the shortcomings of existing spatio-temporal models that struggle to characterize complex dynamic topological coupling. Finally, we design a hypergraph-based spatial information augmentation module to further extract the high-order dependencies among brain regions. Numerous experimental results demonstrate that the proposed method is superior to several state-of-the-art methods, and can provide effective biomarkers for brain disease diagnosis.

23 March 2026 in IEEE Transactions on Neural Systems and Rehabilitation Engineering
Verified
Journal Article
Open access
YJYixin Ji
Yixin Ji
VCVince Calhoun
Vince Calhoun
QZQi Zhu
Qi Zhu
ZXZhengwang Xia
Zhengwang Xia
JZJin Zhang
Jin Zhang
DZDaoqiang Zhang
Daoqiang Zhang
Brain functional networks (BFNs) derived from multi-site fMRI data have been widely explored using transformer-based models to extract discriminative connectivity features for the diagnosis of psychiatric disorders, such as schizophrenia (SZ). However, existing transformer-based methods ignore physiological priors like the scale-free property, causing the attention to fail to capture the intrinsic topology of BFNs. Additionally, multisite heterogeneity makes source-free domain adaptation (DA) essential in clinical practice where data sharing is restricted. However, existing methods mainly rely on heuristic data augmentations or pseudo-labeling, without leveraging the intrinsic inter-regional dependencies, resulting in poor robustness to cross-site variability. To overcome these challenges, we proposed a source-free DA framework based on a scale-free transformer encoder for SZ classification. The transformer encoder was pre-trained on labeled source domains to capture discriminative connectivity patterns while integrating a scale-free prior to bias the attention toward hub nodes. The pretrained encoder and classifier were used to initialize the target model, where the encoder learned latent representations derived from inter-regional interactions for causal graph construction. To enhance robustness, the causal structure was perturbed via random permutation and counterfactual interventions, while entropy minimization jointly optimized the encoder and predictor to learn domain-invariant representations. Results showed that our method outperformed the other 4 transformer, 7 DA, 6 multi-site and 6 state-of-the-art methods across two SZ datasets (87.18%±0.91% and 88.39%±0.13%). Ablation results highlighted the contributions of the causal, permutation, counterfactual, and entropy minimization constraints to the performance improvement. Furthermore, the identified discriminative temporal regions provided insights into the dysfunctional neural-mechanisms in SZ.