In cloud platforms, the occurrence of failures in cloud services is usually accompanied by a substantial number of alarms, posing challenges to the operation, management, and maintenance. To address this issue, this paper proposes an alarm prediction framework based on propagation dependency and reinforcement learning for cloud services. First, a three-layer cloud service alarm knowledge graph is constructed, encompassing the device layer, service layer, and alarm layer, with the objective of integrating alarm data. Then, to filter out irrelevant alarms, an alarm compression schema is designed, taking into account both structural and semantic information. It achieves the preservation of propagation information by mining frequent alarm subtrees. After that, a reinforcement learning-based alarm prediction model is devised, including an alarm-driven reinforcement learning network and a global alarm discriminator. The former formulates the alarm propagation problem as a reinforcement learning inferencing task, and designs an immediate reward function to simulate the local propagation patterns of alarms. The latter guides the agent to learn alarm evolution patterns from a global perspective through an adversarial approach. Experimental results on two datasets demonstrate the effectiveness of our alarm compression and prediction models.

As a cloud-hosted virtual desktop service, cloud desktop supports various fields such as telecommuting, collaborative development, while enabling real-time user interaction through video stream. The stability of this process is determined by bandwidth, which significantly influences the user experience. Therefore, precise bandwidth prediction of video streams is essential in cloud desktops. This work proposes DeskPred for video stream transmission in cloud desktops, focusing on dynamic bandwidth prediction. In the startup stage, the limited data amount poses a challenge for achieving precise bandwidth predictions. We propose an Affinity-based Federated Learning algorithm, which leverages the historical records of high-affinity users for assisted training, all while protecting user privacy. During the long-term adjustment stage, we propose a Fluctuation-based Adaptive Incremental Prediction algorithm for independent training to address the issue of pattern forgetting. The algorithm considers both periodic features and instantaneous features, incorporating new patterns while revisiting previous knowledge through the memory module and Adversarial Elastic Weight Consolidation. We have verified DeskPred through an actual cloud desktop project supported by Lenovo Research. Through experiments conducted on a total of over 18 million data items (approximately 10 GB), DeskPred achieves the highest total score of 71.11%, making it highly suitable for cloud desktop environments.

In autonomous driving and robotic navigation, multi-sensor fusion technology has become increasingly mainstream, with precise sensor calibration as its foundation. Traditional calibration methods rely on manual effort or specific targets, limiting adaptability to complex environments. Learning-based calibration methods still face challenges, such as insufficient overlap between the fields of view (FoV) of multiple sensors and suboptimal cross-modal feature association, which hinder accurate parameter regression. Unlike traditional CNN-based networks, we propose a KansFormer-based self-Calibration Network for camera and LiDAR (KFCalibNet) that replaces fixed activation functions and linear transformations with learnable nonlinear activation functions. This enables the extraction of more fine-grained features from both image and point cloud, significantly enhancing the network's robustness in scenarios with limited FoV overlap. We also employ a multihead attention (MHA) module to compute correlations between image and point cloud features, significantly enhancing cross-modal feature association. To reduce learning complexity, we designed KansFormer with FastKAN as the feedforward network, enabling deep fusion and regression of fine-grained cross-modal features for accurate extrinsic calibration. KFCalibNet achieves an absolute average calibration error of 0.0965 cm in translation and 0.0234° in rotation on the KITTI Odometry dataset, outperforming existing state-of-the-art calibration methods. Moreover, its accuracy and generalization capability have been validated across multiple real-world railway lines.

Industrial environments pose distinctive challenges for anomaly detection, primarily stemming from the complexities associated with high dimensionality and the dynamic nature of data patterns over time. These properties determine that the model’s proper convergence on unlabeled data is unpromising, consequently leading to less efficient discrimination of anomalies in previous anomaly detection (AD) works. To address this problem, we present AnoDual, a novel, meta-learning AD framework. From the perspective of data reconstruction, we introduce the multi-memory enhanced VAE reconstructor M2ER, which learns to extract the most salient patterns in unlabeled noisy data through a self-supervised manner. This design eases impacts from potential anomalous components during data reconstruction, and enhances the discernibility of anomalies. To address performance degradation caused by the numerical deviation based AD scheme in most existing works, we design a dual-source self-supervised discriminator DSD, which examines characteristics in the domain of representations. This model actively assesses discrepancies between data pairs and representation pairs in parallel, and conducts AD on a fine-grained scale. In this way, anomalies that used to be unnoticed due to a less prominent numerical deviation can be spotted. Besides, we propose a meta-learning powered training pipeline to enable model training even when no real label is available, which is common in the industry. Extensive experiments on five large-scale real-world industrial datasets suggest that AnoDual achieves an average F1-Score with a substantial increment of 3.39 %, outperforming the latest state-of-the-art baseline. Note to Practitioners—A generative model plus a numerical threshold based detection approach currently takes a significant share in both academia and the industry. However, the performance of this workflow is not promising in actual applications, with multiple factors contributing to this situation. The proper convergence of such generative models is difficult when the training material contains noisy samples - an over-expressed generative model would result in less significant reconstruction discrepancies for anomalies that are hard to notice. In addition, selecting a numerical threshold, which is used to spot anomalies, requires multiple laborious attempts, and can hardly adapt to an ever-changing pattern in industrial environments. These circumstances make it challenging to apply prior works in practical production, which, in turn, urges the need to develop an effective methodology to address the need for industrial anomaly detection. This manuscript includes a novel, meta-learning powered framework AnoDual, which is tailored for industrial scenarios. This framework discards the conventional design of comparing the reconstruction error numerically, but introduces a solution based on the differentiation of the representations. Besides, the multi-head attention enhanced variational autoencoder also leads to a much more pronounced discrepancy for anomalous samples, which benefits their successful detection. Providing a flexible and robust way to detect anomalies on deployed IoT assets, this work can be further transformed to serve applications in many other domains.

With the widespread application of big data and intelligent information systems, the tenant has become the main form of most scenarios. As a data mining technique, the portrait has been widely used to provide targeted services. Therefore, we transfer the traditional user-driven portrait into tenant driven for churn prediction. To achieve it, this paper first proposes a three-layer architecture and defines the fine-grained features for creating portraits from the perspective of tenants. In a large-scale telecommunication industry dataset of 100,000 tenants, we construct the tenant portrait through the proposed framework, and analyze the influences of the defined features on churn possibility. Then, considering the information missing caused by privacy concerns, we come up with theCrossMatch, a portrait completion model based on semi-supervised and graph convolution, which combines the relation characteristics among tenants for recovering missing information. On this basis, we design the tenant churn prediction method based on a directed attention network. Moreover, we recover missing information on three public node datasets withCrossMatch, achieving around 1-2$\%$improvement. We then apply the directed attention network for churn prediction and achieve an Accuracy of 75.06$\%$, Precision of 77.78$\%$, and F1-score of 71.43$\%$, which outperforms all the baselines.

Most existing works in Industrial Internet of Things (IIoT) anomaly detection either depend on computationally intensive models that exceed the capabilities of multiaccess edge computing (MEC) servers, or lightweight models that lack robustness, making them unadaptable in IIoT infrastructures. To address these challenges, we proposeTHREADS, a hierarchical anomaly detection framework tailored for IIoT applications. TheInstance threadutilizes an efficient variational auto encoder to produce instant feedback and offloads most of the workload to MECs. On the other hand, theShadow threademploys an attention-enhanced transformer discriminator to examine low-confidence results in the cloud. Experimental results on five large-scale datasets showTHREADSachieves an averageF1-Scoreof 0.8537 in the hierarchical mode where most of the workloads are handled by MECs, and the random access memory and CPU usage is reduced by up to 29% and 88%, respectively. Meanwhile,THREADSachieves anF1-Scoreof 0.8563 in a cloud-based mode, consistently outperforming state-of-the-art approaches.

Accurately perceiving service states of the Internet of Things (IoT) is crucial for maintaining system stability and long-term sustainability. Recently, Large Language Models (LLMs) have emerged as a novel technological advancement, offering unprecedented possibilities in various domains due to their advanced comprehension capabilities. However, the limited data availability problem poses a challenge to the integration of domain knowledge into LLMs and the enhancement of their ability to perceive service states. In this article, a novel adversarial augmentation-based domain LLM framework for IoT services is proposed to solve this problem. We first construct an LLM fine-tuning dataset through an elaborate hierarchical prompt template, which integrates task-specific instructions, heterogeneous service data, and statistical features to the service contextual description. Then, inspired by using specialized compact small-models to enhance the capabilities of LLMs, we design two dedicated domain-specific small-models, including a service topology prediction model and a running status prediction model, to capture evolution patterns of services from different perspectives. On this basis, we design an LLM fine-tuning framework based on the domain adversarial distillation, which comprises an LLM parameter tuning module and a small-model adversarial distillation module. The former leverages the Low-Rank Adaptation (LoRA) algorithm to fine-tune the LLM. The latter distills the domain knowledge of small-models into the LLM through aligning the latent vector distributions of the LLM with those of small-models. Following the adversarial training, the LLM will be equipped with domain capabilities of small-models. Experimental results on two datasets demonstrate the effectiveness of our LLM framework.

The popularity of cloud services has provided a new medium for video streams transmission. Cloud desktops, as a representative multimedia application, facilitate interaction between users and cloud via video streams, garnering widespread adoption in various fields. The network condition directly affects the transmission. Therefore, accurate throughput prediction helps guide the allocation of network resources, avoiding a decline in user experience due to insufficient resources and waste caused by excessive resources. Recent works focus more on the temporal characteristics of throughput. However, we believe that throughput of video streaming is significantly influenced by usage scenario. In this paper, we propose a transfer-based autoencoder framework DeskTransfer for throughput prediction in frequent switching cloud desktop scenarios. Specifically, we construct the Scenario Autoencoder and Throughput Autoencoder to respectively learn the scenario and throughput features from historical usage records. By adopting an adversarial mechanism, we design transfer algorithm using latent vectors, enabling the model suitable for multiple scenarios. We collect real-world data from a project cooperated with Lenovo Research for experiment and compare our solution with leading methods on public datasets to validate its effectiveness.

To identify pattern deviations in large-scale industrial infrastructures, anomaly detection is crucial yet challenging. Previous research has not adequately addressed the characteristics and deployment considerations in these complex scenarios. In this paper, we presentInoU, a scalable anomaly detection framework to process unlabeled multivariate time-series data. We incorporate a VAE filter to ease impacts from noisy components in training materials. We propose a scalable trio-attention U-Transformer to construct the typical representation of high-dimensional streams and produce pseudo labels that enable the later training process. The ultra perception and intra-/ inter-flow attention mechanisms are delicately designed to aggregate information from different flows with variable granularities while keeping a global view of the data. Its nested structure helps to maintain high efficiency even when the model is scaled down. We introduce a Siamese discriminator that projects target data into manifolds, and collates discrepancies at the embedding level. This paradigm elevates detection performance far beyond segment-wise error comparison in prior works. We apply contrastive and adversarial learning techniques to optimize manifold projection and detection performance when processing unseen samples. Extensive experiments on five large-scale datasets demonstrate the effectiveness ofInoUwith an averageF1-Scoreimprovement of 5.58%, significantly outperforming the state-of-the-art.

Due to the unrestricted location of cloud resources, an increasing number of users are opting to apply for them. However, determining the appropriate resource quota has always been a challenge for applicants. Excessive quotas can result in resource wastage, while insufficient quotas can pose stability risks. Therefore, it's necessary to propose an adaptive quota scheme for cloud resource. Most existing researches have designed fixed quota schemes for all users, without considering the differences among users. To solve this, we propose an adaptive cloud quota scheme through dynamic portraits and task-resource optimal matching. Specifically, we first aggregate information from text, statistical, and fractal three dimensions to establish dynamic portraits. On this basis, the bidirectional mixture of experts (Bi-MoE) model is designed to match the most suitable resource combinations for tasks. Moreover, we define the time-varying rewards and utilize portrait-based reinforcement learning (PRL) to obtain the optimal quotas, which ensures stability and reduces waste. Extensive simulation results demonstrate that the proposed scheme achieves a memory utilization rate of around 70%. Additionally, it shows improvements in task execution stability, throughput, and the percentage of effective execution time.

Indoor localization availability is still sporadic in industry, especially at the black-hole, i.e., there only exist cellular signals, no GPS or WiFi signals. Based on our 2-year observations at the DiDi ride-hailing platform in China, there are$ 68\,\text{k}$orders everyday created at black-hole. In this paper, we presentTransparentLoc, a large-scale cellular localization system for pickup position recommendation of the DiDi platform. Specifically, we design a CNN model for real-time localization based on a crowdsourcing fingerprint set constructed by outdoor trajectories and abnormal cell tower detection. Then we leverage a DeepFM model to recommend an optimal pickup position for passengers. We share our 2-year experience with 50 million orders across 13 million devices in 4541 cities to address practical challenges including sparse cell towers, unbalanced user fingerprints, temporal variations, and abnormal cell towers in terms of four major service metrics, i.e., pickup position error, over-30-meters ratio, cancel ratio, and call ratio. The large-scale evaluations show that our system achieves a$ 0.54\,\text{m}$lower median pickup position error compared to the iOS built-in cellular localization system, regardless of environmental changes, smartphone brands/models, time, and cellular providers. Additionally, the over-30-meters ratio, cancel ratio, and call ratio have significant reductions of 0.88%, 0.88%, and 5.13%, respectively.

With the rapid development of Server Cluster System technologies, more and more services and tasks are deployed on the Server Cluster Systems. Despite its convenience, a large number of services are mixed together and affect each other, which brings great difficulties to the detection of equipment fault and causes great losses. Nowadays, artificial intelligence for IT operations (AIOps), which utilizes data analysis and machine learning to improve the operation quality of Server Clusters, has been proposed to solve this problem. However, the existing solutions of AIOps cannot be applied to different scenarios. Therefore, in this paper, we propose a novel fault detection method for single indicators which is adaptive to various scenes. To enrich the representation of data, we first propose an interval-volatility-rate method to extract the context features of data in a fixed interval. Based on this, the convolutional neural network and long short-term memory network are employed to get the locally spatial features and temporal features of data. After that, the spatial features and temporal features are combined and input to an MLP network to perform the fault prediction. Additionally, a k − σ principle is designed to promote the sensitivity of fault detection. Experimental results show that our method outperforms other competitive models and has better scalability.

One of the most challenging problems in the question answering (QA) area is the lack of high-quality labeled data. However, the cost of manually labeling a question-answer (QA) pair from the target text is very high. One way to solve this problem is to automatically generate QA pairs from the target text. In this paper, we propose the Batch Normalization conditional VAE-QA pair generation (BNCVAE-QAG) model to generate QA pairs for a given text. First, we employ the Batch Normalization (BN) layer to prevent Kullback-Leibler (KL) divergence from disappearing. We also design modules to extract spatiotemporal features from text contents, in addition, the self-attention mechanism is employed in the question decoder, which improves the accuracy and recall of results. Furthermore, we propose a question generalization mechanism to generate more QA pairs. We evaluate our BNCVAE-QAG model on several datasets. The experimental results show that our model has achieved an impressive performance improvement than the baseline.

With the rapid increase of users, online meeting platforms have accumulated massive meeting transcripts. However, it is still a challenge for users to quickly master the chief information and manage the meetings, despite there are already some useful text summarization models. In this paper, a Knowledge Graph-based Meeting Summarization Framework is proposed to tackle this challenge. First, a two-layers meeting domain Knowledge Graph is developed to integrate more information of meetings. Based on which, an encoder-decoder architecture is utilized to summarize meetings. For encoding meetings, a structural-level and semantic-level embedding strategy is considered, concretely, the Knowledge Graph is embedded to obtain the structural information, an interaction intention recognition model and a two-level transformer mechanism are devised to get the semantic information. Finally, the structural information and semantic information are combined and fed into the decoding network to generate meeting summaries. Extensive experiments on the Chinese meeting dataset show that our summarization framework outperforms other state-of-the-art models.

With the deepening of the digital transformation, enterprises have accumulated plenty of electronic files. For users, how to quickly find their desired files from the massive resources is a big challenge, which is also a critical concern for enterprises. Although a series of enterprise search engines have emerged in the market, there is a lack of search systems for enterprise files. In this paper, we design a novel enterprise file search framework, named QianXun, which supports searching in both Chinese and English. By introducing technologies such as automatic text summarization, online file previewing, and Knowledge Graph-enhanced search, our proposed framework improves the efficiency of file search as well as the user experience, which is of great significance to enterprises.

In recent years, with the growth of network hosts and applications, Abnormal host behavior brings great challenges to the protection of user information privacy. Effective detection of abnormal host behavior is of great significance. Host anomaly detection technology can identify abnormal behaviors of the host by establishing the normal behavior benchmark, which is universal and capable of detecting new network attack behavior. However, the existing host behavior anomaly detection technology confronts with three problems(i.e., a large amount of data, difficulty in labeling, and strong manual dependence). To solve these problems, this paper proposes an unsupervised learning model based on BiGAN(Bidirectional Generative Adversarial Network) to detect abnormal host behavior. Our model introduces several loss functions and modifies them to apply to BiGAN. At the same time, we discuss which loss function is more suitable for host behavior anomaly detection. Experimental results show that the BiGAN model using the Hinge loss function is the most stable and has excellent anomaly detection performance.

Nowadays, Scratch, as a widely-used educational programming platform, has gathered a huge number of programming users all over the world. Facing massive programming resources, how to make satisfactory programming recommendations has attracted increasing attention, especially on explainable recommendations. Existing Scratch recommendation systems overlook to provide why a project is recommended, which prevents users from making better decisions and trusting in the system. To resolve this problem, we design the Scratch-RL, an explainable reinforcement learning framework over knowledge graphs for Scratch recommendation. First, we devise a preference-driven Actor-Critic network to simulate users' local preferences and explore the potential interested projects along the reasoning paths. In the Actor-Critic network, we elaborate a preference state function, a preference-based reward function, and a preference-conditional action pruning strategy for the agent. Then, we leverage a directive discriminator network to help evaluate the correctness of recommendations from the agent and return an extra guidance reward accordingly. A high guidance reward is given when the agent generates correct recommendations, which guarantees that the agent quickly and accurately comprehends the preferences of users. Finally, we jointly train the Actor-Critic network and the discriminator, when the whole training is done, the reasoning paths are taken as the interpretability of the recommendations. Extensive experiments on both the Scratch data set and public data set show that, Scratch-RL obtains favorable recommendation results compared with the state-of-the-art models.

With the rapid increase of users, Scratch, as a popular online social and programming platform, has accumulated massive project resources and complex relations across its social and programming learning network. However, it is challenging to utilize the network information for providing Scratch users with personalized services, despite there are already some useful network representation learning models. In this paper, a network representation learning model with preference-based generative adversarial nets for Scratch (ScratchGAN) is proposed to resolve this problem. In ScratchGAN, we first design a node-vector initialization approach to preserve structure information and side information of Scratch network. Then, considering to learn the fine-grained user preference information of network, we propose a novel Scratch adversarial learning model which includes a Scratch generative adversarial net and a user preference difference constraint component. The former aims to capture user preferences through a new generating strategy based on the delivery nature of preference. The latter attempts to embed users' detailed preference differences according to their interaction behaviors. ScratchGAN can mine user preferences while preserving network structure information and side information. Extensive experiments on the Scratch network show that ScratchGAN outperforms other state-of-the-art models in link prediction and recommendation tasks.

In recent years, Scratch has been a popular programming platform for young children. To help children express emotions for projects, Scratch provides children with a music module to create desirable background music. However, in Scratch, there is not a tool helping recognize the emotion of music. Besides, as Scratch-generated music differs from regular music, existing music emotion recognition models perform poor in Scratch-generated music. To overcome it, in this paper, we propose a novel music emotion recognition model for Scratch-generated music. First, we build a Scratch-generated dataset by the main melody extraction algorithm. Then, for each music, we extract their underlying features and input them to the CNN module. After that, the features learned by CNN are input to RNN to get the final classification results. In our model, the CNN module can learn the important features of music while RNN can learn the sequential features. The experimental results show that the proposed model performs better than traditional music emotion recognition models.

Scratch is a popular programming platform with plenty of learning resources. However, it is quite difficult for users to find suitable resources. In this paper, in order to provide the resources which the users require, we propose a Scratch Recommendation Framework Adaptive to User Interest Drifts (SRFA-UID). First, a user interest drifts model is designed, which adopts the time decay factor and the weights of operation behaviors to track users' dynamic interest. Then, on the basis of users' current and historical interest, we calculate their combined user similarity. Next, we present a novel two-hop-algorithm to update users' friend community. Considering the preferences of the whole friend community, the Computational Thinking (CT) skills of users and the impact factors of items, we put forward a recommendation function to obtain items that are related to users. For an item, the function can calculate its F value to determine if we can recommend it to the users. Experimental results show that SRFA-UID performs better than other state-of-the-art methods in the Scratch dataset.

Poetry is a very popular literary form. Currently, its sentiment analysis is one of the hottest research trends. However, there are few relevant studies focusing on the sentiment analysis of ancient Chinese poetry, especially for Tang Poetry. In this paper, we propose a deep learning-based method to solve the above problem. Specifically, we combine Convolutional Neural Network and Gate Recurrent Unit to better extract the characteristics of Tang poetry. In addition, considering the special structural characteristics of Tang poetry, a multi-channel processing model is used to reshape the feature vector of sentences. Finally, in order to verify the rationality and superiority of the proposed methods, we construct a dataset by labeling more than 2500 representative Tang poems. The experimental results prove that our scheme has a higher accuracy rate, up to 64%, compared against three other competing methods.

BlockPy is a block-based program language which has both block-based interface and traditional text-based interface. It fills the gap between block-based programming and language coding. But there is little work that focuses on the Computational Thinking(CT) evaluation of BlockPy programs. In this paper, we design and implement a BlockPy Analysis Tool to assess the CT skills of BlockPy programs automatically. We use Python's built-in AST module to analyse each node in the abstract syntax tree(AST) of each BlockPy program. Then, considering the characteristics of Python language, we propose a new CT Evaluation Criteria based on Scratch Analysis Tool(SAT). Under the guidance of the CT Evaluation Criteria, we propose a detailed scoring program to analyze each node of the program and get the CT score. Experimental results show the superiority of our tool compared with other analysis tools.

Review spam filtering of Chinese is a hot research topic in the field of natural language processing. In recent years, there have aroused a lot of neural network models for review spam filtering of Chinese, but these models mainly focus on utilizing text semantics or part-of-speech of a review without considering the variants of a sensitive word. In this paper, based on TextCNN, we propose a novel review spam filtering model of Chinese that embeds various features of the text. Considering the common types of variant words, we first extract three sets of vectors for each Chinese character of the review, including the single character vector, the pinyin vector, and the pinyin vector of each single character. Then, we use TextCNN to learn the features of these vectors, respectively. After that, the learned features are added as the comprehensive features of review. Finally, we input it to the softmax layer to get the final classification results. The experimental results show that the joint model performs better than classic classification methods in review spam filtering of Chinese. The recall and F1-score on review spam reach 92.4% and 93.6%.

In recent years, Scratch has been widely used in helping teenagers learn to program. Before programming with Scratch, users usually need to select a background image with a proper style to set the environment of role activities and emotional tone of works. In order to make the image styles rich and diverse, users can use the fast neural style transfer method to generate the image. However, when users use this method, they are usually puzzled with the optional styles, which makes it a common phenomenon that the contents of transformed images do not match the chosen styles. To resolve this problem, we design a novel image classification model to help recognize the scene of the image, which can guide users to select optional styles for the image. First, we improve the residual network with two attention modules and reconstruct the residual module structure, which improves the accuracy of model classification. Then, we propose a self-adjusting learning rate module, which can accelerate the convergence of the model and reduce fluctuations of the loss function. The experiment results show that our model outperforms other classic image classification methods in classifying the background image of Scratch. The classification accuracy on the testset reaches 98%.