Information Systems最新文献

Process model comprehension is essential for a variety of technical and managerial tasks. To facilitate comprehension, process models are often divided into subprocesses when they reach a certain size. However, depending on the task type this can either support or impede comprehension. To investigate this hypothesis, we conduct a comprehensive eye-tracking study, where we test two different types of comprehension tasks. These are local tasks focusing on a single subprocess, thereby benefiting from abstraction (i.e., irrelevant information is hidden), and global tasks comprising multiple subprocesses, thereby also benefiting from abstraction but impeded by fragmentation (i.e., relevant information is distributed across multiple fragments). Our subsequent analysis at task (coarse-grained) and phase (fine-grained) levels confirms the opposing effects of abstraction and fragmentation. For global tasks, we observe lower task comprehension, higher cognitive load, as well as more complex search and inference behaviors, when compared to local ones. An additional qualitative analysis of search and inference phases, based on process maps and time series, provides additional insights into the evolution of information processing and confirms the differences between the two task types. The fine-grained analysis at the phase level is based on a novel research method, allowing to clearly separate information search from information inference. We provide an extensive validation of this research method. The outcome of this work provides a more thorough understanding of the effects of fragmentation, in the context of modularized process models, at a coarse-grained level as well as at a fine-grained level, allowing for the development of task- and user-centric support, and opening up future research opportunities to further investigate information processing during process comprehension.

Interaction prediction is a crucial task in the Social Internet of Things (SIoT), serving diverse applications including social network analysis and recommendation systems. However, the dynamic nature of items, users, and their interactions over time poses challenges in effectively capturing and analyzing these changes. Existing interaction prediction models often overlook the temporal aspect and lack the ability to model multi-relational user-item interactions over time. To address these limitations, in this paper, we propose a Structure, Facticity, and Temporal information preservation embedding model (SFTe) to predict future interaction. Our model leverages the advantages of Temporal Knowledge Graphs (TKGs) that can capture both the multi-relations and evolution. We begin by modeling user-item interactions over time by constructing a Temporal Interaction Knowledge Graph (TIKG). We then employ Structure Embedding (SE), Facticity Embedding (FE), and Temporal Embedding (TE) to capture topological structure, facticity consistency, and temporal dependence, respectively. In SE, we focus on preserving the first-order relationships to capture the topological structure of TIKG. In the FE component, given the distinct nature of SIoT, we introduce an attention mechanism to capture the effect of entities with the same additional information for generating subgraph embeddings. Lastly, TE utilizes recurrent neural networks to model the temporal dependencies among subgraphs and capture the evolving dynamics of the interactions over time. Experimental results on standard future interaction prediction demonstrate the superiority of the SFTe model compared with the state-of-the-art methods. Our model effectively addresses the challenges of time-aware interaction prediction, showcasing the potential of TKGs to enhance prediction performance.

Graph entity dependencies (GEDs) are novel graph constraints, unifying keys and functional dependencies, for property graphs. They have been found useful in many real-world data quality and data management tasks, including fact checking on social media networks and entity resolution. In this paper, we study the discovery problem of GEDs—finding a minimal cover of valid GEDs in a given graph data. We formalise the problem, and propose an effective and efficient approach to overcome major bottlenecks in GED discovery. In particular, we leverage existing graph partitioning algorithms to enable fast GED-scope discovery, and employ effective pruning strategies over the prohibitively large space of candidate dependencies. Furthermore, we define an interestingness measure for GEDs based on the minimum description length principle, to score and rank the mined cover set of GEDs. Finally, we demonstrate the scalability and effectiveness of our GED discovery approach through extensive experiments on real-world benchmark graph data sets; and present the usefulness of the discovered rules in different downstream data quality management applications.

The Resource Description Framework (RDF) is widely used to model web data. The scale and complexity of the modeled data emphasized performance challenges on the RDF-triple stores. Workload adaption is one important strategy to deal with those challenges on the storage level. Current workload-adaption approaches lack the necessary generalization of the problem and only optimize part of the storage layer with the workload (mostly the replication). This creates a big performance gap within other data structures (e.g. indexes and cache) that could heavily benefit from the same workload adaption strategy. Moreover, the workload statistics are built collectively in most of the current approaches. Thus, the analysis process is unaware of whether workloads’ items are old or recent. However, that does not simulate the temporal trends that exist naturally in user queries which causes the analysis process to lag behind the rapid workload development. We present a novel universal adaption approach to the storage management of a distributed RDF store. The system aims to find optimal data assignments to the different indexes, replications, and join cache within the limited storage space. We present a cost model based on the workload that often contains frequent patterns. The workload is dynamically and continuously analyzed to evaluate predefined rules considering the benefits and costs of all options of assigning data to the storage structures. The objective is to reduce query execution time by letting different data containers compete on the limited storage space. By modeling the workload statistics as time series, we can apply well-known smoothing techniques allowing the importance of the workload to decay over time. That allows the universal adaption to stay tuned with potential changes in the workload trends.

Over the past two decades, many studies have devoted a good deal of attention to detect injection attacks in recommender systems. However, most of the studies mainly focus on detecting the heuristically-generated injection attacks, which are heuristically fabricated by hand-engineering. In practice, the adversarially-learned injection attacks have been proposed based on optimization methods and enhanced the ability in the camouflage and threat. Under the adversarially-learned injection attacks, the traditional detection models are likely to be fooled. In this paper, a detection method is proposed for the adversarially-learned injection attacks via knowledge graphs. Firstly, with the advantages of wealth information from knowledge graphs, item-pairs on the extension hops of knowledge graphs are regarded as the implicit preferences for users. Also, the item-pair popularity series and user item-pair matrix are constructed to express the user's preferences. Secondly, the word embedding model and principal component analysis are utilized to extract the user's initial vector representations from the item-pair popularity series and item-pair matrix, respectively. Moreover, the Variational Autoencoders with the improved R-drop regularization are used to reconstruct the embedding vectors and further identify the shilling profiles. Finally, the experiments on three real-world datasets indicate that the proposed detector has superior performance than benchmark methods when detecting the adversarially-learned injection attacks. In addition, the detector is evaluated under the heuristically-generated injection attacks and demonstrates the outstanding performance.

Incident detection (ID), or the automatic discovery of anomalies from road traffic data (e.g., road sensor and GPS data), enables emergency actions (e.g., rescuing injured people) to be carried out in a timely fashion. Existing ID solutions based on data mining or machine learning often rely on dense traffic data; for instance, sensors installed in highways provide frequent updates of road information. In this paper, we ask the question: can ID be performed on sparse traffic data (e.g., location data obtained from GPS devices equipped on vehicles)? As these data may not be enough to describe the state of the roads involved, they can undermine the effectiveness of existing ID solutions. To tackle this challenge, we borrow an important insight from the transportation area, which uses trajectories (i.e., moving histories of vehicles) to derive incident patterns. We study how to obtain incident patterns from trajectories and devise a new solution (called Filter-Discovery-Match (FDM)) to detect anomalies in sparse traffic data. We have also developed a fast algorithm to support FDM. Experiments on a taxi dataset in Hong Kong and a simulated dataset show that FDM is more effective than state-of-the-art ID solutions on sparse traffic data, and is also efficient.

When solving the problem of identifying similar records in different datasets (known as Entity Resolution or ER), one big challenge is the lack of enough labeled data. Which is crucial for building strong machine learning models, but getting this data can be expensive and time-consuming. Active Machine Learning (ActiveML) is a helpful approach because it cleverly picks the most useful pieces of data to learn from. It uses two main ideas: informativeness and representativeness. Typical ActiveML methods used in ER usually depend too much on just one of these ideas, which can make them less effective, especially when starting with very little data. Our research introduces a new combined method that uses both ideas together. We created two versions of this method, called DPQ and STQ, and tested them on eleven different real-world datasets. The results showed that our new method improves ER by producing better scores, more stable models, and faster learning with less training data compared to existing methods.

The growth of data-driven approaches typical of Machine Learning leads to an ever-increasing need for large quantities of labeled data. Unfortunately, these attributions are often made automatically and/or crudely, thus destroying the very concept of “ground truth” they are supposed to represent. To address this problem, we introduce HUM-CARD, a dataset of human trajectories in crowded contexts manually annotated by nine experts in engineering and psychology, totaling approximately $5000$ hours. Our multidisciplinary labeling process has enabled the creation of a well-structured ontology, accounting for both individual and contextual factors influencing human movement dynamics in shared environments. Preliminary and descriptive analyzes are presented, highlighting the potential benefits of this dataset and its methodology in various research challenges.

Heart failure, as a critical symptom or terminal stage of assorted heart diseases, is a world-class public health problem. Establishing a prognostic model can help identify high dangerous patients, save their lives promptly, and reduce medical burden. Although integrating structured indicators and unstructured text for complementary information has been proven effective in disease prediction tasks, there are still certain limitations. Firstly, the processing of single branch modes is easily overlooked, which can affect the final fusion result. Secondly, simple fusion will lose complementary information between modalities, limiting the network’s learning ability. Thirdly, incomplete interpretability can affect the practical application and development of the model. To overcome these challenges, this paper proposes the MDL-HFP multimodal model for predicting patient prognosis using the MIMIC-III public database. Firstly, the ADASYN algorithm is used to handle the imbalance of data categories. Then, the proposed improved Deep&Cross Network is used for automatic feature selection to encode structured sparse features, and implicit graph structure information is introduced to encode unstructured clinical notes based on the HR-BGCN model. Finally, the information of the two modalities is fused through a cross-modal dynamic interaction layer. By comparing multiple advanced multimodal deep learning models, the model’s effectiveness is verified, with an average F1 score of 90.42% and an average accuracy of 90.70%. The model proposed in this paper can accurately classify the readmission status of patients, thereby assisting doctors in making judgments and improving the patient’s prognosis. Further visual analysis demonstrates the usability of the model, providing a comprehensive explanation for clinical decision-making.

Anomalies in business processes are inevitable for various reasons such as system failures and operator errors. Detecting anomalies is important for the management and optimization of business processes. However, prevailing anomaly detection approaches often fail to capture crucial structural information about the underlying process. To address this, we propose a multi-Graph based Anomaly detection fraMework for business processes via grAph neural networks, named GAMA. GAMA makes use of structural process information and attribute information in a more integrated way. In GAMA, multiple graphs are applied to model a trace in which each attribute is modeled as a separate graph. In particular, the graph constructed for the special attribute activity reflects the control flow. Then GAMA employs a multi-graph encoder and a multi-sequence decoder on multiple graphs to detect anomalies in terms of the reconstruction errors. Moreover, three teacher forcing styles are designed to enhance GAMA’s ability to reconstruct normal behaviors and thus improve detection performance. We conduct extensive experiments on both synthetic logs and real-life logs. The experiment results demonstrate that GAMA outperforms state-of-the-art methods for both trace-level and attribute-level anomaly detection.