As a key enabler of location-based services, next location prediction models user mobility to improve service personalization, benefiting applications such as Point-of-Interest (POI) recommendation and route planning. Although Large Language Model (LLM)-based prediction methods have attracted increasing attention in recent years, their performance remains suboptimal due to the difficulty of compactly and effectively modeling both semantic information and mobility patterns in trajectories, as well as the stepwise-biased training and inference that hinder the capture of long-term behavioral regularities. To address these challenges, a novel framework named MoveAlign is proposed to learn compact location representations and multi-step behavioral correlations in user mobility, thereby enhancing LLMs for next location prediction. First, a new mobility-aware semantic tokenization method is proposed to represent locations as compact mobility-token tuples, preserving semantic structure and mobility patterns as concise and structured inputs. Second, a new multi-step imitation learning strategy is proposed for supervised LLM fine-tuning, enabling the LLM to learn stable mobility patterns with higher-order behavioral correlations across multiple steps, thereby improving the accuracy of next location prediction. Extensive experiments on three real-world urban datasets demonstrate that MoveAlign consistently outperforms the existing baselines across standard evaluation metrics.
In addition to providing a selling channel for manufacturer’s national brand (NB), the e-tailer also sells substitute store brand (SB), which can be sourced from an outside contract manufacturer or NB manufacturer. The NB manufacturer cooperates with the e-tailer through the agency or resale format. Our study analyzes the e-tailer’s source decision on the SB while considering the NB manufacturer’s decisions on whether to produce the SB and selling format. Meanwhile, we analyze how these decisions affect each other. We find that, first, the NB manufacturer prefers agency format if the commission rate is low. Meanwhile, sourcing the SB from the outside contract manufacturer encourages the NB manufacturer to choose the agency format. Second, the NB manufacturer always prefers to produce the SB. Third, the e-tailer’s source decision depends on the competition and the production cost of the outside contract manufacturer. When the competition is weak, the e-tailer prefers to source the SB from the outside contract manufacturer if the production cost is low. When the competition is intense, the e-tailer under resale format prefers to source the SB from the outside contract manufacturer regardless of the cost; the e-tailer under agency format prefers to source the SB from the NB manufacturer regardless of the cost.
Digital governance represents a pivotal shift in governmental administration and a fundamental pathway toward achieving national governance modernization in the digital age. Leveraging the National Pilot Policy of Information Benefiting the People (NPPIP) as a quasi-natural experiment, this study employs panel data from 237 prefecture-level and above cities in China between 2010 and 2022 to empirically investigate the impact of government digital governance on fiscal expenditure efficiency. The results show that the NPPIP significantly enhances fiscal expenditure efficiency, and this conclusion remains valid after undergoing multiple robustness tests. Mechanism analyses further uncover that the enhancement stems primarily from improved resource allocation efficiency and greater fiscal transparency. Heterogeneity analysis indicates that the effects are driven mainly by gains in technical efficiency change and are more substantial in cities with higher administrative levels. This study offers innovative causal evidence on how government governance reshapes fiscal expenditure efficiency from a digital transformation perspective. It provides essential theoretical underpinnings and empirical evidence to further advance governance model innovation and fiscal sustainability.
The widespread sharing of point-of-interest (POI) visitation experiences through location-based social networks (LBSNs) has led to massive mobility check-in data, making next POI recommendation a critical task for personalized services. Although existing deep learning-based methods have achieved impressive performance in this task, there exist several issues, including difficulty in recommending new POIs, limited exploitation of heterogeneous semantic relations, and insufficient consideration of user mobility intent. These issues ultimately constrain the quality of personalized recommendations. To overcome these issues, this study formulates the next POI recommendation framework from an intent-aware exploration perspective, and proposes a novel intent-aware recommendation model based on multi-relational heterogeneous graph, named SeekNew. Specifically, a multi-relational heterogeneous graph module and an intent-aware tree module are proposed and integrated to improve the accuracy of the next POI recommendation, especially the recommendation of new POIs, and a sample-wise loss weighting strategy is proposed to amplify the influence of rare samples for balanced model optimization. The proposed multi-relational heterogeneous graph module models multiple heterogeneous relations among entities in LBSNs for enhancing the quality of trajectory representations. The proposed intent-aware tree module hierarchically infers users’ next mobility intent for effectively narrowing the solution space of candidate POIs. Experimental results on check-in data from three cities demonstrate that SeekNew outperforms state-of-the-art baselines, with notable improvements in recommending new POIs.
This paper constructs an index of multidimensional attention from distal stakeholders, including government agencies, media, competitors, industry leaders, social critics, higher education institutions, and examines its influence on enterprise innovation. The results indicate that distal stakeholder attention has a significant positive effect on both innovation input and output. Mechanism analyses reveal that it promotes enterprise R&D investment by alleviating financing constraints and reducing executive short-sightedness, whereas it fosters invention patent applications primarily by mitigating executive short-sightedness.
