Accurate cloud resource forecasting is essential for proactive resource provisioning, maintaining Quality of Service (QoS), and reducing operational costs in dynamic cloud environments. The existing forecasting approaches predominantly estimate future CPU workload directly from historical resource traces, which often overlook the relationship between customer service demand and subsequent resource consumption. This study proposes a two-stage integrated forecasting model that explicitly models this dependency by first forecasting customer service requests, expressed as Transactions Per Second (TPS), and subsequently estimating future CPU workload from the TPS forecast. Both the forecasting component and resource prediction component employed the XGBoost model within a cascaded learning architecture, complemented by adaptive online retraining using an expanding-window strategy to address concept drift in continuously evolving cloud workloads. The proposed work was evaluated using real-world traces collected from a private cloud environment comprising ten applications. Experimental results demonstrate robust forecasting performance by achieving Symmetric Mean Absolute Percentage Error (SMAPE) below 7% for most applications, with the best-performing application achieving an MAE of 0.7372, RMSE of 1.1866, SMAPE of 3.57%, and an R2 of 0.9185. Horizon-wise drift analysis confirmed stable recursive forecasting behavior with controlled error accumulation across a 60-step prediction horizon. Compared with the conventional direct CPU forecasting method, the proposed two-stage integrated model gives improved forecasting robustness, computational efficiency, and interpretability, making it well-suited for proactive resource management and intelligent auto-scaling in cloud computing environments.

AI accelerators improve performance and efficiency for neural networks; evolving them for general-purpose computing requires new hardware, software, and programming models. In this article, we argue for the potential of general-purpose AI computing, highlight some challenges, and present initial results to motivate the feasibility.

Variable workloads and dynamic Quality of Service (QoS) demands make it difficult to manage cloud resources efficiently, especially in terms of CPU utilization. Over-provisioning of cloud resources leads to unnecessary energy consumption and costs, while under-provisioning decreases QoS and affects customer satisfaction. This study explores lightweight machine learning models to predict CPU utilization in cloud environments based on the services requested by cloud customers. Previous works mainly focus on traditional time-series-based CPU workload prediction, mostly using deep learning models, and did not address the problem of future CPU workload prediction for heterogeneous applications running in cloud environments. This study presents a predictive framework for CPU usage across multiple applications running in a private cloud. The proposed framework consists of several steps, including pre-processing, handling missing values and standardizing data, refining the target variable (CPU) through a rolling window approach, and evaluating various lightweight machine learning models (Bayes Ridge, Elastic Net, Lasso, Ridge, and MLP) on a dataset with 10 heterogeneous cloud applications. The experimental results demonstrate that the Ridge model shows robust performance in predicting CPU loads for most of the applications compared to the rest of the models, supporting intelligent resource management in cloud computing environments.

Accurate classification of underwater boulders is essential for a range of marine applications. Manual classificationmapproaches are labor-intensive, prone to subjectivity, and unsuitable for large-scale survey applications. In this study, we introduce a new benchmark dataset, SonarBoulder, constructed from high-resolution sonar imagery collected off the coast of Sweden. The dataset includes both balanced and imbalanced versions to evaluate model performance. We conducted a comparative analysis of seventeen classification models, including traditional machine learning models (e.g., Random Forest, KNN, SVM, Logistic Regression, AdaBoost, Naive Bayes, ANN, and XGBoost), deep learning methods (e.g., CNN, VGG16, ResNet50, ResNet101, EfficientNetB0, and InceptionV3), and transformer-based frameworks (e.g., DeepViT and ViT). On the balanced SonarBoulder dataset, the deep learning model ResNet101 achieved the best overall performance with 97.75% accuracy, 97.92% precision, and 97.66% F1 score, while the transformer model DeepViT achieved the highest recall at 98.59%. In the imbalanced scenario, InceptionV3 achieved the highest precision (97.26%), recall (95.83%), and F1 score (96.54%) as compared to the other methods. Our results demonstrate the effectiveness of using deep learning and transformer-based models for underwater boulder classification.

This study introduces and evaluates GraphTrace, a graph-based method for identifying crime hotspots suitable for CCTV placement. The method addresses key limitations in traditional spatial crime analysis techniques, such as rigid spatial divisions and reliance on heuristics, by dynamically modeling crime clusters with guaranteed distance constraints. We evaluate GraphTrace using five years of official crime data (N = 125,512) from Malmö, Sweden, and compare its performance against four established spatial methods: Grid+KDE, K-Means, HDBScan, and Greedy PAI Maximization. Each method uses crime data from one year to identify high-crime locations used as suggested CCTV camera placements, which are then evaluated based on their ability to capture crimes occurring within a specified radius in the following year. For example, hotspots identified from 2019 data are assessed against 2020 crime data by counting how many crimes that fall within the radius of each location. Performance is measured using total crime counts and the Predictive Accuracy Index (PAI). GraphTrace significantly outperforms all comparison methods (p<0.05) in terms of both crime capture and PAI. Effect sizes using Cohen’s d range from 0.14 to 1.98, demonstrating up to very large improvements in PAI. Despite its performance, GraphTrace maintains feasible runtimes and scales well. GraphTrace balances precision and computational efficiency by avoiding exhaustive pairwise comparisons while preserving spatial flexibility. Unlike grid-based methods, it does not segment the study area arbitrarily, and unlike many clustering heuristics, it enforces strict distance constraints. This study presents an initial evaluation and open-source implementation of GraphTrace for hotspot detection and CCTV placement, showing strong promise for spatial crime analysis.

With the rapid growth of internet technologies, IT businesses are transferring to cloud-based systems, and cloud-based services are in high demand among internet users. Therefore, appropriate allocation of resources in cloud computing environments is essential. The companies can reduce costs by saving energy by dynamically scaling up or down the number of active servers. In this context, this study presents a machine learning-based model for accurate prediction of CPU utilization. Previous studies employed timestamp-based data to predict CPU utilization in cloud computing, while the proposed work uses incoming user requests to predict CPU workload so that a timely decision can be made to scale up or scale down the servers in a cloud computing environment. The proposed model is based on several machine learning algorithms that are stacked into a single model called the stacking model for CPU workload prediction. The effectiveness of the proposed stacking model was tested on several evaluation metrics to validate its performance. Furthermore, the performance of the proposed stacking model is also compared with other state-of-the-art machine learning models such as support vector machines (SVM), decision trees (DT), random forests (RF), gradient boosting, and extreme gradient boosting (XGBoost).

In smart cities, Closed-Circuit Television (CCTV) cameras are crucial for enhancing law enforcement and public safety. Since CCTV cameras include both state-controlled and private-sector units in large numbers, it poses a considerable challenge to manage them. Automating the calculation of their coverage areas enables law enforcement and city planners to efficiently adapt surveillance strategies to the evolving needs of urban safety and dynamics. This paper proposes a prototype that automates the digitization of a vast number of camera coverage areas, such as for a large city. Given the positions for each camera, its sector width, and length-of-view, the prototype automatically identifies each camera's capture area represented as a polygon of positions. The prototype's generated capture areas are validated to the ground truth representing the true capture areas for 51 cameras (in the city of Malmo, Sweden) provided by the Swedish law enforcement agencies.

Preventing customer churn, i.e., termination of business commitments, is essential for companies operating in saturated markets, especially for subscription-based models such as telecommunication. Knowing when customers decide to terminate services is instrumental to effective churn prevention. In this study, we investigate how churn prediction performs in practice when training models on different time intervals of historic data (1-4 weeks back) and predicting churn at different numbers of weeks ahead (1-4 weeks). We use a real-world, time-series dataset of mobile subscription usage to examine churn prediction for business-to-business (B2B) customers. We utilize the time-series data at a higher temporal resolution than prior studies and investigate different forecasting horizons. Leveraging popular machine learning algorithms such as Random Forests, Gradient Boosting, Neural Networks, and Gated Recurrent Unit, we show that the best model achieves an average F-1-score of 79.3% for one-week ahead predictions. However, the average F-1-score decreases to 63.3% and 61.8% for two and four weeks ahead, respectively. A model interpretation framework (SHAP) evaluates the feature impact on the models' internal decision logic. We also discuss the challenges in applying churn prediction for the B2B segment.

We present a method, including tool support, for bibliometric mining of trends in large and dynamic research areas. The method is applied to the machine learning research area for the years 2013 to 2022. A total number of 398,782 documents from Scopus were analyzed. A taxonomy containing 26 research directions within machine learning was defined by four experts with the help of a Python program and existing taxonomies. The trends in terms of productivity, growth rate, and citations were analyzed for the research directions in the taxonomy. Our results show that the two directions, Applications and Algorithms, are the largest, and that the direction Convolutional Neural Networks is the one that grows the fastest and has the highest average number of citations per document. It also turns out that there is a clear correlation between the growth rate and the average number of citations per document, i.e., documents in fast-growing research directions have more citations. The trends for machine learning research in four geographic regions (North America, Europe, the BRICS countries, and The Rest of the World) were also analyzed. The number of documents during the time period considered is approximately the same for all regions. BRICS has the highest growth rate, and, on average, North America has the highest number of citations per document. Using our tool and method, we expect that one could perform a similar study in some other large and dynamic research area in a relatively short time.

The aim of the current study is to estimate whether spatial data on place features from OpenStreetMap (OSM) produce results similar to those when employing register data to predict future violent crime in public across three Swedish cities of varying sizes. Using violent crime in public as an outcome, four models for each city are produced using a Risk Terrain Modeling approach. One using spatial data on place features from register data and one from OSM, one model with prior violent crime excluded and one with prior crime included. The results show that several place features are significantly associated with violent crime in public independent of using register or OSM data as input. While models using register data seem to produce more accurate and efficient predictions than OSM data for the two smaller cities, the difference for the largest city is negligible indicating that the models provide similar results. As such, OSM place feature data may be of value when predicting the spatial distribution of future violent crime in public and provide results similar to those when using register data, at least when employed in larger compared to smaller cities. Possibilities, limitations, and avenues for future research when using OSM data in place-based criminological research are discussed.

Domain-Specific Hardware Accelerators (DSHA) are natural components in the evolution of general computers. However, designing and simulating hardware in Hardware Description Languages (MTh) often requires more effort for the developers and might not be suitable in all scenarios, which makes high-level language-based software simulators for computer hardware attractive. Yet, choosing which simulation framework to use can he challenging due to the lack of comparative studies of high-level language-based simulators. This paper presents a comparative evaluation of slate-of-the-art simulation frameworks that simulate computer hardware in high-level languages like C++. The contemporary simulators used in this study were selected from the 79 articles introducing novel AI accelerators referenced in our previous survey. We have identified six simulators that are suitable for AI accelerator evaluation, and provide a deeper analysis of three of them.

EU directives stipulate a systematic follow-up of train delays. In Sweden, the Swedish Transport Administration registers and assigns an appropriate delay attribution code. However, this delay attribution code is assigned manually, which is a complex task. In this paper, a machine learning-based decision support for assigning delay attribution codes based on event descriptions is investigated. The text is transformed using TF-IDF, and two models, Random Forest and Support Vector Machine, are evaluated against a random uniform classifier and the classification performance of the Swedish Transport Administration. Further, the problem is modeled as both a hierarchical and flat approach. The results indicate that a hierarchical approach performs better than a flat approach. Both approaches perform better than the random uniform classifier but perform worse than the manual classification.

During the last decade we have witnessed how artificial intelligence (AI) have changed businesses all over the world. The customer life cycle framework is widely used in businesses and AI plays a role in each stage. However, implementing and generating value from AI in the customer life cycle is not always simple. When evaluating the AI against business impact and value it is critical to consider both the model performance and the policy outcome. Proper analysis of AI-derived policies must not be overlooked in order to ensure ethical and trustworthy AI. This paper presents a comprehensive analysis of the literature on AI in customer life cycles (CLV) from an industry perspective. The study included 31 of 224 analyzed peer-reviewed articles from Scopus search result. The results show a significant research gap regarding outcome evaluations of AI implementations in practice. This paper proposes that policy evaluation is an important tool in the AI pipeline and empathizes the significance of validating both policy outputs and outcomes to ensure reliable and trustworthy AI.

Network anomaly detection for critical infrastructure supervisory control and data acquisition (SCADA) systems is the first line of defense against cyber-attacks. Often hybrid methods, such as machine learning with signature-based intrusion detection methods, are employed to improve the detection results. Here an attempt is made to enhance the support vector-based outlier detection method by leveraging behavioural attribute extension of the network nodes. The network nodes are modeled as graph vertices to construct related attributes that enhance network characterisation and potentially improve unsupervised anomaly detection ability for SCADA network. IEC 104 SCADA protocol communication data with good domain fidelity is utilised for empirical testing. The results demonstrate that the proposed approach achieves significant improvements over the baseline approach (average F_1 score increased from 0.6 to 0.9, and Matthews correlation coefficient (MCC) from 0.3 to 0.8). The achieved outcome also surpasses the unsupervised scores of related literature. For critical networks, the identification of attacks is indispensable. The result shows an insignificant missed-alert rate ( 0.3% on average), the lowest among related works. The gathered results show that the proposed approach can expose rouge SCADA nodes reasonably and assist in further pruning the identified unusual instances.

The evidence that burglaries cluster spatio-temporally is strong. However, research is unclear on whether clustered burglaries (repeats/near-repeats) should be treated as qualitatively different crimes compared to spatio-temporally unrelated burglaries (non-repeats). This study, therefore, investigated if there were differences in modus operandi-signatures (MOs, the habits and methods employed by criminals) between near-repeat and non-repeat burglaries across 10 Swedish cities, as well as whether MO-signatures can aid in predicting if a burglary is classified as a near-repeat or a non-repeat crime. Data consisted of 5744 residential burglaries, with 137 MO features characterizing each case. Descriptive data of repeats/non-repeats is provided together with Wilcoxon tests of MO-differences between crime pairs, while logistic regressions were used to train models to predict if a crime scene was classified as a near-repeat or a non-repeat crime. Near-repeat crimes were rather stylized, showing heterogeneity in MOs across cities, but showing homogeneity within cities at the same time, as there were significant differences between near-repeat and non-repeat burglaries, including subgroups of features, such as differences in mode of entering, target selection, types of goods stolen, as well the traces that were left at the crime scene. Furthermore, using logistic regression models, it was possible to predict near-repeat and non-repeat crimes with a mean F1-score of 0.8155 (0.0866) based on the MO. Potential policy implications are discussed in terms of how data-driven procedures can facilitate analysis of spatio-temporal phenomena based on the MO-signatures of offenders, as well as how law enforcement agencies can provide differentiated advice and response when there is suspicion that a crime is part of a series as opposed to an isolated event.

As machine learning and AI continue to rapidly develop, and with the ever-closer end of Moore’s law, new avenues and novel ideas in architecture design are being created and utilized. One avenue is accelerating AI as close to the user as possible, i.e., at the edge, to reduce latency and increase performance. Therefore, researchers have developed low-power AI accelerators, designed specifically to accelerate machine learning and AI at edge devices. In this paper, we present an overview of low-power AI accelerators between 2019–2022. Low-power AI accelerators are defined in this paper based on their acceleration target and power consumption. In this survey, 79 low-power AI accelerators are presented and discussed. The reviewed accelerators are discussed based on five criteria: (i) power, performance, and power efficiency, (ii) acceleration targets, (iii) arithmetic precision, (iv) neuromorphic accelerators, and (v) industry vs. academic accelerators. CNNs and DNNs are the most popular accelerator targets, while Transformers and SNNs are on the rise.

Customer satisfaction is an important aspect for any corporations customer support process. One important factor keeping the time customers’ wait for a reply at acceptable levels.By utilizing learning models based on the Random Forest Algorithm, the extent to which it is possible to predict e-Mail time-to-respond is investigated. This is investigated both for customers, but also for customer support agents. The former focusing on how long until customers reply, and the latter focusing on how long until a customer receives an answer.The models are trained on a data set consisting of 51, 682 customer support e-Mails. The e-Mails covers various topics from a large telecom operator. The models are able to predict the time-to-respond for customer support agents with an AUC of 0.90, and for customers with an AUC of 0.85. These results indicate that it is possible to predict the TTR for both groups. The approach were also implemented in an initial trial in a live environment.How the predictions can be applied to improve communication efficiency, e.g. by anticipating the staff needs in customer support, is discussed in more detail in the paper. Further, insights gained from an initial implementation are provided.

This paper describes work in progress of an interdisciplinary research project that focuses on the placement and analysis of public close-circuit television (CCTV) cameras using data-driven analysis of crime data. A novel web-based prototype that acts as a framework for the camera placement analysis with regards to historical crime occurrence is presented. The web-based prototype enables various analyses involving public CCTV cameras e.g., to determine suitable locations for both stationary CCTV cameras as well as temporary cameras that are moved around after a few months to address crime seasonality. The framework also opens up for other analyses, e.g. automatically highlighting crimes that are carried out closed by at least one camera. The research also investigates to what extent it is possible to generate estimates on the amount of detail captured by a camera given the distance to the crime light conditions. The research project includes interdisciplinary competences from various areas such as criminology, computer and data science as well as the Swedish Police.

Customer support can affect customer churn both positively and negatively. By identify non-routine e-mails to be handled by senior customer support agents, the customer support experience can potentially be improved. Complex e-mails, i.e. non-routine, might require longer time to handle, being more suitable for senior staff. Non-routine e-mails can be considered anomalous. This paper investigates an approach for context-based unsupervised anomaly detection that can assign each e-mail an anomaly score. This is investigated in customer support setting with 43523 e-mails. Context-based anomalies are investigated over different time resolutions, by multiple algorithms. The likelihood of anomalous e-mails can be considered increased when identified by several algorithms or over multiple time resolutions. The approach is suitable to implement as a decision support system for customer support agents in detecting e-mails that should be handled by senior staff.