Source-linked AI summary
Machine learning for Internet of Things data analysis: A survey
Mohammad Saeid Mahdavinejad, Mohammadreza Rezvan, Mohammadamin Barekatain, Peyman Adibi, Payam Barnaghi, Amit P. Sheth
TL;DR
IoT produces large, heterogeneous, time- and location-dependent data whose intelligent analysis is important for smart applications. This survey reviews machine learning methods for IoT, develops an algorithm taxonomy, and discusses their application and challenges; it concludes that matching algorithms to data characteristics is central, while future research issues remain.
Problem
IoT data has substantial scale, heterogeneous characteristics, and application-specific analytical needs, requiring suitable methods for extracting useful knowledge.
Method
The paper surveys machine learning approaches for IoT and Smart City data, organizes them into an algorithm taxonomy, and discusses clustering applications and algorithm selection.
Results
The survey concludes that algorithms should handle IoT data characteristics and that selecting a proper algorithm for a particular problem is important.
Takeaways & Limitations
The taxonomy makes choosing a suitable analytic algorithm for a specific IoT or Smart City problem easier.
Takeaways & Limitations
The paper identifies recent issues and future research directions for smart data analytics as an open research area.
Abstract
from arXiv · showhide
Rapid developments in hardware, software, and communication technologies have allowed the emergence of Internet-connected sensory devices that provide observation and data measurement from the physical world. By 2020, it is estimated that the total number of Internet-connected devices being used will be between 25 and 50 billion. As the numbers grow and technologies become more mature, the volume of data published will increase. Internet-connected devices technology, referred to as Internet of Things (IoT), continues to extend the current Internet by providing connectivity and interaction between the physical and cyber worlds. In addition to increased volume, the IoT generates Big Data characterized by velocity in terms of time and location dependency, with a variety of multiple modalities and varying data quality. Intelligent processing and analysis of this Big Data is the key to developing smart IoT applications. This article assesses the different machine learning methods that deal with the challenges in IoT data by considering smart cities as the main use case. The key contribution of this study is presentation of a taxonomy of machine learning algorithms explaining how different techniques are applied to the data in order to extract higher level information. The potential and challenges of machine learning for IoT data analytics will also be discussed. A use case of applying Support Vector Machine (SVM) on Aarhus Smart City traffic data is presented for a more detailed exploration.
1. Introduction
IoT generates expanding Big Data that requires intelligent analysis to extract knowledge and support smarter services. The paper frames machine learning and data analytics as ways to match algorithms to IoT data characteristics and address scale, velocity, variety, and quality challenges.
- IoT connects physical objects through wired and wireless communications, sensors, actuators, and Internet connectivity.
- IoT data analysis accesses raw data from different resources and analyzes it to extract knowledge for intelligent services and improved framework performance.
- IoT data analytics must account for volume, variety, velocity, differing data quality, and real-time generation from many sources.
- Selecting algorithms and data models that fit data characteristics is important for pattern recognition and better IoT data analysis.
- The paper asks how machine learning algorithms can be applied to IoT smart data and frames Smart Data as transforming raw data into useful information.
data?
The paper asks what taxonomy of machine learning algorithms can be used for IoT smart data.
- The paper asks what taxonomy of machine learning algorithms can be applied to IoT smart data.
be adopted in IoT?
The survey organizes IoT machine-learning analysis around data characteristics, algorithm selection, Smart City applications, and implementation, with a taxonomy of eight algorithm groups and an Aarhus traffic-data SVM use case.
- The survey asks which machine-learning algorithms are appropriate for processing IoT data and handling its characteristics.
- It reviews IoT applications, data characteristics, machine-learning algorithms, and related issues as the survey’s main analytical components.
- About 70 articles are reviewed, revealing eight major groups of algorithms applicable to IoT data.
- The survey also discusses cognitive algorithms as methods for extracting and interpreting patterns from IoT smart data.
- Smart City is selected as the primary IoT use case because it is prominent in reviewed studies, encompasses many IoT use cases, and has accessible open datasets.
- The survey provides technical and simple implementations, including an SVM applied to Aarhus City smart-traffic data to predict traffic hours during one day.
2. Literature Review
The literature review frames IoT as a challenging computer-science area whose researchers must prepare and process data effectively.
- IoT is described as a challenging area in computer science because it represents a new concept for the Internet and smart data.
- Preparing and processing data are identified as important challenges for IoT researchers.
3. Internet of Things
IoT connects devices that collect, analyze, and exchange environmental data, creating processing challenges from scale, communication, and heterogeneous workloads. Computing frameworks distribute processing across cloud, fog, and edge locations, with migration toward distributed processing reducing several resource burdens.
- Internet of Things: IoT comprises connected devices that transfer data automatically to optimize performance and includes sensing, processing networks, analysis, and monitoring.
- Communication protocols: IoT communication is organized into device-to-device, device-to-server, and server-to-server components.
- Data processing: Preparing and processing IoT data is critical, requiring approaches such as edge analytics, stream analysis, database analysis, fog, and cloud processing.
- Computing frameworks: Fog and edge computing move processing toward the network edge, where data can be initially processed, filtered, cleaned, and stored locally.
- Computing frameworks: Cloud-centered processing can have high latency, high load balancing, and increased CPU usage when handling high-volume IoT Big Data.
- Distributed computing: Migrating from cloud to fog and distributed computing is associated with lower network loading, faster processing, lower CPU usage, lower energy consumption, and higher data-volume processing.
4. Smart City
Smart City is presented as a central IoT use case spanning services such as energy, mobility, public transportation, environment monitoring, and urban planning. Its data are continuous, dynamic, heterogeneous, and variable in quality, requiring analysis for real-time and long-term decisions.
- Smart City applications: IoT enables Smart Cities to collect environmental data, use live status reports, and operate smart monitoring systems for more intelligent responses.
- Smart City motivation: Smart City applications aim to improve quality of life and services including traffic, water, energy, health, safety, transportation, and resource management.
- Smart Energy: Smart Energy includes time-series analysis, leak monitoring, smart grids, renewable resources, and measures intended to reduce power consumption.
- Smart mobility: Smart mobility covers autonomous vehicles, traffic control, and public transportation using sensor, location, routing, and historical data.
- Urban planning: Urban planning uses Smart City data analysis to predict future crowding and identify solutions for potential problems.
- Data characteristics: Smart City data are continuously generated, dynamic across locations and times, and collected from heterogeneous sources with different quality levels.
- Data characteristics: Different sensor update rates and heterogeneous integration can cause important information loss, making data integration challenging.
- Quality of Information: Improving Quality of Information requires extracting higher-level abstractions and providing actionable information, supported by trustworthy-source selection, data combination, and denser sampling.
5. Taxonomy of machine learning algorithms
Machine learning is presented as a form of artificial intelligence that enables machines to learn without explicit programming. The section frames smart-city data characteristics as part of the taxonomy context.
- Machine learning is a subfield of computer science and a type of artificial intelligence.
- It provides machines with the ability to learn without explicit programming.
- The taxonomy is situated around the characteristics of smart data in smart cities.
5.1. Classification
The classification section surveys frequently used machine-learning approaches for smart data, including instance-based, probabilistic, margin-based, regression, and tree-based methods. It also describes their operating principles, scalability, flexibility, and limitations.
- K-Nearest Neighbors: KNN classifies an unseen point using the labels of its K closest training points, selected through a distance metric.The discussed metrics include Euclidean distance, L∞ norm, angle, Mahalanobis distance, and Hamming distance.
- K-Nearest Neighbors: KNN requires storing the entire training set, which makes it unable to scale large data sets.Tree-based search and online variants are mentioned as ways to address this issue.
- Naive Bayes: Naive Bayes classifies inputs with Bayes’ theorem under a naive independence assumption between features given the class.The classifiers estimate the class prior and feature-conditional probabilities.
- Support Vector Machine: SVMs find separating hyperplanes using margin constraints and slack variables, with parameter C controlling the penalty for violations.Kernel methods extend SVMs to nonlinear classification, while support vectors reduce the points used for prediction.
- Support Vector Machine: SVMs handle high-dimensional data and support linear, nonlinear, binary, and multiclass classification, but do not directly provide probability estimates.Their solved-model parameters are also difficult to interpret.
- Regression and trees: Regression approaches include linear regression, LMS, SVR, and tree-based regression, while CART partitions the input space into axis-aligned regions.LMS is described as fast, scalable, and capable of online parameter learning.
5.4.2. Density-Based Spatial Clustering of Applications with Noise
DBSCAN clusters unlabeled IoT data by density, identifying dense regions as clusters and low-density regions as outliers. It supports arbitrary cluster shapes but is sensitive to density variation and distance-metric choice.
- DBSCAN determines clusters from data-point density rather than requiring the number of clusters in advance.This differs from K-means, which requires a preset cluster count.
- Dense points with many close neighbors form clusters, while points in low-density regions are treated as outliers.
- Large differences in density can produce poor clusters, and performance is highly sensitive to the distance metric used to define density.
- DBSCAN can detect clusters with arbitrary shapes, including spherical, elongated, and linear forms.
6. Discussion on taxonomy of machine learning algorithms
The taxonomy matches IoT analytics tasks to machine-learning algorithms according to data structure, anomaly status, sequence, category, and feature-extraction needs. It highlights algorithms used across smart-city and smart-data applications.
- Algorithm selection is organized around tasks including structure discovery, anomaly detection, prediction, classification, and feature extraction.
- K-means and DBSCAN are discussed as clustering tools for discovering structure in unlabeled data and handling smart-city or smart-home data.
- One-class SVM and PCA-based methods are applied to detect unusual or anomalous smart-data points.
- Linear regression and SVR are frequently applied to prediction and classification of sequenced data, while linear regression also supports real-time prediction.
- Neural networks are presented as suitable for category prediction and function approximation, including multiclass prediction for smart data.
- PCA and CCA are used for feature extraction, with CCA additionally modeling correlations between two data categories.
7. Research trends and open issues
IoT analytics must address data-quality, scale, velocity, heterogeneity, privacy, and security challenges. The paper also identifies interpretability and robustness limitations in current learning approaches.
- IoT data analysis is constrained by high volume, fast velocity, varied sources, and difficult data-quality preservation.
- Existing solutions do not accurately handle all data characteristics because Big Data management and real-time processing are distributed.
- Semantic technologies improve IoT-data abstraction through annotation but require further work to handle velocity and volume.
- IoT analytics must address privacy risks from personal or business data and security requirements such as network security and encryption.
- Deep learning can achieve high accuracy with enough data and time, but noisy smart data can influence it and neural-network results may lack interpretation.
8. Conclusions
The conclusion frames IoT analytics as a data-driven basis for services such as energy, mobility, and urban planning. It emphasizes matching algorithms to application-specific data characteristics through a taxonomy and reviewed smart-city use cases.
- Analyzing smart data can enhance and optimize IoT services in energy, mobility, and urban planning.
- Choosing an appropriate algorithm for a specific IoT or smart-city application is identified as an important issue.
- The article reviews IoT data-analytics studies and considers application-specific data characteristics when discussing algorithm choice.
- Its taxonomy describes eight categories of analytic algorithms and is intended to make algorithm choice for particular problems easier.
- The paper discusses recent issues and future research directions for smart-data analytics.