Mirror-assisted strategies are commonly used in the rehabilitation training of patients with hemiparesis in the upper limbs following a stroke. Traditional robotic mirror assistance focuses on achieving high-precision mirror trajectory tracking, often neglecting the issue of active movement in the affected side. This paper proposes a task performance-based adaptive impedance control, where the robot assists the affected side in an assist-as-needed manner, thereby encouraging the patient to perform active movements. To account for inter-individual variability, a method for assessing the affected side’s motor performance, based on the healthy side’s movement level, is introduced. Adaptive impedance control is then constructed based on the motor performance of the affected side, enabling the robot to provide adaptive assistance force. Eight healthy participants were recruited for experimental testing. Experimental results show that when the robot provides mirror-based assist-as-needed to the affected side, the robot’s stiffness coefficient and assistance force are positively correlated with the motor assessment coefficient of the affected side, thereby verifying the feasibility of the proposed strategy. This study offers a robotic-assisted rehabilitation strategy for stroke patients that balances active participation and individual adaptability, with the potential to enhance rehabilitation outcomes and enable precise rehabilitation interventions.

The rapid expansion of online programming exercise platforms has brought abundant learning resources for programming education, but also presents challenges for personalized exercise recommendation due to missing or imprecise knowledge annotations and the sparsity of learner-exercise interaction records, which together lead to reduced accuracy and a lack of explainability in the recommendation results. Existing natural language processing and large language models (LLM)-based annotation methods struggle to capture implicit knowledge and often generate redundant or inconsistent results. Moreover, mainstream recommendation systems are ineffective at handling the severe sparsity of interaction sequences in programming exercise datasets, and typically model student preferences in a single dimension—overlooking key educational factors such as knowledge gaps, difficulty tolerance, and preferred learning rhythm. To address these challenges, we propose KDG-Rec, a novel framework that first introduces AgentCo-KAS, a multiagent LLM-based collaborative annotation method with role-specific fine-tuning and cross-agent verification for high-precision, fine-grained knowledge annotation. Building on these enriched annotations, we develop a disentangled graph neural network model that constructs dual exercise-interaction graphs to effectively capture learning patterns at multiple granularities in interaction sequences and explicitly decouples student preferences into four interpretable dimensions for adaptive fusion. Extensive experiments on real-world datasets demonstrate that KDG-Rec outperforms nine state-of-the-art methods across multiple metrics, significantly advancing personalized programming exercise recommendation.

Dynamic interactive object search in large-scale human environments presents substantial challenges for existing methods. Current scene representations like 3D Scene Graphs (3DSG) only provide coarse-grained spatial segmentation and cannot identify functional areas such as storage or leisure areas. Without functional area understanding, existing methods are constrained to exhaustive sequential exploration at large scales, resulting in inefficient search behaviors—particularly in open-layout environments with numerous interactive objects such as drawers and cabinets. Moreover, these methods lack adaptability to environmental dynamics such as object relocations. To address these limitations, this paper proposes CMAR-search, a novel framework built upon Commonsense and Memory Augmented Reasoning (CMAR). Our approach leverages commonsense about area functionalities and aggregates environmental memory to construct a Functional 3D Scene Graph (F3DSG), which organizes the environment into functional areas with their associated containers. Through this structured representation, CMAR enables hierarchical action planning at both macro-area and micro-container levels, empowering the system to efficiently identify and inspect semantically relevant areas for effective object search. Notably, CMAR continuously integrates real-time perception, accumulated memory, and commonsense to dynamically relocalize objects in changing environments. Extensive experiments in simulation and real-world settings demonstrate that CMAR-search significantly surpasses state-of-the-art baselines in both success rate and search efficiency for object search in dynamic interactive environments.

In recent times, programming knowledge tracing, as a fundamental task in programming learning analytics, has been widely applied in education, enabling accurate diagnosis of learners' mastery of programming knowledge concepts across diverse learning contexts. Existing programming knowledge tracing methods often overlook two key challenges: knowledge concept annotation deficiencies caused by the multiple-solution nature of programming problems and inadequate modeling of the actual progression of knowledge states within students' continuous submission sequences for targeted problems. Consequently, the information contained in the submitted code is underutilized, negatively impacting the estimation of students' evolving knowledge states. In this paper, we propose HGAKT-Code, a code-aware knowledge tracing framework for programming education assisted by Online Judge (OJ) systems. Our model integrates three novel components: (1) a Hierarchical Multi-label Code Classification (HMCC) algorithm to mitigate Q-matrix inaccuracies caused by diverse code solutions; (2) a code similarity-enhanced attention mechanism to better capture critical submissions reflecting subtle changes in students' knowledge states; and (3) a heterogeneous GNN with designed graph attention, modeling both conceptual and temporal learning dynamics. We evaluated HGAKT-Code on three OJ datasets and achieve AUC improvements of 4.2%, 3.0%(public OJ dataset) and 2.9% (institutional OJ dataset) over state-of-the-art baselines, respectively, demonstrating its effectiveness across educational settings.

This paper proposes a force-field training strategy for stimulating upper-limb rehabilitation movements in hemiparetic patients. The approach integrates virtual channels, and assistance and correction force fields, and allows a robot to assist patients with movements based on their individual motor performance and rehabilitation needs. A dynamic model of the robot and a human-robot physical interaction model are established using impedance control, ensuring compliant interaction between the human and the robot. An adaptive assistance force field is created within the virtual channel, which adjusts according to the patient’s movement performance, while a correction force field is applied outside the virtual channel to guide the patient’s movements toward the desired trajectory. Experimental results validate that the proposed dual-field strategy dynamically adjusts to the patient’s moment-to-moment performance, successfully preventing the slacking effect and promoting active engagement. This approach offers promising potential for the clinical application of rehabilitation robots.

Knowledge Tracing (KT) aims to predict students' future performances based on their former exercises and additional information in educational settings. KT has received significant attention since it facilitates personalized experiences in educational situations. Simultaneously, the autoregressive modeling on the sequence of former exercises has been proven effective for this task. One of the primary challenges in autoregressive modeling for Knowledge Tracing is effectively representing the anterior (pre-response) and posterior (post-response) states of learners across exercises. Existing methods often employ complex model architectures to update learner states using question and response records. In this study, we propose a novel perspective on knowledge tracing task by treating it as a generative process, consistent with the principles of autoregressive models. We demonstrate that knowledge states can be directly represented through autoregressive encodings on a question-response alternate sequence, where model generate the most probable representation in hidden state space by analyzing history interactions. This approach underpins our framework, termed Alternate Autoregressive Knowledge Tracing (AAKT). Additionally, we incorporate supplementary educational information, such as question-related skills, into our framework through an auxiliary task, and include extra exercise details, like response time, as additional inputs. Our proposed framework is implemented using advanced autoregressive technologies from Natural Language Generation (NLG) for both training and prediction. Empirical evaluations on four real-world KT datasets indicate that AAKT consistently outperforms all baseline models in terms of AUC, ACC, and RMSE. Furthermore, extensive ablation studies and visualized analysis validate the effectiveness of key components in AAKT.

When rehabilitation robots assist patients with unilateral upper-limb motor dysfunction in mirror-assisted exercise training, the lack of bilateral tactile feedback may lead to secondary injury on the affected side. To address this issue, a force feedback mirror-aided strategy based on adaptive impedance control is proposed, utilizing a bimanual robot. This strategy establishes a force feedback mirror between the unaffected and affected sides, allowing the unaffected side to perceive the assistive force when the manipulator assists the affected side's movement, thereby ensuring the safety of the mirror-assisted exercise. To achieve mirror trajectory tracking during rehabilitation, a human-robot physical interaction model is developed based on the robotic dynamic model and impedance control. For tactile feedback between the unaffected and affected sides, adaptive impedance control is employed based on the interaction force between the affected side and the manipulator, ensuring that the interaction forces on both sides are proportional and equal in real time. Experimental results demonstrate that, during bilateral mirror-assisted exercise, force feedback mirrors are effectively formed between the two arms, confirming that the proposed strategy not only enables mirror trajectory tracking but also facilitates force feedback mirroring. This study lays the foundation for future safety optimization of robot-assisted mirror rehabilitation training systems.

Subjective questions are crucial to assess students’ ability to analyze, synthesize, evaluate and create knowledge. In the massive online education scenarios, the manually scoring of subjective questions is time-consuming. Instead, it could be supported by the task of Short Answer Grading in Natural Language Process. However, it is worth noting that most existing automatic scoring system does not perform well on domain-specific and long questions. In this paper we address the challenges of automated short answer grading (ASAG) by proposing a novel scoring approach that strategically integrates a fine-tuned large language model (LLM), a neural network (NN) for feature extraction, and an answer-question relevance assessment module (RELEVANCE). Our method effectively scores student responses based on a set of predefined rubrics and reference answers. Our experiments on the ASAP-SAS dataset demonstrate that our method achieves an average Quadratic Weighted Kappa (QWK) score of 0.797, surpassing current state-of-the-art AutoSAS model, particularly excelling in longer tasks with a 11.9

Sequential recommendation represents a well-explored yet challenging domain within research. Despite significant advancements in GNN-based methods for modeling intricate patterns in user-item interaction sequences, these methods face difficulties in capturing nuanced semantics in sequences with sparse dependencies and noise, and often struggle with short sequences. Additionally, distinguishing higher-order semantic distinctions among diverse user interests is still challenging, and existing GNN-based methods can be computationally intensive. To address these challenges, we propose a Dual-Graph approach for Sequential Recommendation, DGSR. We construct individual interaction graphs for each sequence, and a heterogeneous global interaction graph that incorporates user identity as an attribute of user edges. Last-item augmented GGNN is employed within individual interaction graphs to mitigate the impact of sparse dependencies and noise, thereby extracting the most recent interests for each user more effectively. Moreover, within the global graph, we propose a parameter-efficient heterogeneous GNN to extract high-order interest distinctions among diverse users while maintaining low computational complexity. Finally, we utilize vanilla transform mechanism to integrate intra- and inter-user interests from both types of graphs. Experiments on four publicly available datasets demonstrate that our method achieves state-of-the-art performance, surpassing all baseline methods.

Application Programming Interfaces (APIs) have become common in contemporary software development. Many automated API recommendation methods have been proposed. However, these methods suffer from a deficit of using domain knowledge, giving rise to challenges like the “cold start” and “semantic gap” problems. Consequently, they are unsuitable for test function recommendation, which recommends test functions for test engineers to implement test cases formed with various test steps. This paper introduces an approach named TOKTER, which recommends test functions leveraging test-oriented knowledge graphs. Such a graph contains domain concepts and their relationships related to the system under test and the test harness, which is constructed from the corpus data of the concerned test project. TOKTER harnesses the semantic associations between test steps (or queries) and test functions by considering literal descriptions, test function parameters, and historical data. We evaluated TOKTER with an industrial dataset and compared it with three state-of-the-art approaches. Results show that TOKTER significantly outperformed the baseline by margins of at least 36.6% in mean average precision (MAP), 19.6% in mean reciprocal rank (MRR), and 1.9% in mean recall (MR) for the top-10 recommendations.

In practices, test cases are often designed by test engineers based on the functionalities of the System under Test (SUT) in parallel and independently. This can lead to a lack of a comprehensive overview of the test architecture, hampering the reuse of test functions when implementing new test cases. To address this challenge, we propose ATAG, an automated test architecture generation approach, which employs an optimization algorithm to retrieve highly cohesive and loosely coupled test functions based on control flows and data flows of test cases. We also equip ATAG with a newly proposed BERT-based model, i.e., FunBERT, for generating test function names. We conducted an empirical study with three industrial datasets to evaluate the effectiveness of ATAG and FunBERT. Results show that test architectures generated with ATAG (benefiting from both control and data flows) improved, on average, ≈ 26–35 ≈ 28–50

In the realm of contemporary educational data mining, aspect-based sentiment analysis plays a crucial role in deciphering students’ nuanced perceptions of MOOC courses. However, sentiment analysis in educational context often encounters the prevalent challenge of cold start issues. This paper proposes a novel methodology for aspect-level sentiment analysis of course reviews, beginning with the identification of critical aspects in course reviews, followed by a comprehensive sentiment analysis at the aspect level. We introduce a Dual-Track Sentiment Analysis model (DTSA), which dynamically integrates two analytical tracks: one utilizing fine-tuned BERT model and the other employing sentiment dictionaries to effectively mitigate the cold start problem. Experimental results demonstrate the superiority of our approach over baseline models in various key metrics, particularly in addressing cold start challenges with limited review data. By incorporating a matching strategy, our model ensures reliable and timely sentiment analysis of course reviews, even with small amount of course reviews. This methodology effectively alleviates the cold start problem in aspect-level sentiment analysis in educational evaluation text, providing accurate insights when lacking sufficient initial learners’ review data and enhancing the robustness of MOOC course evaluation processes.

To achieve the primary teaching goal of a programming course, the typical teaching method is to improve students' programming skills by assigning them a number of programming tasks and to check the quality of programs finished by students by using online judge system. Currently, these judge systems mainly contribute to evaluating how well a program satisfies its task requirement. Besides this basic role, a program judge system should also help students to progressively improve their programs' quality throughout the entire development lifecycle, rather than simply assessing the correctness of their programs. This assistance encompasses program construction, testing, and bug fixing stages. To meet the diverse needs of program evaluation and handle the high workload generated by a large number of evaluation tasks, this paper proposes a distributed microservice based program judge system called FLOD (Full Lifecycle Online juDge) built on time driven interactive judge cores, together with the configurable testing services. FLOD provides teachers with configurable judge workflow and running environment, upon which various programming tasks can be assigned and implemented. FLOD can automatically access the code repositories of students by integrating with Gitlab and provide interactive code judge services to feedback students instantly with the bugs and bad code styles found such that students can improve their programs at their own paces. FLOD has been successfully deployed to serve the Object Oriented Design and Construction (OODC) course offered in Beihang university since 2019. To assess its effectiveness, we propose the metrics and collect the usage data from 2019 to 2022. The results show that FLOD effectively supports the accompanying programming teaching with considerable reliability and scalability, which is very important to achieve the teaching goal for our course with over 300 students.

针对学生的学习基础、学习能力、学习效率等差异,从进行课程顶层设计、重构教学内容体系、采取线上线下混合教学、构建多元过程化考核体系4个方面,提出开展大学计算机课程面向赋能教育的个性化教学思路和方法,为学生全面赋能.

Implementing test cases as programs to automate test execution is a popular testing practice. Current industrial practices usually use test functions to implement the test steps of a test case and then to compose the executable test case by choosing the test functions to call manually. It is time-consuming and could lead to invalid test results by selecting inappropriate test functions. In this article, we propose an automatic test function recommendation approach named Scenario-based Recommendation of TEst Function (SRTEF). Given a test step of a test case, SRTEF uses the weighted description similarity and the scenario similarity to recommend test functions. The description similarity utilizes the deep structured semantic model (DSSM) to measure the relatedness between a test step and a test function by their literal descriptions. The test scenario and the test function usage scenario are considered to calculate the scenario similarity. SRTEF has been successfully applied in Huawei. The systematic experiments have been conducted to evaluate SRTEF by using the dataset from Huawei and comparing with BiInformation source-based KnowledgE Recommendation (BIKER), reported as the best approach so far. The results show that SRTEF outperforms BIKER with significant positive ratios consistently in all the three selection strategies, i.e., Top-3, Top-5, and Top-10. The DSSM shows its advantage over word embedding by the double performance of capturing the semantic relatedness in SRTEF.

In recent years, Convolutional Neural Network (CNN) has achieved a great success in computer vision. However, at present, for an image classification task, there is no CNN model that can perform 100% accurately due to insufficient or excessive feature learning. Once a CNN model deployed to perform tasks online, misclassified samples might lead the system with the CNN model deployed to enter an unsafe state such as collisions. To assess the performance of such online models, we, in this paper, propose Parallel Signal Routing Paths (PSRP) method to identify misclassified samples by extracting execution paths for each sample and comparing inherent feature differences in terms of CNN nodes between misclassified and well-classified samples, for the ultimate aim of addressing the challenge of test data not having ground-truth labels in online environment where the CNN models are deployed, and give availability results for applying PSRP on 3 public datasets and 3 typical CNN models.

Machine learning (ML) system is highly data-dependent. It turns out different behavior and performance by using different training data. Since its behavior and performance are more unpredictable than traditional software. The requirement analysis of ML system should focus on model development and its training. This paper reveals how the model development and training are conducted by reviewing the published papers. Based on the principles by Kitchenham, we propose a concern driven systematic literature review method and choose face recognition facilitated ML system as case study. We identify the concerns and then research questions from data perspective, and answer the questions based on the collected literature data. The results show that the current studies included in review have already recognized the role of data played in ML system. Since there is no standard template and method to report data and the related processing procedure, the content and detail about the data reported vary from study to study and are not systematic, which is a great challenge to both practitioners and users.

Implementing test cases to automate test execution is a popular testing practice currently. A stepwise test case consists of several sequential test steps. Given a test function library, the typical way to implement a test case is calling the existing test functions in the library to reduce test cost. How to find the appropriate test function(s) to implement a test step in a given test case thus becomes an important problem. However, in current testing practices, test engineers usually select the appropriate test function manually by experience. It is time-consuming and could lead to invalid test results by selecting inappropriate or wrong test functions to call. In this paper, we propose an automatic test function recommendation approach with scenario named SRTEF (Scenario-based Recommendation of TEst Function). Given a test step, SRTEF uses two levels of similarities to recommend test functions, description similarity and scenario similarity. The description similarity measures the semantic relatedness between the test step and test function by their literal descriptions. To calculate the scenario similarity, SRTEF at first retrieves a set of historical test cases that contains test step(s) semantically similar to the given test step; then the scenario similarity between test step and test function is calculated according to the calling relation between retrieved test case and test function, and the co-occurrence relation among test functions. SRTEF has been successfully applied in Huawei. We evaluate SRTEF by using the dataset from Huawei and comparing with BIKER, reported as the best recommendation approach so far. The results show that SRTEF outperforms the BIKER approach by at least 49% in Mean Average Precision, 33% in Mean Reciprocal Rank, and 25% in Mean Recall.

As one mainstream of current software development, Object-oriented programming has become one key course for undergraduate students in Computer Science. Since Object-oriented concepts are difficult to understand for students, small programming exercises are used to train and help the students, and the study performance is evaluated based on the quality of the submitted source code. The common practice of code assess-ment in programming courses is checking whether the submitted projects pass carefully-designed test cases. However, even some projects pass all test cases, they may have bad software design and do not use the knowledge of Object-oriented programming well, especially in the early stage of courses. Therefore, we propose an anomaly detection approach for early warning in Object-oriented programming courses, which can automatically find the abnormal application of Object-oriented knowledge. In our approach, we conduct static analysis on the code submitted by students. Typical Objected-oriented metrics are extracted, and students are divided into two groups by K-means clustering: being good at Object-objected knowledge or not, and finally detect anomalous students based on the distance from cluster centers. We evaluate our approach on the realistic data sets collected from our Object-oriented programming course, and experimental results show the effectiveness of our method.

This study conducts a data-driven statistic analysis for comparatively examining early programmers' academic performance and learning behavior between a SPOC blended teaching class and a traditional teaching class in an introductory C programming course. Teaching activities in the experimental class are implemented by a SPOC blended supporting tool -"rainy class", while the control group implements traditional classroom teaching. Both the two groups are required to complete assigned labs in the automatic assessment system-the online judge, and log data in terms of students' learning procedure is collected. Group analysis is employed to verify whether the two group differ in academic performance and learning behavior. The results show significant difference between the two groups on certain indicators. The findings confirm that the middle-lower programming foundation students in SPOC blended teaching class are positively regarding better academic performance, learning efforts and are able to obtain good achievements during the tests. Then, correlation analysis is conducted in the SPOC blended learning group to explore the impact of the performance in "rainy classroom" on students' academic performance and learning behavior. Conclusions show that effective "rainy classroom" engagement is positively correlated to students' final examination score and meanwhile promotes early programmers' performance in the overall online learning process.