Legal information retrieval demands high precision, yet traditional ''Retrieve-then-Rerank'' pipelines with two separate models suffer from cascading error propagation and knowledge disconnects between stages. To address these issues, we propose RULER, a Robust Unified LLM-based Efficient Retrieval that integrates efficient Bi-Encoder retrieval and high-precision Cross-Encoder reranking within a parameter-sharing architecture. To mitigate the Phantom Hits problem that irrelevant documents are assigned unreasonably high confidence, we introduce a Distribution-Robust Data Construction strategy that explicitly simulates pure-negative candidate groups. This is coupled with a Dynamic Margin Ranking Objective and Maximum Entropy Regularization, which collectively enforce uncertainty on irrelevant samples and enhance robustness. Extensive experiments on the JuDGE and LeCaRDv2 benchmarks demonstrate that RULER achieves state-of-the-art performance, outperforming all independent retrievers in retrieval tasks and surpassing competing unified architectures—where retriever and reranker parameters are shared—in high-precision reranking.

Image classification, a fundamental task in computer vision, often suffers from limited labeled samples, a problem commonly referred to as few-shot learning. Model-Agnostic Meta-Learning (MAML) addresses this issue by learning a set of meta-parameters from a variety of similar tasks, enabling rapid adaptation to new tasks within the same distribution using only a few gradient updates. However, traditional MAML frameworks usually encounter two major limitations: one is that they seek a common initialization shared across the entire task distribution, which restricts their capacity to capture the diversity inherent in varying task distributions. The other is that they resort to conventional optimizers (e.g., SGD), which limits their ability to adaptively learn across different domain tasks. To address these limitations, we propose a Task-adaptive Model-Agnostic MetaLearning (TMAML) framework. TMAML improves the existing MAML framework by jointly learning both the initialization and the parameter update rules, tailored for each individual task. The model first learns the domain information of the task as prior parameters, enabling more efficient and effective adaptation in both the initialization and update stages. During initialization, TMAML comprehensively considers the prior parameters of the specific task as well as the intrinsic information of the current samples. When learning the parameter update rules, it fully incorporates the prior parameters of the specific task along with the model's current learning state to guide adaptation more precisely. Experimental results demonstrate that TMAML effectively improves upon existing methods, boosting image classification performance across a wide range of settings. Notably, compared to the state-of-the-art cross-domain few-shot learning method, TMAML achieves an accuracy improvement of up to 3.56 % on 3-domain 5-way 4-shot tasks and up to 5.41 % on 5-domain 5-way 5-shot tasks.

Knowledge Tracing (KT) has emerged as a critical enabling technology for intelligent educational services, aiming to predict learners' knowledge states by analyzing their interaction histories. As personalized education evolves into a service-oriented paradigm, KT plays a pivotal role in optimizing service delivery, tailoring content recommendations, and enhancing learner engagement through adaptive learning services. However, existing KT methods face significant computational and modeling challenges, such as the insufficient integration of multi-dimensional question attributes, the inability to dynamically model knowledge decay in service-driven environments, and the neglect of conjectural bias, which leads to overestimated learner performance and suboptimal service outcomes. To address these challenges, we propose HDKT (Hypergraph-based De-conjecture Knowledge Tracing), a novel KT framework designed for service-oriented educational platforms. First, the HDKT approach incorporates a Hypergraph-Enhanced Question Representation Module, which explores multiple attributes of questions to obtain comprehensive representations, improving data fidelity for personalized learning services. Second, HDKT utilizes a novel Knowledge State Evolution Module to trace students' knowledge states precisely. This module is a sequential neural network that is augmented with a session-based forgetting mechanism to dynamically model knowledge decay and a de-conjecture mechanism to adjust learners' knowledge states by mitigating conjectural influences, ensuring more accurate assessments. Extensive experiments on three benchmark educational datasets demonstrate that HDKT significantly outperforms eleven state-of-the-art methods in terms of AUC, RMSE, and ACC. The results highlight HDKT's potential as a robust and scalable solution for AI-powered, service-oriented educational platforms, paving the way for more adaptive and personalized learning services. The full code and data can be found at https://anonymous.4open.science/r/HDKT-Code.

GPT has demonstrated impressive capabilities in executing various natural language processing (NLP) and reasoning tasks, showcasing its potential for deductive coding in social annotations. This research explored the effectiveness of prompt engineering and fine-tuning approaches of GPT for deductive coding of context-dependent and context-independent dimensions. Coding context-dependent dimensions (i.e., Theorizing, Integration, Reflection) requires a contextualized understanding that connects the target comment with reading materials and previous comments, whereas coding context-independent dimensions (i.e., Appraisal, Questioning, Social, Curiosity, Surprise) relies more on the comment itself. Utilizing strategies such as prompt decomposition, multi-prompt learning, and a codebook-centered approach, we found that prompt engineering can achieve fair to substantial agreement with expert-labeled data across various coding dimensions. These results affirm GPT's potential for effective application in real-world coding tasks. Compared to context-independent coding, context-dependent dimensions had lower agreement with expert-labeled data. To enhance accuracy, GPT models were fine-tuned using 102 pieces of expert-labeled data, with an additional 102 cases used for validation. The fine-tuned models demonstrated substantial agreement with ground truth in context-independent dimensions and elevated the inter-rater reliability of context-dependent categories to moderate levels. This approach represents a promising path for significantly reducing human labor and time, especially with large unstructured datasets, without sacrificing the accuracy and reliability of deductive coding tasks in social annotation. The study marks a step toward optimizing and streamlining coding processes in social annotation. Our findings suggest the promise of using GPT to analyze qualitative data and provide detailed, immediate feedback for students to elicit deepening inquiries.

Abstract In call centers, the service level is an important metric to measure the reasonability of the staffing schedule. Traditional service level calculation methods are based on the queue theory, which has very strict restrictions and is not suitable for real scenarios. Therefore, in this paper, a data‐driven method to solve the service level prediction problem is proposed to be used. To this end, the relationship between service level and other factors, such as number of calls, number of agents, time, is explored. Then some features are extracted based on empirical analyses and propose to use decision tree based ensemble methods, like random forest and GBDT, to model the relationship between service level and input features. Finally, extensive experimental results show that the proposed method outperforms other baselines significantly. Especially compared with the traditional queue theory methods, our method improves the performance by 6% and 9% in terms of MAE and MAPE.

Service level is an important metric to measure the reasonability of service systems. However, traditional mathematical service level prediction models have strict restrictions and suffer from accuracy degradation in complex real scenarios. In this paper, we propose to use a Siamese neural network to solve the service level prediction problem. This task is challenging due to two reasons: insufficient information and prior knowledge constraint. To tackle these issues, we develop a method entitled SiaSL based on the Siamese neural network. Our model consists of three key modules: 1) a time embedding module to embed the time period into low-dimensional vectors to model the time-independent characteristics of the service level; 2) a feature extractor to extract deep feature from raw scheduling information; 3) an output module to predict the final service level. Besides, to make SiaSL learn the prior knowledge, we propose a data augmentation method and an iterative training mechanism. Extensive experiments on a real-world dataset validate the effectiveness and efficiency of our method. On the premise of satisfying prior knowledge, our model achieves state-of-the-art performance on this problem.

Delivery stations play important roles in logistics systems. Well-designed delivery station planning can improve delivery efficiency significantly. However, existing delivery station locations are decided by experts, which requires much preliminary research and data collection work. It is not only time consuming but also expensive for logistics companies. Therefore, in this article, we propose a data-driven pipeline that can transfer expert knowledge among cities and automatically allocate delivery stations. Based on existing well-designed station location planning in the source city, we first train a model to learn the expert knowledge about delivery range selection for each station. Then we transfer the learned knowledge to a new city and design three strategies to select delivery stations for the new city. Due to the differences in characteristics among different cities, we adopt a transfer learning method to eliminate the domain difference so that the model can be adapted to a new city well. Finally, we conduct extensive experiments based on real-world datasets and find the proposed method can solve the problem well.

To improve the classification performance of support vector machines (SVMs) on imbalanced datasets, cost-sensitive learning methods have been proposed, e.g., Different Error Costs (DEC) and Fuzzy SVM for Class Imbalance Learning (FSVM-CIL). They relocate the hyperplane by adjusting the costs associated with misclassifying samples. However, the error costs are determined either empirically or by performing an exhaustive search in the parameter space. Both strategies can not guarantee effectiveness and efficiency simultaneously. In this paper, we proposeATEC, a solution that can efficiently find a preferable hyperplane by automatically tuning the error cost for between-class samples.ATECdistinguishes itself from all existing parameter tuning strategies by two main features: (1) it can evaluate how effective an error cost is in terms of classification accuracy; and (2) it changes the error cost in the right direction if it is not effective. Extensive experiments show that compared with the state-of-art methods, SVMs that are equipped withATECcan not only obtain comparable improvements in terms of F1 score of minority class, area under the precision-recall curve (AUC-PR) and area under the ROC curve (AUC-ROC) scores, but also outperform the grid-search parameter tuning strategy by two orders of magnitude in terms of the training time when a high F1 score is required.

Nowadays, although there exists many ridesharing services and dynamic matching algorithms for passengers and drivers, there is no service or algorithm that can balance the benefit of passengers and drivers while taking their time and cost constraints into consideration. In this paper, we try to solve the dynamic ridesharing problem by considering all above factors for all the participants. To this end, we present URoad, an efficient algorithm for large-scale dynamic ridesharing service, where a new price cost model is carefully designed to make up for the shortcomings of existing algorithms, and in the meantime a corresponding efficient matching algorithm is proposed to satisfy both the time and cost constraints of passengers and drivers. Specifically, for a given passenger, URoad will find out the optimal driver who can satisfy all the constraints of the passenger and the driver with the minimum detour distance. We design a series of data structures to speed up URoad for large scale ridesharing service application, e.g., Time Index, Grid Index and Greedy Strategy. Through extensive experiments, we prove that URoad can find the optimal driver for a given passenger from more than one hundred thousand drivers within 0.5 second in average.

Nowadays, many location-based applications require the ability of querying k-nearest neighbors over a very large scale of moving objects in road networks, e.g., taxi-calling and ride-sharing services. Traditional grid index with equal-sized cells can not adapt to the skewed distribution of moving objects in real scenarios. Thus, to obtain the fast querying response time, the grid needs to be split into more smaller cells which introduces the side-effect of higher memory cost, i.e., maintaining such a large volume of cells requires a much larger memory space at the server side. In this paper, we present SIMkNN, a scalable and in-memory kNN query processing technique. SIMkNN is dual-index driven, where we adopt a R-tree to store the topology of the road network and a hierarchical grid model to manage the moving objects in non-uniform distribution. To answer a kNN query in real time, SIMkNN adopts the strategy that incrementally enlarges the search area for network distance based nearest neighbor evaluation. It is far from trivial to perform the space expansion within the hierarchical grid index. For a given cell, we first define its neighbors in different directions, then propose a cell communication technique which allows each cell in the hierarchical grid index to be aware of its neighbors at anytime. Accordingly, an efficient space expansion algorithm to generate the estimation area is proposed. The experimental evaluation shows that SIMkNN outperforms the baseline algorithm in terms of time and memory efficiency.

This paper studies dynamic pricing for cloud service where different resources are consumed by different users. The traditional cloud resource pricing models can be divided into two categories: on‐demand service and reserved service. The former only takes the using time into account and is unfair for the users with long using time and little concurrency. The latter charges the same price to all the users and does not consider the resource consumption of users. Therefore, in this paper, we propose a flexible dynamic pricing model for cloud resources, which not only takes into account the occupying time and resource consumption of different users but also considers the maximal concurrency of resource consumption. As a result, on the one hand, this dynamic pricing model can help users save the cost of cloud resources. On the other hand, the profits of service providers are guaranteed. The key of the pricing model is how to efficiently calculate the maximal concurrency of resource consumption since the cost of providers is dynamically varied based on the maximal concurrency. To support this function in real time, we propose a data structure based on the classical B+ tree and the implementation for its corresponding basic operations like insertion, deletion, split, and query. Finally, the experiment results show that we can complete the dynamic pricing query on 10 million cloud resource usage records within 0.2 seconds on average.

In this paper, we propose a new problem: covering the optimal time window over temporal data. Given a duration constraint d and a set of users where each user has multiple time intervals, the goal is to find all time windows which (1) are greater than or equal to the duration d, and (2) can be covered by the intervals from as many as possible users. This problem can be applied to real scenarios where people need to determine the best time for maximizing the number of people to be involved in an activity, e.g., the meeting organization and the online live video broadcasting. As far as we know, there is no existing algorithm that can solve the problem directly. In this paper, we propose two algorithms to solve the problem, the first one is considered as a baseline algorithm called sliding time window (STW), where we utilize the start and end points of all users - intervals to construct time windows satisfying duration d. And then we calculate the number of users whose intervals can cover the current time window. The second method, named TLI, is designed based on the the data structures from the Timeline Index in SAP HANA. In TLI algorithm, we conduct three consecutive phases to achieve the purpose of efficiency improvement, namely construction of Timeline Index, calculation of valid user set and calculation of time windows. Within the third phase, we prune the number of time windows by keeping track of the number of users in current optimal time window, which can help shrink the search space. Through extensive experimental evaluations, we find TLI algorithm outperforms STW two orders of magnitude in terms of querying time.