Pub Date : 2026-01-22DOI: 10.1016/j.ipm.2026.104637
Xiangjun Dong , Yicong Zhen , Ping Qiu , Jing Chi , Lei Guo , Wenpeng Lu , Long Zhao , Yongshun Gong , Yuhai Zhao
Mining high utility repeated negative sequential patterns (HURNSPs) from data streams is an important method for data stream analysis. However, the existing methods in this topic don’t consider negative events and repeated events, which results in weak decision-making effectiveness. So in this paper, we propose an effective algorithm DS_HURNSP for mining HURNSPs from data streams based on a sliding window model. First, we propose an effective utility-list prefix tree structure to store high utility repeated positive sequential patterns (HURPSPs). Second, we construct a utility mapping set based on a hash-table structure to enable rapid querying of HURPSPs information. Finally, we propose a two-stage computation method to compute the utility of high utility repeated negative sequential candidates (HURNSCs) by mapping them to the set of HURPSPs, avoiding rescanning database. Extensive experiments on six datasets show that the DS_HURNSP algorithm generates tens to thousands of times as many HURNSPs as the baseline method, and reduces the average runtime by more than half.
{"title":"DS_HURNSP: An effective method for mining high utility repeated negative sequential patterns from data streams","authors":"Xiangjun Dong , Yicong Zhen , Ping Qiu , Jing Chi , Lei Guo , Wenpeng Lu , Long Zhao , Yongshun Gong , Yuhai Zhao","doi":"10.1016/j.ipm.2026.104637","DOIUrl":"10.1016/j.ipm.2026.104637","url":null,"abstract":"<div><div>Mining high utility repeated negative sequential patterns (HURNSPs) from data streams is an important method for data stream analysis. However, the existing methods in this topic don’t consider negative events and repeated events, which results in weak decision-making effectiveness. So in this paper, we propose an effective algorithm DS_HURNSP for mining HURNSPs from data streams based on a sliding window model. First, we propose an effective utility-list prefix tree structure to store high utility repeated positive sequential patterns (HURPSPs). Second, we construct a utility mapping set based on a hash-table structure to enable rapid querying of HURPSPs information. Finally, we propose a two-stage computation method to compute the utility of high utility repeated negative sequential candidates (HURNSCs) by mapping them to the set of HURPSPs, avoiding rescanning database. Extensive experiments on six datasets show that the DS_HURNSP algorithm generates tens to thousands of times as many HURNSPs as the baseline method, and reduces the average runtime by more than half.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104637"},"PeriodicalIF":6.9,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023259","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-22DOI: 10.1016/j.ipm.2026.104627
Paerhati Tulajiang , Jinzhong Ning , Yuanyuan Sun , Liang Yang , Yuanyu Zhang , Kelaiti Xiao , Zhixing Lu , Yijia Zhang , Hongfei Lin
Multilingual named entity recognition (NER) is especially challenging in low-resource and typologically diverse languages, where translation drift, morphological variation, and noisy alignments degrade performance. Existing encoder-based methods often rely on dense attention or uniform alignment, which tends to propagate irrelevant signals across languages. We present SEGA, a lightweight and typology-aware framework that incorporates sparse guided attention to select auxiliary signals, alongside a weighted fusion layer that balances representations between cross-lingual and monolingual contexts. Unlike prior approaches, SEGA requires no parallel corpora and supports fully monolingual inference. We evaluate SEGA on six multilingual NER benchmarks spanning over 60 languages, including CoNLL, WikiANN, MasakhaNER 2.0, XTREME-40, WikiNEuRal, and MultiNERD. SEGA achieves new state-of-the-art results on five datasets, with absolute gains of up to +24.2 F1 over strong encoder baselines, and outperforming prompt-based large language models by up to +18.9 F1 in low-resource scenarios. Efficiency analyses show that SEGA adds only ∼ 30M parameters beyond a standard dual encoder, making it lightweight and deployable on a single GPU. Comprehensive ablation, visualization, and error analyses confirm that SEGA is robust to alignment noise, morphological complexity, and boundary ambiguity, offering a practical and scalable solution for real-world multilingual NER.
{"title":"SEGA: Selective cross-lingual representation via sparse guided attention for low-resource multilingual named entity recognition","authors":"Paerhati Tulajiang , Jinzhong Ning , Yuanyuan Sun , Liang Yang , Yuanyu Zhang , Kelaiti Xiao , Zhixing Lu , Yijia Zhang , Hongfei Lin","doi":"10.1016/j.ipm.2026.104627","DOIUrl":"10.1016/j.ipm.2026.104627","url":null,"abstract":"<div><div>Multilingual named entity recognition (NER) is especially challenging in low-resource and typologically diverse languages, where translation drift, morphological variation, and noisy alignments degrade performance. Existing encoder-based methods often rely on dense attention or uniform alignment, which tends to propagate irrelevant signals across languages. We present SEGA, a lightweight and typology-aware framework that incorporates sparse guided attention to select auxiliary signals, alongside a weighted fusion layer that balances representations between cross-lingual and monolingual contexts. Unlike prior approaches, SEGA requires no parallel corpora and supports fully monolingual inference. We evaluate SEGA on six multilingual NER benchmarks spanning over 60 languages, including CoNLL, WikiANN, MasakhaNER 2.0, XTREME-40, WikiNEuRal, and MultiNERD. SEGA achieves new state-of-the-art results on five datasets, with absolute gains of up to +24.2 F1 over strong encoder baselines, and outperforming prompt-based large language models by up to +18.9 F1 in low-resource scenarios. Efficiency analyses show that SEGA adds only ∼ 30M parameters beyond a standard dual encoder, making it lightweight and deployable on a single GPU. Comprehensive ablation, visualization, and error analyses confirm that SEGA is robust to alignment noise, morphological complexity, and boundary ambiguity, offering a practical and scalable solution for real-world multilingual NER.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104627"},"PeriodicalIF":6.9,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023254","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-22DOI: 10.1016/j.ipm.2026.104643
Xiaowen Cui , Xue Dong , Ping Qiu , Chuanhou Sun , Yuhai Zhao , Wenpeng Lu , Xiangjun Dong
Target pattern mining (TPM) aims to return sets of target patterns related to a user-queried target sequence. However, existing TPM research is confined to positive sequential patterns, overlooking negative sequential patterns, which results in limited decision support capabilities. Moreover, introducing negative sequential patterns faces challenges of low mining efficiency and limited pruning techniques. To address these issues, we propose an efficient target pattern mining algorithm based on negative sequential pattern, called TaNSP, to achieve TPM for negative sequence as the user-queried target sequence and output negative sequential patterns containing the target query sequence, while also supporting positive sequential patterns. Specifically, we propose a pruning strategy based on a triple bitmap to guide pattern generation and improve mining efficiency. Then, we propose a pruning strategy to address the limitations of pruning techniques when the negative sequence is the target query sequence. The experimental results on six datasets demonstrate that, compared to the baseline method, TaNSP can increase operational efficiency by more than twice, demonstrating excellent scalability and practicality.
{"title":"TaNSP: An efficient target pattern mining algorithm based on negative sequential pattern","authors":"Xiaowen Cui , Xue Dong , Ping Qiu , Chuanhou Sun , Yuhai Zhao , Wenpeng Lu , Xiangjun Dong","doi":"10.1016/j.ipm.2026.104643","DOIUrl":"10.1016/j.ipm.2026.104643","url":null,"abstract":"<div><div>Target pattern mining (TPM) aims to return sets of target patterns related to a user-queried target sequence. However, existing TPM research is confined to positive sequential patterns, overlooking negative sequential patterns, which results in limited decision support capabilities. Moreover, introducing negative sequential patterns faces challenges of low mining efficiency and limited pruning techniques. To address these issues, we propose an efficient <u>ta</u>rget pattern mining algorithm based on <u>n</u>egative <u>s</u>equential <u>p</u>attern, called TaNSP, to achieve TPM for negative sequence as the user-queried target sequence and output negative sequential patterns containing the target query sequence, while also supporting positive sequential patterns. Specifically, we propose a pruning strategy based on a triple bitmap to guide pattern generation and improve mining efficiency. Then, we propose a pruning strategy to address the limitations of pruning techniques when the negative sequence is the target query sequence. The experimental results on six datasets demonstrate that, compared to the baseline method, TaNSP can increase operational efficiency by more than twice, demonstrating excellent scalability and practicality.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104643"},"PeriodicalIF":6.9,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023309","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-21DOI: 10.1016/j.ipm.2026.104633
Wenbo Zhang , Xi Chen , Mingwei Wang , Rui Hou , Ning Li
The prediction of patient no-shows has emerged as a critical research topic in healthcare service management. To address the challenges of behavioral heterogeneity and factor diversity in patient no-shows, this study proposes a data-driven three-stage decision analytics method integrating personalized preference learning and explainable prediction. First, to capture patient no-show behavioral heterogeneity, we applied community detection algorithms to identify distinct patient subgroups, enabling subgroup-specific exploration of no-show patterns. Second, within each subgroup, we employ a preference learning model to quantify the relative importance of the influencing factors tailored to each group’s characteristics. Finally, we developed an XGBoost predictive model based on SHAP to achieve highly accurate and interpretable no-show probability predictions. Experimental results based on 2431 real outpatient appointment records from the Department of Ophthalmology at General Medical Harbin 242 Hospital showed that air quality emerged as a critical factor influencing patient attendance, with poorer air quality significantly increasing the likelihood of no-shows. In addition, the proposed PL-XGBoost-SHAP model achieved an average accuracy of 90.08%, precision of 92.36%, recall of 95.65% and F1 score of 93.96% across the four subgroups. These results showed an accuracy improvement of 2–12% compared with the five baseline models, demonstrating the scientific validity and feasibility of the proposed approach. Sensitivity analyses and statistical tests confirmed the robustness and generalizability of the proposed method. Consequently, this study offers significant practical implications for healthcare providers, enabling the design of personalized appointment reminders and effective allocate resources.
{"title":"From personalized learning to explainable prediction: A data-driven framework for patient no-shows","authors":"Wenbo Zhang , Xi Chen , Mingwei Wang , Rui Hou , Ning Li","doi":"10.1016/j.ipm.2026.104633","DOIUrl":"10.1016/j.ipm.2026.104633","url":null,"abstract":"<div><div>The prediction of patient no-shows has emerged as a critical research topic in healthcare service management. To address the challenges of behavioral heterogeneity and factor diversity in patient no-shows, this study proposes a data-driven three-stage decision analytics method integrating personalized preference learning and explainable prediction. First, to capture patient no-show behavioral heterogeneity, we applied community detection algorithms to identify distinct patient subgroups, enabling subgroup-specific exploration of no-show patterns. Second, within each subgroup, we employ a preference learning model to quantify the relative importance of the influencing factors tailored to each group’s characteristics. Finally, we developed an XGBoost predictive model based on SHAP to achieve highly accurate and interpretable no-show probability predictions. Experimental results based on 2431 real outpatient appointment records from the Department of Ophthalmology at General Medical Harbin 242 Hospital showed that air quality emerged as a critical factor influencing patient attendance, with poorer air quality significantly increasing the likelihood of no-shows. In addition, the proposed PL-XGBoost-SHAP model achieved an average accuracy of 90.08%, precision of 92.36%, recall of 95.65% and F1 score of 93.96% across the four subgroups. These results showed an accuracy improvement of 2–12% compared with the five baseline models, demonstrating the scientific validity and feasibility of the proposed approach. Sensitivity analyses and statistical tests confirmed the robustness and generalizability of the proposed method. Consequently, this study offers significant practical implications for healthcare providers, enabling the design of personalized appointment reminders and effective allocate resources.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104633"},"PeriodicalIF":6.9,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023257","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-21DOI: 10.1016/j.ipm.2026.104609
Yuan Tian, Shi Ying, Tiangang Li
Log parsing transforms unstructured log messages into structured formats, serving as a critical step for various log analysis tasks. Large language models (LLMs) have recently shown strong performance in this task. However, they tend to rely on their experiential knowledge as shortcuts, introducing bias and reducing parsing accuracy. To address this issue, we propose CausalLog, a lightweight and flexible debiasing framework for log parsing. CausalLog is inspired by the Structural Causal Model and the front-door adjustment principle. On this basis, counterfactual rewriting is implemented through tailored prompt engineering, aiming to mitigate biases without accessing LLM internals. In addition, n-gram statistics of log data are integrated as a bias-free reference for an adjustment score, which helps improve both parsing accuracy and interpretability. Experiments on public log datasets show that CausalLog outperforms state-of-the-art methods, providing observational evidence that it improves both log grouping and template extraction accuracy.
{"title":"CausalLog: Log parsing using LLMs with causal intervention for bias mitigation","authors":"Yuan Tian, Shi Ying, Tiangang Li","doi":"10.1016/j.ipm.2026.104609","DOIUrl":"10.1016/j.ipm.2026.104609","url":null,"abstract":"<div><div>Log parsing transforms unstructured log messages into structured formats, serving as a critical step for various log analysis tasks. Large language models (LLMs) have recently shown strong performance in this task. However, they tend to rely on their experiential knowledge as shortcuts, introducing bias and reducing parsing accuracy. To address this issue, we propose CausalLog, a lightweight and flexible debiasing framework for log parsing. CausalLog is inspired by the Structural Causal Model and the front-door adjustment principle. On this basis, counterfactual rewriting is implemented through tailored prompt engineering, aiming to mitigate biases without accessing LLM internals. In addition, n-gram statistics of log data are integrated as a bias-free reference for an adjustment score, which helps improve both parsing accuracy and interpretability. Experiments on public log datasets show that CausalLog outperforms state-of-the-art methods, providing observational evidence that it improves both log grouping and template extraction accuracy.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104609"},"PeriodicalIF":6.9,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023258","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-20DOI: 10.1016/j.ipm.2026.104634
Chengyu Li , Wenlong Yang , Meiling Li , Yang Wang
Scientific collaboration is a key driver of breakthroughs in contemporary science. While prior research has primarily focused on team diversity based on individual attributes, quantifying structural diversity based on collaborative relationships at the individual paper level remains underexplored. In this study, we apply neural embedding techniques to quantify structural diversity at the paper level and examine its association with research outcomes. Leveraging the node2vec and GraphSAGE algorithms, we embed each scientist into a dense vector space using the Microsoft Academic Graph dataset, covering 4,772,781 papers written by 5,162,397 scientists from 1995 to 2016. We show that our embedding model effectively classifies scientists across scientific domains and accurately predicts their primary scientific domains. Crucially, our analysis reveals that structural diversity is strongly associated with citation impact, novelty, and disruption at the individual paper level. Specifically, a 1SD increase in structural diversity is associated with a 7.6%, 24.7%, and 2.9% increase in citation impact, and the odds of publishing novel and disruptive papers, respectively. These findings are generalizable across multiple model specifications, time spans, different team sizes, and scientific domains. Finally, our analysis reveals that structural diversity exhibits the strongest correlation with citation impact, novelty, and disruption among all examined diversity measures. Our study highlights the importance of fostering structurally diverse collaborations and has policy implications for institutions, funders, and governments aiming to support impactful and innovative research.
{"title":"Neural embeddings of collaboration networks predict citation impact and innovation","authors":"Chengyu Li , Wenlong Yang , Meiling Li , Yang Wang","doi":"10.1016/j.ipm.2026.104634","DOIUrl":"10.1016/j.ipm.2026.104634","url":null,"abstract":"<div><div>Scientific collaboration is a key driver of breakthroughs in contemporary science. While prior research has primarily focused on team diversity based on individual attributes, quantifying structural diversity based on collaborative relationships at the individual paper level remains underexplored. In this study, we apply neural embedding techniques to quantify structural diversity at the paper level and examine its association with research outcomes. Leveraging the node2vec and GraphSAGE algorithms, we embed each scientist into a dense vector space using the Microsoft Academic Graph dataset, covering 4,772,781 papers written by 5,162,397 scientists from 1995 to 2016. We show that our embedding model effectively classifies scientists across scientific domains and accurately predicts their primary scientific domains. Crucially, our analysis reveals that structural diversity is strongly associated with citation impact, novelty, and disruption at the individual paper level. Specifically, a 1SD increase in structural diversity is associated with a 7.6%, 24.7%, and 2.9% increase in citation impact, and the odds of publishing novel and disruptive papers, respectively. These findings are generalizable across multiple model specifications, time spans, different team sizes, and scientific domains. Finally, our analysis reveals that structural diversity exhibits the strongest correlation with citation impact, novelty, and disruption among all examined diversity measures. Our study highlights the importance of fostering structurally diverse collaborations and has policy implications for institutions, funders, and governments aiming to support impactful and innovative research.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104634"},"PeriodicalIF":6.9,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023317","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-20DOI: 10.1016/j.ipm.2026.104635
Lihua Fu , Luo Jin , Yaokuang Li
Artificial intelligence (AI) is driving technological and industrial transformation, reshaping enterprise structure and innovation. Despite its importance, research lacks insight into AI’s impact on innovation ambidexterity and the underexplored role of risk-taking. Drawing on information processing theory, this study constructs a theoretical model to examine the impact of AI applications on enterprise exploration, exploitation, and innovation ambidexterity, as well as incorporates the level of risk-taking as a moderating variable. Using Stata 17 to perform panel data regression and a series of robustness tests, we analyze 1,908 firm-year observations from the information transmission, software, and information technology service industries, as well as the scientific research and technology service industries listed on the Shanghai and Shenzhen Stock Exchanges from 2012 to 2022. The empirical analyses propose that AI applications remarkably enhance innovation ambidexterity at the 1 % level. This positive effect is the most pronounced when risk-taking is moderate. This study extends information processing theory to the AI-enabled innovation context and further enriches its boundary conditions by introducing risk-taking as a nonlinear moderator. Managerially, the findings suggest that enterprises should calibrate risk-taking levels to complement AI deployment, enabling a balanced approach to exploration and exploitation.
{"title":"Artificial intelligence applications and innovation ambidexterity: The “inverted U-shaped” regulating effect of risk-taking","authors":"Lihua Fu , Luo Jin , Yaokuang Li","doi":"10.1016/j.ipm.2026.104635","DOIUrl":"10.1016/j.ipm.2026.104635","url":null,"abstract":"<div><div>Artificial intelligence (AI) is driving technological and industrial transformation, reshaping enterprise structure and innovation. Despite its importance, research lacks insight into AI’s impact on innovation ambidexterity and the underexplored role of risk-taking. Drawing on information processing theory, this study constructs a theoretical model to examine the impact of AI applications on enterprise exploration, exploitation, and innovation ambidexterity, as well as incorporates the level of risk-taking as a moderating variable. Using Stata 17 to perform panel data regression and a series of robustness tests, we analyze 1,908 firm-year observations from the information transmission, software, and information technology service industries, as well as the scientific research and technology service industries listed on the Shanghai and Shenzhen Stock Exchanges from 2012 to 2022. The empirical analyses propose that AI applications remarkably enhance innovation ambidexterity at the 1 % level. This positive effect is the most pronounced when risk-taking is moderate. This study extends information processing theory to the AI-enabled innovation context and further enriches its boundary conditions by introducing risk-taking as a nonlinear moderator. Managerially, the findings suggest that enterprises should calibrate risk-taking levels to complement AI deployment, enabling a balanced approach to exploration and exploitation.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104635"},"PeriodicalIF":6.9,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023315","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-20DOI: 10.1016/j.ipm.2026.104640
Abdelfatah Ahmed , Mohammad Irshaid , Mohamad Alansari , Divya Velayudhan , Mohammed Tarnini , Mohammed El-Amine Azz , Naser Al-Khalayleh , Taimur Hassan , Ernesto Damiani , Naoufel Werghi
X-ray baggage screening demands robust automated detection systems that can perform reliably under limited annotation. We introduce PRISM-X, a progressive semi-supervised framework that integrates pseudo-label generation, region-level multimodal grounding, and contrastive refinement into a unified detection pipeline. Evaluated on the SIXray and CLCXray benchmarks, PRISM-X consistently outperforms both vision-only and vision-language baselines. On SIXray, it achieves 66.0% mAP and 87.4% AP50, improving over the strongest vision-only method by +6.8% mAP and the leading vision-language model by +6.7% grounding accuracy. On CLCXray, PRISM-X reaches 45.7% mAP, 64.7% AP50, and 51.6% AP75, surpassing Cascade R-CNN by +3.3% mAP and GDINO by +3.4% AP75. These results demonstrate the effectiveness of PRISM-X in low-label, cluttered X-ray scenarios and its superiority over existing weakly and semi-supervised approaches.
{"title":"PRISM-X: Progressive semi-supervised threat detection in X-ray scans with self-guided multimodal refinement","authors":"Abdelfatah Ahmed , Mohammad Irshaid , Mohamad Alansari , Divya Velayudhan , Mohammed Tarnini , Mohammed El-Amine Azz , Naser Al-Khalayleh , Taimur Hassan , Ernesto Damiani , Naoufel Werghi","doi":"10.1016/j.ipm.2026.104640","DOIUrl":"10.1016/j.ipm.2026.104640","url":null,"abstract":"<div><div>X-ray baggage screening demands robust automated detection systems that can perform reliably under limited annotation. We introduce PRISM-X, a progressive semi-supervised framework that integrates pseudo-label generation, region-level multimodal grounding, and contrastive refinement into a unified detection pipeline. Evaluated on the SIXray and CLCXray benchmarks, PRISM-X consistently outperforms both vision-only and vision-language baselines. On SIXray, it achieves 66.0% mAP and 87.4% AP<sub>50</sub>, improving over the strongest vision-only method by +6.8% mAP and the leading vision-language model by +6.7% grounding accuracy. On CLCXray, PRISM-X reaches 45.7% mAP, 64.7% AP<sub>50</sub>, and 51.6% AP<sub>75</sub>, surpassing Cascade R-CNN by +3.3% mAP and GDINO by +3.4% AP<sub>75</sub>. These results demonstrate the effectiveness of PRISM-X in low-label, cluttered X-ray scenarios and its superiority over existing weakly and semi-supervised approaches.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104640"},"PeriodicalIF":6.9,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023318","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Financial sentiment analysis (FSA) has garnered considerable attention for its potential to detect bullish and bearish sentiments that drive stock market fluctuations. Nonetheless, extracting salient sentiments from analyst reports encounters two main challenges. First, the highly specialized terms and expressions prevalent in these reports make it difficult for general Large Language Models (LLMs) to interpret financial expertise. Second, sentiment cues are implicit and dispersed across long-range dependencies, whereas existing LLM-based FSA methods relying on a single fine-tuning strategy lack fine-grained control during adaptation, thus leading to key information loss. To tackle these issues, we propose SEHLP, the first LLM that integrates summary information with a hybrid adaptation strategy that combines Low-rank Adaptation (LoRA) and dynamic Prefix Tuning to enhance FSA. Specifically, we employ prompt engineering on Qwen-2.5-14B to generate concise summaries that distill salient insights of each report, and construct FinLLaMA as SEHLP’s backbone through Supervised Fine-tuning (SFT) on extensive domain-specific instructions, enhancing financial knowledge comprehension. To inject summary information and enable fine-grained control during fine-tuning, we propose a hybrid adaptation strategy that concatenates LoRA-updated attention projections with dynamic summary-enhanced key-value prefixes, thereby fully utilizing sentiment cues in analyst reports and their summaries. Moreover, we construct a large-scale LCFR-Instruct corpus with 16,912 samples to address the lack of high-quality FSA instruction data. Comprehensive experiments on the LCFR-Instruct and FinTHUC-Instruct benchmark datasets indicate that SEHLP, with only 1.3B parameters, consistently surpasses competing LLMs, exhibiting ACC gains of 1.89% and 1.59% over the larger FinGPT-7B model on both datasets while maintaining superior efficiency. Our code is publicly accessible at https://github.com/lhz9999/SEHLP.
{"title":"SEHLP: A summary-enhanced large language model for financial report sentiment analysis via hybrid LoRA and dynamic prefix tuning","authors":"Haozhou Li, Qinke Peng, Xu Mou, Zeyuan Zeng, Ruimeng Li, Jinzhi Wang, Wentong Sun","doi":"10.1016/j.ipm.2026.104639","DOIUrl":"10.1016/j.ipm.2026.104639","url":null,"abstract":"<div><div>Financial sentiment analysis (FSA) has garnered considerable attention for its potential to detect bullish and bearish sentiments that drive stock market fluctuations. Nonetheless, extracting salient sentiments from analyst reports encounters two main challenges. First, the highly specialized terms and expressions prevalent in these reports make it difficult for general Large Language Models (LLMs) to interpret financial expertise. Second, sentiment cues are implicit and dispersed across long-range dependencies, whereas existing LLM-based FSA methods relying on a single fine-tuning strategy lack fine-grained control during adaptation, thus leading to key information loss. To tackle these issues, we propose SEHLP, the first LLM that integrates summary information with a hybrid adaptation strategy that combines Low-rank Adaptation (LoRA) and dynamic Prefix Tuning to enhance FSA. Specifically, we employ prompt engineering on Qwen-2.5-14B to generate concise summaries that distill salient insights of each report, and construct FinLLaMA as SEHLP’s backbone through Supervised Fine-tuning (SFT) on extensive domain-specific instructions, enhancing financial knowledge comprehension. To inject summary information and enable fine-grained control during fine-tuning, we propose a hybrid adaptation strategy that concatenates LoRA-updated attention projections with dynamic summary-enhanced key-value prefixes, thereby fully utilizing sentiment cues in analyst reports and their summaries. Moreover, we construct a large-scale LCFR-Instruct corpus with 16,912 samples to address the lack of high-quality FSA instruction data. Comprehensive experiments on the LCFR-Instruct and FinTHUC-Instruct benchmark datasets indicate that SEHLP, with only 1.3B parameters, consistently surpasses competing LLMs, exhibiting ACC gains of 1.89% and 1.59% over the larger FinGPT-7B model on both datasets while maintaining superior efficiency. Our code is publicly accessible at <span><span>https://github.com/lhz9999/SEHLP</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104639"},"PeriodicalIF":6.9,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023256","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-20DOI: 10.1016/j.ipm.2026.104628
Hui Ye , Xuri Ge , Junqi Wang , Junchen Fu , Xin Xin , Jiao Xue , Yao Chen , Pengjie Ren , Zhumin Chen
To address the inefficiency of full fine-tuning of Contrastive Language-Image Pre-training (CLIP) models and the performance loss of adapter-based methods, we propose a novel efficient hybrid fine-tuning strategy (called HFLIP) to achieve a balance of efficiency and performance. HFLIP fine-tunes the key selected ViT blocks with interpretable semantic attention supervision on selected transformer heads via machine-learning methods’ selection, while keeping other blocks adapter-based for efficiency. Specifically, HFLIP introduces two key components: (1) a Dynamic Block-selection Genetic Algorithm (DBGA) that automatically selects a small subset of critical blocks in the ViT for full tuning, while keeping the rest adapter-tuned, ensuring a proper trade-off between fine-tuning effectiveness and efficiency; and (2) a Clustering-based Head-selection with Explainable-attention Guidance (CHEG), where hierarchical clustering is employed to identify representative attention heads, which are then fine-tuned under guidance from explainable attention maps, encouraging semantically consistent and globally diverse attention patterns. Extensive experiments on multiple downstream tasks show that HFLIP achieves comparable or even better performance than full fine-tuning by updating only 30% of the training parameters, while reducing GPU memory consumption by about 16%. In addition, HFLIP makes the CLIP-based ViT attention mechanism more interpretable compared to both the pretrained CLIP and other fine-tuned variants. We release our code at https://github.com/huiye8870/HFLIP.
{"title":"Beyond efficient fine-tuning: Efficient hybrid fine-tuning of CLIP models guided by explainable ViT attention","authors":"Hui Ye , Xuri Ge , Junqi Wang , Junchen Fu , Xin Xin , Jiao Xue , Yao Chen , Pengjie Ren , Zhumin Chen","doi":"10.1016/j.ipm.2026.104628","DOIUrl":"10.1016/j.ipm.2026.104628","url":null,"abstract":"<div><div>To address the inefficiency of full fine-tuning of Contrastive Language-Image Pre-training (CLIP) models and the performance loss of adapter-based methods, we propose a novel efficient hybrid fine-tuning strategy (called HFLIP) to achieve a balance of efficiency and performance. HFLIP fine-tunes the key selected ViT blocks with interpretable semantic attention supervision on selected transformer heads via machine-learning methods’ selection, while keeping other blocks adapter-based for efficiency. Specifically, HFLIP introduces two key components: (1) a Dynamic Block-selection Genetic Algorithm (DBGA) that automatically selects a small subset of critical blocks in the ViT for full tuning, while keeping the rest adapter-tuned, ensuring a proper trade-off between fine-tuning effectiveness and efficiency; and (2) a Clustering-based Head-selection with Explainable-attention Guidance (CHEG), where hierarchical clustering is employed to identify representative attention heads, which are then fine-tuned under guidance from explainable attention maps, encouraging semantically consistent and globally diverse attention patterns. Extensive experiments on multiple downstream tasks show that HFLIP achieves comparable or even better performance than full fine-tuning by updating only 30% of the training parameters, while reducing GPU memory consumption by about 16%. In addition, HFLIP makes the CLIP-based ViT attention mechanism more interpretable compared to both the pretrained CLIP and other fine-tuned variants. We release our code at <span><span>https://github.com/huiye8870/HFLIP</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 4","pages":"Article 104628"},"PeriodicalIF":6.9,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146023316","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号