Machine learning for Internet of Things data analysis: A survey

1. Introduction

• IoT combines communications, sensors, actuators, and Internet-connected physical objects to generate data for improved services and intelligent system performance.Systems must access raw data from diverse networked sources and analyze it to extract knowledge.Source: §1 · p. 2
• Data science, including data mining and machine learning, is positioned as a means to discover patterns and insights from IoT’s growing data resources across volume, variety, and velocity dimensions.Analytics requires defining data types, selecting models such as neural networks, classification, and clustering, and applying efficient algorithms suited to the data.Source: §1 · p. 2
• The review identifies three central challenges: matching algorithms to data characteristics, handling real-time scale and velocity, and finding suitable models for IoT pattern recognition and analysis.These challenges also create opportunities for new developments.Source: §1 · pp. 2–3

1.1. The Contribution of this paper

• The paper’s contribution is framed by questions about how machine-learning algorithms can be applied to IoT smart applications.The supplied passage introduces this objective but does not provide the remaining questions or their answers.Source: §1.1 · p. 3

• It asks what taxonomy of machine learning algorithms can be used.Source: data? · p. 3

be adopted in IoT?

• The survey connects IoT applications, IoT data characteristics, and data-driven machine-learning algorithms to determine suitable processing and decision-making methods.It also discusses issues arising from these topics.Source: be adopted in IoT? · p. 3
• About 70 reviewed articles reveal eight major algorithm groups applicable to IoT data, categorized by structural similarity, data type, and practical processing scale.Source: be adopted in IoT? · p. 4
• The survey analyzes real-world IoT data characteristics and reports findings from more than 20 authors, discussed in Section 6 and Table 1.Source: be adopted in IoT? · p. 4
• Smart City is the primary IoT use case because 60 percent of reviewed articles focus on it, it incorporates many IoT use cases, and accessible datasets support research.Source: be adopted in IoT? · p. 4
• An SVM implementation uses Aarhus City smart traffic data to predict traffic hours during one day.Source: be adopted in IoT? · p. 4

1.2. Organization

• The paper reviews related work, IoT applications and architectures, data characteristics, eight machine-learning categories, Smart City matching, and future issues across Sections 2–7.Source: §1.2 · p. 5

• IoT research faces important challenges in preparing and processing data because IoT is a developing area involving smart data.Source: §2 · p. 5
• [9] proposed four IoT data-mining models: multilayer, distributed, grid-based, and multi-technology integration approaches.The models address layered processing, distributed deposition, heterogeneous large-scale high-performance applications, and future-Internet integration.Source: §2 · p. 6
• [10] investigated warehousing and mining RFID stream data, while [11] reviewed classification, clustering, association, time-series, and outlier-detection functions.[11] connected data from e-commerce, industry, healthcare, and city governance with IoT-like characteristics and mapped mining functionality to applications.Source: §2 · p. 6
• [12] organized IoT data-mining challenges around explaining data, algorithms, potential, and open issues, identifying three concerns for mining IoT data.These concerns are demonstrating that processing solves the chosen problem, extracting data characteristics, and selecting algorithms according to data characteristics and an algorithm taxonomy.Source: §2 · p. 6
• [13] reviewed Smart City IoT infrastructure, including enabling technologies, protocols, architectures, and advanced communication for value-added services.The review addresses services for city administrations and citizens and includes discussion of Padova Smart City data.Source: §2 · p. 7

3. Internet of Things

• IoT comprises connected devices that automatically transfer data to optimize performance, using sensors, processing networks, data analysis, and system monitoring.Its broader purpose is to create smarter environments and simplify lifestyles by saving time, energy, and money.Source: §3 · p. 7

3.1. Computing Framework

• IoT processing uses fog and cloud frameworks according to application needs and whether data must be processed immediately or later.Processing and preparation may involve edge analytics, stream analysis, database analysis, fog, or cloud computing before data transfer.Source: §3.1 · p. 8, §3 · p. 8

3.1.1. Fog Computing:

• Fog computing moves information-processing tasks toward edge servers, providing limited computing, storage, networking, logical intelligence, and data filtering for data centers.The architecture has been implemented in eHealth and military applications.Source: §3.1.1 · p. 8

3.1.2. Edge Computing:

• Edge computing processes data initially on edge devices, which may require local master-data copies for offline operation and support security, cleaning, filtering, and local storage.This architecture runs processing away from the core and toward the network edge.Source: §3.1.2 · p. 9

3.1.3. Cloud Computing:

• Cloud computing sends IoT data to data centers for analysis, but its high latency, load balancing, and CPU demands make it insufficient for much high-speed IoT processing.Big Data volume increases cloud-server CPU usage [23].Source: §3.1.3 · p. 9

3.1.4. Distributed Computing

• Distributed computing divides high-volume IoT data into packets processed by different computers, reducing network loading, CPU usage, and energy consumption while increasing speed and volume capacity.Frameworks such as Hadoop and Spark support this approach, which addresses repeated sensor-generated Big Data challenges.Source: §3.1.4 · p. 10

4. Smart City

• IoT connects and monitors city assets, enabling live status reports and smarter responses while improving traffic, water, energy, safety, transportation, and resource management.Smart-city objectives include improving citizens’ quality of life, reducing public-service expenses, and supporting economic activity.Source: §4 · p. 10

4.1. Use Case

• Smart-city use cases include Smart Energy, smart mobility, Smart Citizens, and urban planning, supporting more comfortable lives and data-informed urban services.These use cases span energy, transportation, citizen-focused monitoring, and long-term city decisions.Source: §4.1 · p. 11

4.1.1. Smart Energy

• Smart Energy targets lower power consumption through smart grids, time-series monitoring, renewable resources, and leak-monitoring systems for water or gas management.Smart grids require re-engineering electricity services and generate high-speed time-series data for monitoring key devices.Source: §4.1.1 · p. 11

4.1.2. Smart Mobility

• Smart mobility uses vehicle, camera, and counter-sensor data for autonomous driving, congestion-aware traffic control, and more accurate, real-time public transportation.Public transportation systems can provide location and routing information, support passenger scheduling, and combine real-time decisions with historical-data analysis.Source: §4.1.2 · p. 12

4.1.3. Smart Citizen

• Smart Citizen applications cover environmental, crime, and social-health monitoring, with closer observation intended to enhance security and overall social health.Source: §4.1.3 · pp. 12–13

4.1.4. Urban Planning

• IoT-based urban planning combines data from different sources to predict future crowding, address potential problems, and schedule infrastructure improvements.Data-informed decisions affect city infrastructure, design, and functionality.Source: §4.1.4 · p. 13

4.2. Smart City Data Characteristics

• Smart-city data is continuous, large-scale, dynamic across time and location, heterogeneous in quality, and generated at different rates, creating integration and information-loss challenges.GPS sensors may update every few seconds while temperature sensors may update hourly; quality is affected by measurement error, environmental noise, and discrete observations.Source: §4.2 · pp. 13–14
• Smart-city data require higher-level abstraction to provide actionable information and improve Quality of Information (QoI).QoI depends on the application and the data characteristics.Source: §4.2 · p. 14
• Trustworthy sources, multi-resource integration, and increased sampling frequency and density improve data accuracy, precision, and environmental-noise reduction.These measures improve the quality of observations and measurements.Source: §4.2 · p. 14
• Figure 2 presents the data characteristics of IoT and smart-city data.Source: §4.2 · p. 14
• Semantic annotation enhances data quality by making raw, low-level device data interpretable and supporting knowledge-based information fusion [28, 42].Semantic models describe data, quality, and original attributes, while smart-data characteristics are summarized in Table 1.Source: §4.2 · p. 14

• Machine learning is a subfield of computer science.Source: §5 · p. 14
• It is a type of artificial intelligence that enables machines to learn without explicit programming.Source: §5 · p. 14
• Machine learning evolved from pattern recognition and computing.Source: §5 · p. 14

Smart City

• The article focuses on supervised and unsupervised learning for IoT smart-data analysis, covering classification, regression, clustering, and feature extraction.Supervised learning uses labeled inputs, whereas unsupervised learning identifies clusters or internal representations; reinforcement learning is not the focus.Source: Smart City · pp. 16–17

5.1.1. K-Nearest Neighbors

• KNN classifies unseen points using the majority label among K nearest training neighbors, selected with metrics such as Euclidean, Mahalanobis, or Hamming distance.KNN requires storing the entire training set, limiting scalability, although tree-based searches and online variants address this issue.Source: §5.1.1 · pp. 17, 19

5.1.2. Naive Bayes

• Naive Bayes classifies new points with Bayes’ theorem under conditional feature independence, using different distributions to estimate class priors and feature likelihoods.The classifiers require few training points, handle high-dimensional data, and are fast and highly scalable.Source: §5.1.2 · pp. 19–20

5.1.3. Support Vector Machine

• SVMs find maximum-margin separating hyperplanes, use slack variables and penalty parameter C for outliers, and support nonlinear and multiclass classification.Nonlinear SVMs use the kernel trick to map inputs into high-dimensional feature spaces, while multiclass approaches include one-vs-all and all-vs-all.Source: §5.1.3 · pp. 20–21
• SVMs effectively handle high-dimensional data sets and use memory efficiently through support vectors, but do not directly provide probability estimates and have difficult-to-interpret parameters.Source: §5.1.3 · p. 21

5.2.1. Linear Regression

• Linear regression learns a function by mapping basis-function representations φ(x) to y as a linear combination of a fixed set of basis functions.Examples include polynomial, Gaussian radial, and sigmoidal basis functions.Source: §5.2.1 · p. 22

5.2.2. Support Vector Regression

• Support Vector Regression (SVR) extends SVMs to regression, with models depending only on a subset of training points after rejecting points close to the model prediction; implementations include epsilon-support and nu-support regression.Source: §5.2.2 · pp. 22–23

5.3.1. Classification and Regression Trees

• CART partitions the input space into axis-aligned cuboid regions, assigns each region a separate classification or regression model, and uses a greedy top-down heuristic because finding optimal tree structure is NP-complete.Source: §5.3.1 · pp. 23–24

5.3.2. Random Forests

• Random forests train many trees on bootstrap samples and random feature subsets, aggregating predictions by mode for classification and mean for regression.They offer strong accuracy, speed, and scalability at the cost of human interpretability, while the feature-subset size trades off weak and overly similar trees.Source: §5.3.2 · p. 25

5.3.3. Bagging

• Bagging improves model accuracy and stability while reducing overfitting by training models on bootstrap-resampled datasets and using majority voting for classification.Each generated training set is formed by randomly sampling data points with replacement.Source: §5.3.3 · pp. 25–26
• Bagging combines model outputs using the mean for regression tasks.Source: §5.3.3 · p. 26
• Bagging can improve results for models such as CART and neural networks.Source: §5.3.3 · p. 26
• Bagging can degrade stable models such as KNN; applications include customer attrition prediction and preimage learning.Source: §5.3.3 · p. 26

5.4.1. K-means

• K-means partitions unlabeled data into K similarity-based groups by minimizing distances between points and their nearest cluster centers.Binary indicators represent assignments, while the algorithm learns the centers and assignments iteratively; it is described as fast and highly scalable.Source: §5.4.1 · p. 26
• K-means has an online stochastic version, while related IoT applications include MapReduce clustering of many small datasets, travel-pattern regularities, and real-time sensor-event processing.The latter application uses OpenIoT middleware as an interface.Source: §5.4.1 · pp. 26–27
• K-means has limitations from Euclidean similarity, including restricted variable types, non-robust centers under outliers, and forced single-cluster assignments.These properties can produce inappropriate clusters in some cases.Source: §5.4.1 · p. 27

5.4.2. Density-Based Spatial Clustering of Applications with Noise

• DBSCAN clusters dense regions, treats low-density points as outliers, detects arbitrary-shaped clusters, and determines their number from density rather than requiring K.It is efficient, fast, and robust against outliers, but struggles with differing densities and is sensitive to the distance metric.Source: §5.4.2 · p. 28

5.5.1. Principal Component Analysis

• PCA projects data onto an L-dimensional linear subspace whose basis vectors maximize projected variance and equivalently minimize average reconstruction error.The projected data and basis vectors are obtained from eigenvectors associated with the largest eigenvalues.Source: §5.5.1 · pp. 28–29
• PCA runtime depends on how its basis is calculated, with reported complexities of O(M^3), O(LM^2), O(NM^2), and O(N^3).An iterative Expectation Maximization variant avoids explicitly calculating the covariance matrix, has demanding steps of O(NML), and supports online implementation for large M and N.Source: §5.5.1 · pp. 29–30
• PCA is presented as an important preprocessing technique supporting compression, whitening, visualization, and nonlinear extension through kernelized PCA.Its cited applications include face recognition, interest-rate derivatives portfolios, and neuroscience.Source: §5.5.1 · p. 30

5.5.2. Canonical Correlation Analysis

• CCA extends PCA from one variable to two or more by finding paired, highly cross-correlated linear subspaces through a generalized eigenvector problem.The paper notes a comparison of PCA and CCA for detecting intermittent faults and masking failures in indoor environments.Source: §5.5.2 · p. 30

5.6. Neural Network

• Neural networks fix the number of basis functions, process new data quickly as compact models, but usually require substantial computation for training and readily adapt to regression and classification problems.Source: §5.6 · p. 30

5.6.1. Feed Forward Neural Network

• Feed-forward neural networks use nonlinear unit activities; for regression tasks they use linear activation, whereas multiclass classification uses softmax activation.Source: §5.6.1 · p. 31

5.7. Time Series and Sequential Data

• Unlike sets of data points, many sequential data points are not i.i.d., and sequences can exhibit significant sequential correlation; one related problem is called time-series prediction.Source: §5.7 · pp. 32–33

5.8. Anomaly Detection

• Anomaly detection identifies items or patterns in a dataset that differ from normal behavior; anomalies are often severely underrepresented, sparse, and more diverse than normal behavior.Source: §5.8 · p. 34
• Anomaly detection is used in health monitoring, credit card fraud detection, intrusion detection, ecosystem disturbance detection, and military surveillance.Source: §5.8 · p. 34
• Removing outliers through anomaly detection preprocessing can significantly improve subsequent machine-learning performance, especially for supervised learning tasks.Source: §5.8 · p. 34

5.8.1. One-class Support Vector Machines

• One-class support vector machines (OCSVMs) are a semi-supervised anomaly-detection technique for unlabeled data sets that estimates an input-space subset S whose outside-sample probability is bounded by a fixed value between 0 and 1.Source: §5.8.1 · p. 35
• One-class SVMs can detect anomalies by modeling complex normal patterns in data.Related work also reviews stream-data outlier-detection techniques and their issues.Source: §5.8.1 · p. 35

• Algorithm selection should begin by identifying whether the task is structure discovery, anomaly detection, value prediction, category prediction, or feature extraction.This task-oriented framing supports decisions for smart-data analysis.Source: §6 · p. 35
• For anomaly detection, one-class SVM- and PCA-based methods can train on anomalous and noisy data with high performance, including human-activity monitoring.Source: §6 · p. 36
• Neural networks are suitable for predicting data categories and solving function-approximation problems.Source: §6 · p. 36

7. Research trends and open issues

• IoT data analysis must overcome major problems in data characteristics, applications, and analytics algorithms.These problems are categorized into three types.Source: §7 · p. 37

7.1. IoT Data Characteristics

• High-volume, fast, varied IoT data make preserving quality difficult, while existing solutions cannot accurately handle all characteristics across distributed management and real-time platforms.IoT data are often raw and insufficient for analysis; semantic technologies improve abstraction through annotation but still need to address velocity and volume.Source: §7.1 · p. 37

7.2. IoT Applications

• IoT applications must address privacy and security because collected data may be personal or business-critical, while insecure designs can allow infected device networks to cause a crisis.Relevant security parameters include network security and data encryption.Source: §7.2 · p. 38

7.3. IoT Data Analytics Algorithms

• Analytics algorithms must process diverse IoT Big Data in real time; deep learning can achieve high accuracy with sufficient data and time but is noise-sensitive and difficult to interpret.Semi-supervised algorithms can assist by modeling small labeled datasets with larger unlabeled datasets.Source: §7.3 · p. 38

Conclusions

• Smart City IoT services in energy, mobility, and urban planning can be enhanced and optimized by analyzing the smart data collected from these domains.Smart City is presented as a major IoT application providing services across these areas.Source: §8 · p. 38
• Algorithm selection must account for each application’s device numbers and data types, the specific features of generated data, and the taxonomy of available algorithms.These three considerations are identified as important when applying data-analytic algorithms to IoT and Smart City data.Source: §8 · p. 40
• The review makes choosing a proper algorithm for a particular IoT or Smart City problem easier by organizing analytic algorithms into eight categories.It presents the taxonomy as a central contribution for applying data analysis to smart data.Source: §8 · p. 40
• The article reviews Smart City use cases, describes smart-data characteristics and quality, and relates application specifics to the choice of analytic algorithms.These discussions connect the data and application context to selecting an appropriate analysis method.Source: §8 · p. 40
• The future-trend discussion identifies recent issues and future research paths for smart data analytics.This section frames unresolved issues and directions for continued work in the field.Source: §8 · p. 40