Recommender Systems for the Internet of Things: A Survey

RS proactively recommends items that users may prefer. It has evolved through three main generations from RS for E-commerce, context-and social-aware RSs, and RSs that seek to handle IoT data [102]. Several approaches are used to build RSs; however, the conventional approach comprises three main categories: collaborative filtering, content-based, and hybrid RSs. Collaborative filtering recommends items for a particular user based on the ratings of previous users. Content-based methods recommend items from the same category as the items that the user has targeted before. The hybrid approach combines two or more recommendation methods. There are three major requirements for an effective service recommendation as discussed below.

• •

  Accuracy. This is a critical metric in any recommendation system as it ensures the appositive experience of users. An RS is defined as accurate if it recommends relevant services and a few irrelevant ones, especially when there is a lack of information to support the recommendation.

• •

  Productivity.One of the critical requirements for RSs is productive ability. Productive recommendations are those that are able to produce without the user’s explicit request. Everything can connect to the Internet from any place and for anyone, especially with IoT. Therefore, rich information is available to decide if this situation needs a recommendation or not, or to define which kinds of recommendation techniques are to be employed.

• •

  Diversity.Unlike traditional recommendations, RSs in the IoT environment need to handle heterogeneous and interdependent relationships between entities, data, information, and knowledge. Consequently, it becomes crucial to provide the techniques in order to distinguish between the various kinds of relations and to finally conduct an accurate recommendation

• •

  Newly deployed services. It is critical that newly-deployed services can address the cold start problem, which is considered to be one of the main issues for RSs.

IoT allows internet-enabled physical things to connect, communicate and exchange data. RFID, wireless sensor networks and embedded objects, known as smart things, form a network that bridges the physical and virtual worlds. Smart mobiles, multimedia appliances, toys, and all kinds of other devices can be embedded with sensors to participate in the network. The IoT can be considered a combination of ubiquitous computing, pervasive computing, and mobile computing. The ultimate goal of IoT is to provide a seamlessly integrated platform through which applications involving either physical things or traditional virtual resources can be developed and integrated through the internet. In other words, we can treat physical things as traditional web resources so as to access and interact with them. The IoT relies on the existing internet standards and architectures, and researchers are seeking ways of reusing and adapting current internet standards, such as TCP/IP, HTTP, and Web services, to physical things. However, designing an RS based on IoT is far more complicated than designing traditional RSs because of several drawbacks and limitations that are associated with the use of IoT, as described below.

• •

  Big data management.IoT devices generate massive amounts of data, so an efficient big data management system is needed to deal with various kinds of data and variable velocity. Providing reliability and scalability for this system is important to ensure that it works with no downtime.

• •

  Trust management. This challenge arises when the RS deals with a large distributed sensors network. The system should have the ability to defend against malicious nodes by establishing a technique to detect untrustworthy entities. One technique to ensure this trust and reputation among IoT devices is allowing each IoT device to evaluate the trustworthiness of the others [107, 6].

• •

  Privacy. RSIoTs deal with a huge amount of sensitive data concerning individuals, people, organizations, businesses, and health providers. Authentication, authorization, data encryption and integration, and fine-grained access control are crucial to protect these data [36].

• •

  Security. Key security risks can be grouped into three categories: risks connected with physical components, such as counterfeit attacks, false attacks, information tampering and network damage; communication risks, such as DoS and DDoS attacks; and risks associated with applications, such as information disclosure, authentication, illegal human intervention, and unstable platforms [75]. However, dealing with security issues for RSIoT is more complex than traditional RS because of the heterogeneity of the objects and the large scale of the network.

• •

  Interoperability. Enabling communications among IoT devices that have different standards and protocols can be a critical issue. Networking protocols should be adapted to eliminate the restrictions between constrained and unconstrained entities in the IoT environment [96].

• •

  Quality of services. Quality of services is one of the main requirements of any system in IoT that ensures availability, efficiency, scalability, and adaptability. It is crucial for IoT systems to plan effectively, use resources efficiently, and respond to queries immediately and adaptively in a highly dynamic environment [40].

• •

  Heterogeneity. There might be diverse resources in IoT, so significant distinctions among these resources need to be hidden to provide a consistent presence.

In this section, we highlight several challenges faced by RSIoT, which constitute the main aspects of concern when designing an RSIoT.

• •

  Diverse relations. A common challenge for an RSIoT is that it needs to consider heterogeneous relationships among users, things, data, information, and knowledge due to the poor interoperability between things and data. Therefore, a recommendation technique should be able to conduct accurate recommendations by discovering and leveraging all the different relations.

• •

  Scalability. An RS should be able to perform consistently when the whole IoT-based system scales out, and the related data accumulate. An IoT-based system is subject to scaling out in both hardware and software, as more and more individual things are connected into the IoT and continuously contribute to the explosive increase in the number of things. Along with this process, the amount of data may also increase dramatically. For this reason, the recommendation framework should maintain stable performance in terms of providing a reasonable response time and consistent accuracy.

• •

  Dynamicity. An RSIoT needs to handle three aspects of dynamicity: dynamic discovery of things at the network level, dynamic discovery of user preferences based on their situation, and on-demand real-time recommendation. Besides, the location of things, environmental parameters, and related resources can dynamically change.

We focus on RSIoT. We describe the techniques to build RSIoT and introduce applications of things of interest to recommendations. To analyse the growth of RSs in the IoT environment, this section reviews the major techniques that are exploited to construct RSIoT, including conventional recommender techniques, context awareness, the social IoT technique, multi-agent algorithms, recommendations with a graph database model, recommendations with machine learning and deep learning techniques, and recommendations with reinforcement learning techniques.

The content-based (CB) approach recommends items that are similar to the items previously targeted by the user. Instead of relying on ratings, it uses the existing interest history to predict user interest in the target and match the content of similar profiles with the target content (see Figure 3). Therefore, it may suffer from the cold start problem [93]. The CB approach has been used for recommendations in the IoT.

In [31], the authors proposed two RSs for their AGILE project, which aims to improve users’ health conditions. As a result, two new apps conduct the patients’ healthcare devices and physical activity plans. The CB approach is used to build the recommender engine for a second app. the idea of the app is to collect the patient’s data through medical sensors that can be worn, and a virtual nurse is illustrated as a case to explain this. Based on these data, the app can recommend a suitable activity plan to the patient. In another illustration, in [60], the authors adapted a CB approach to building a recommendation module for their smart restaurant, which aims to provide dish recommendations based on the customers’ tracking history. The Jaccard measure was adapted to measure the similarity between the customers and the dishes to conduct recommendations based on user preferences. The CB approach has also been applied to parking applications. However, based on this technique, the applications will only make recommendations that are based on the similarities among the customers’ parking profile without considering their user preferences. Therefore, accurate recommendations resulted in parking issues. To address this problem, the work in [104] proposed a periodical recommendation that is conducted during three periods: recommendations at the entry gate, recommendations for the nearest parking zone and recommendations for the new zone, when the recommended parking is taken. The system uses three main dimensions as data sources: user profile, user preferences and nearest zone. The distance matrix is adapted to calculate the distance between the user and a zone. To minimize the computation time, the system can remove unrelated zones from the zone list. The system was also tested using a simulation dataset and the result shows it outperforms other systems that focus on user profile or user preferences only.

The hybrid (HB) approach combines two or more approaches to build an RS. For example, by combining the collaborative and content-based methods, the limitations of each can be addressed. The different ways of merging collaborative and CB approaches into a hybrid RS are as follows [17]:

• •

  Building a hybrid RS by implementing content-based and collaborative features separately and mixing their predictions.

• •

  Combining some content-based features into a collaborative approach.

• •

  Combining some collaborative features into a CB approach.

• •

  Building a general unifying model that combines collaborative and content-based characteristics.

Significant research efforts have been devoted to exploiting the hybrid approach to improving recommendations and search performance in IoT. In particular, the framework of the recommendations for Digital Signage (DS) in Urban Space [110] exploited the hybrid approaches to make DS more attractive. The framework has five components: IoT sensing, data pre-processing, recommender engine, DS user interface, and DS data storage. The recommender engine plays the main role in investigating product taxonomy, advertisement taxonomy, condition feature groups, and the DS recommender model of the framework. It can deal with multiple reviewers by switching to various reviewer’s modes. The implemented system’s results were based on a digital signage system deployed in a Taipei shipping mall and showed that both the demographic and context features are crucial factors in providing accurate recommendations. To reduce the latency of recommendations, in [58], a user-space customized RS platform system is proposed that will exploit the mobile edge environment. The conceptual architecture combines several components to build an RS that produces high-quality recommendations. The system produces the recommendations by exploiting user feedback data and sensor data that are closed to the user. This system deals with two user scenarios: (1) if this is the first time the user has dealt with the system, the recommendations are based on similarity with other user’s preferences (collaborative filtering); (2) if this is not the first time the user has engaged with the system, the user’s feedback data in the repository is considered in the recommendations (content-based). In [46, 89], the authors built their RS engine using a hybrid recommendation algorithm which combines CF and CB approaches.

The RS framework is based on the cognitive process. It consists of three main layers. In the first layer (Requirement layer), the goal of the system is described by using a Cognitive Specification Language (CSL), which it then sends to the next layer. The Things layer observes the environment and provides the recommendations list. The main layer is the cognitive process layer, which contains the cognitive recommender engine that is responsible for receiving the sensory data from the Things layer and executing the recommendation algorithm.

This approach recommends items to the user, based on knowledge about the users, items, and their relationships. The Knowledge-based (KB) approach has a functional knowledge base whose performance is based on the identified relationship between a user’s need and a possible recommendation [17]. It does not depend on user rating or gathering information about specific users to provide recommendations. The KB approach addresses several limitations of other approaches, such as the ramp-up (cold start) problem. Ontology is a formal method of representing knowledge that is central to building RSIoT.

Creating and evaluating ontologies are complex and time-consuming processes, but several studies have used them to build RS s in IoT. For example, the authors in [51] proposed a method for generating automatic rules and recommending the best rule, which requires no complicated configuration or extra efforts at the user end. Also, the user has the opportunity to add new rules for newly connected devices. To increase performance, the authors further applied two steps with three ontology models and collected open web data, related to rules by a data pipeline. First, three ontology models were created: (1) Things (knowledge base), which provides all information for things; (2) Context (knowledge base), which provides contextual information about people, environment and things; and (3) Functionality (knowledge base), which links the functionality of context and things. Second, data about the automatic tasks were gathered from the web (web crawler) to be stored in the IFTTT crawled data. Then the collected data from the web were mapped with the data of the three ontology models in order to produce the home rules. The user can add new automatic tasks when a new device is added (sensor) by recommending some tasks to the system. However, adding new automatic tasks might conflict with current tasks. For this reason, the authors defined four potential conflict situations to be solved. In [61], a conceptual framework called the web of objects (WoO) is proposed that provides smart spaces services recommendations by using semantic ontology. It combines three main layers: The virtual objects (VO) layer is used to represent real-world objects, a composite virtual objects (CVO) layer that accumulates VOs and their information, and the services layer, which is responsible for providing recommendations based on the results of the previous layers. Also, the authors in [113] adapted ontology formats for smart space technology to implement an RS for historical tourism. Also, for smart health applications, ontology’s main role is to supervise chronic patients and to provide long-term care [27]. However, a classic ontology cannot define the membership value of risk and uncertain factors precisely; thus, the system provides inefficient results. To address this problem, [5] adapted a fuzzy ontology to build a system which monitors diabetes patients and recommends specific food and drugs. It combines 4 kinds of fuzzy ontology: patient ontology which contains the history of a diabetic patient, sensor ontology, which includes the data from sensors, knowledge and rules-based ontology, which consists of the rules that help to define the patients’ conditions, and drugs and food recommendation ontology to provide recommendation services. The experiment’s results show that the performance of the proposed system outpaces that of classical ontology.

Context-aware computing is well known in computer science research. In [26], context is defined as "any information that can be used to characterize the situation of entities (i.e. whether a person, place, or object) that are considered relevant to the interaction between a user and an application, including the user and the application themselves. Context is typically the location, identity, and state of people, groups, and computational and physical objects".

The application of context-recommender systems in the IoT environment faces several challenges, such as context discovery, security, privacy issues, and sharing context, but some researchers have used context awareness in their frameworks to build RSIoTs. Proactive multi-type context-aware recommender system [92] exploited the user context to define when the recommendations must be pushed to the user, and which kinds of recommendations are to be pushed. The proposed approach is demonstrated in three domains: gas stations, restaurants, and attractions. The authors built the system in two main phases: identifying the situation of the user and the type of recommendations phase, and the item recommender phase. In the first phase, a context-aware management system (CAMS) performs the following tasks: (1) data acquisition, where data about IoT devices, such as sensors, GPS, etc. are collected; (2) data modelling to convert the raw data into understandable form; (3) context reasoning to deduce new knowledge based on the results of the previous stages – the neural network is applied at this stage to generate several scores for the three types of recommendations; and (4) scoring algorithm to define which types of recommendations will be moved to the next phase. In the second phase, a CF approach is used to recommend the items for the user. The proposed framework was adopted as an application, and then tested and evaluated with good results. The authors in [48] proposed a Context-as-a-Service (CoaaS) recommender system that uses an IoT context service to enable applications to provide and consume contextual information seamlessly without the need for manual integration among IoT devices. There two core components are context service description language (CSDL) and context service matchmaking (CSM). CSDL enables developers to describe their services regarding semantic signature and contextual behaviour in standard language. CSM is responsible for matching between context request and context services. The CoaaS framework was tested by adapting it to produce smart car park recommendations via an application on a smartphone; this is explained in detail in section 4.2. The EW4 model [125] exploited contextual information and the mobility of the user’s tweets to provide accurate recommendations. The framework combines both offline and online components. The author has proposed an algorithm using a generative process to model the offline components, which is based on four aspects: who (User), where (Geo), when (Time) and what (Words). Several applications based on this model could be built, such as location recommendations, use prediction and location prediction. The EW4 model’s performance for some tasks was evaluated using two datasets.

The authors in [121] proposed an RS that could provide accurate recommendations by considering contextualization with IoT data. They defined two operations for contextualization: contextual filter and contextual aggregation. In the contextual filter step, data on IoT devices and services are filtered, based on the current context. Contextual aggregation involves combining the filtered data, based on both contextual similarities and relevance. The authors adapted the framework to design a smart parking application. Zhou et al. [128] proposed a model which could improve recommendations accuracy by exploiting context-awareness. The model consists of three main components: firstly, the server, which is responsible for providing services recommendations to users; secondly, the user, who provides the context to the server in order to get recommendations; and thirdly, the services provided by the server. The authors designed a Hierarchical Social Contextual Tree (HSCT) algorithm that uses three sources to provide accurate recommendations: service usage frequency, contextual bandit feedback, and social intimacy. The experiment’s result showed that the system outperforms other algorithms, especially in two dimensions, namely scalability, and the cold start problem. Also, in [56] the context is adapted as a source to build a recommender engine for recipe recommendations in a smart kitchen. The recommender engine consists of four main components: food item taxonomy, which represents the food items and their attributes in the refrigerator and cupboard; recipe taxonomy, which includes all the recipes and their ingredients; conditional groups, which combines environment and context features; and the recommender model, which utilizes the historical data to conduct recommendations. Based on the previous components, the authors described a motivation scenario to explain the importance of the proposed system. Also, context awareness was employed in building an RS for smart healthcare [19]. In the same vein, the authors [50] proposed an RS that provides social recommendations for cultural heritage. The architecture consists of three main parts: socializing, social-based recommendations and context-based recommendations. The system can deal with the sparsity issue, where the recommendations are conducted, even when the new user uses the system.

The concept of social networking is applied to IoT to create a social relationship among things. The authors in [49] introduced the concept of socialization between things, which focuses on finding solutions to allow smart wireless devices to establish temporary relationships. The main idea behind SIoT is that social objects can select, discover and compose services. Five main kinds of relationships are produced, based on the object’s activities and features, which are explained in detail in [13].These relationships depend on the object’s activities and the features that are to be created and updated. Moreover, they require some modules to utilize these relationships, such as a service discovery module and relationship management module [90] SIoT enables IoT applications to exploit IoT services efficiently. Hence the RS can exploit the data that are gathered from various applications. For example, the authors in [90] proposed a framework that could exploit the SIoT network to produce services recommendations. The SIoT network includes the profiles of all the IoT applications. The data profiles can be used to provide service recommendations or could be exploited by other IoT applications to search for similar conditions that have been solved in the past. The proposed framework has three main layers: the SIoT perception layer, which collects the data of all IoT devices and their relationships, which are extracted by using SIoT technique; the network layer, which connects the previous layer with the upper layer by using several communications protocols; and the interoperability layer, which enables all IoT applications to use the IoT data. The authors provide an example of how their framework could be adapted for various applications. This would be done by exploiting the availability of social correlations among things and between users and things to provide services recommendations.

However, facilitating access to quality services and trustworthy devices in a SIoT environment has become a critical issue. To address this problem, a scheme of access service recommendation [22] is proposed that provides recommendations about trustworthy nodes. Trustworthiness is evaluated, based on three parameters: a feedback-based reputation system, social relationship and an energy awareness mechanism. Recommendations are subject to an unreliable wireless environment and limited battery capacity in smartphone applications. In [86], the mobile IoT is exploited to build an RS for services and social partners. The RS has two main modules: a recommendation module and a physical layer module. The recommendation module combines three components: recommendation-database, service/interest-matching, and service/connection time-prediction. The Recommendation database performs three main functions: (1) it collects all the information on users and service providers, including user preferences, features, and services; (2) it creates a social network page based on this information; and (3) records the historical information. The service/interest-matching (S/I-M) matches the services of the providers with user interest using different recommendation and data mining techniques. The service/connection time-prediction (S/CT-P) is used to detect a suitable time to create a D2D link, taking the quality of the wireless channel, battery capacity and the mobility of the wireless device into account. The results of (S/I-M) and (S/CT-P) are evaluated to provide recommendations by updating the social network pages. The physical layer module is responsible for managing all the resources, such as the physical link with the partner, power consumption and codeword rate, as well as providing communication recommendations to improve the performance. The authors give a detailed example to explain the role of this module.

A multi-agent system (MAS) has two main goals: to provide rules to construct a complex system and provide techniques to coordinate each agent’s behaviour. In [105], several reasons for using a multi-agent system are presented; some of these are summarized in Table 1. These considerations have led some studies to focus on the use of MAS to build recommendations systems in IoT.

In [35], a multi-agent algorithm is exploited to improve the speed of the recommendations, with each thing described by thing descriptors (i.e. bit vector). Things with similar features will then be brought together and managed only by neighbour’s cyber agents. Note that all cyber agents are connected by hops. The probability function is used to decide if the thing descriptor will be moved towards a linked cyber agent. The cyber agent evaluates the thing descriptor if it arrives from another linked cyber agent. The algorithm is designed to make it both intuitive and straightforward to produce a recommendable thing. When a query is established, a cyber agent matches thing descriptors and then forwards the query to a neighbour cyber agent with a maximum value of similarity. This process continues with other cyber agents until the node agent has a maximum similarity result, compared with the current agent, and the query process then finishes. After that, the query is forwarded to the asking cyber agent to produce the user’s recommendations. The algorithm was tested and evaluated, and the results showed that the recommendations were produced faster.

The authors in [28] exploited the IoT data to build a Mobility Recommender System (MRS) for parking, especially in urban areas, by incorporating features such as an ideal door-to-door route. The system has two main parts: the data sources, which are divided into static and dynamic data as input to the system; and the recommender system, which is responsible for dealing with these data. A MAS is exploited here to design a mobility query engine (MQE) that can distribute IoT data based on location, type, and complexity. The goal of using a MAS is to recommend parking spaces based on the consideration of different input data. For example, when the door-to-door route requires the user to use public transport, one agent who is aware of city policy may provide sets of recommendations to the route calculation planner (RCP) that take account the city’s regulations. Another agent who knows about the user’s parking preferences may also provide several recommendations to the RCP. Finally, the MRS is responsible for collecting and aggregating the queries into a ranked list of recommendations that will be considered in subsequent recommendations, where a utility RS is used to calculate user preferences.

Twardowski et al. [111] built an RS that uses data from mobile devices and other IoT devices to provide personalized recommendations in real-time. The system consists of two main parts: the Big Data Server Side and Mobile Devices Side. The Big Data Server Side is responsible for the collecting and processing of mobile devices by using a multi-agent system and the Mobile Devices Side, which combines the numbers of mobile devices, such as smartphone, sensors, and beacons. The main features of this system are that it uses edge services to reduce direct communication between the mobile devices and the server, deals with the sparse data by using matrix and tensor factorization techniques and lastly, addresses the cold start problem by using the current context information and calculating recommendations in real-time.

The authors in [54] proposed an RS based on the multiagent technique to optimize electricity consumption and save costs in a smart home. The system consists of three main modules with specific agents: a device module to collect data from each appliance; a crawler module, which contains the information about electricity prices, and a recommendation module that uses data from the other two modules to suggest recommendations. There is also a control agent which manages the interaction between the modules. The recommendations part is built, based on two conventional recommendation techniques, which are KB and UB. However, the system has only been evaluated in a theoretical case study using the UKDALE dataset.

A graphical database model is defined to represent a structural schema. Models and diagrams, such as entity relationship diagrams, are tools that designers/architects can use to test the different data structure relationships or display validity constraints graphically before implementation [9]. Graph data models are used in applications that consider the importance of the data and the relations among data at the same level. There are several benefits of using graphs as a modelling tool [43]:

• •

  It allows the user to show all the information about an object in a single node, while the related information is referred to by arcs.

• •

  Queries can be referred directly to a graph structure.

• •

  Graph databases can provide efficient graph algorithms to investigate specific operations and efficient graph storage structures.

The authors in [84] exploited a Neo4j graph database to address one of the main challenges in IoT, namely, big data management. The sensory data from a smart city, which are used to produce service recommendations, are stored in a graph database. The proposed architecture has three main components: node-red, Neo4j graph database, and the recommender. Node-Red combines two parts: (1) data source flow, which is responsible for processing all data before they are transferred to the next part; (2) heating manager flow, which uses the received data to control the heating schedule for each house, as well as enabling users to interact with the system. The data stored in the Neo4j graph database define the relations among them in order to answer the user’s questions. Finally, the RS exploits not only the predefined existing data (historical data) but also the real-time data stored in Neo4j. CF is suggested as an approach to defining the similarity among the users by using similarity metrics (Euclidean and cosine). The main feature of the architecture is that each component is independent, which enables the extension and scalability of the system’s features.

Also, the author in [81] exploited graph techniques to build an RS which provides IoT services recommendations to the user based on their own IoT devices. The system combines three main steps: IoT services modelling, IoT service catalogue (classification algorithm) and service matching algorithms (Recommendations). The IoT services model their functions by using Typed Attributed Graphs to describe all the objects and their relations. IoT services are catalogued by using an algorithm that entails the following three steps: (1) Defining physical interfaces and their location in spaces in order to build the space’s profile; (2) determining how many smart spaces share the same profile in order to distinguish the relations between the objects; and (3), a signature is defined for the list of profiles, based on the result of step 2, with the spaces that share each profile. Service matching algorithms are used to correlate between the user request and signature of the service classes. The authors discuss two different scenarios where the system can use these algorithms. One is when the user is an expert, which means he can add a new service to the catalogue, and the other is when the system can make a service recommendation based on the requests of a typical user.

Machine Learning (ML) can be generally considered a sub-field of Artificial Intelligence (AI). It entails the use of computers to simulate human learning and can collect and use real-world knowledge to enhance performance. The field of ML has various definitions that are informed by its application across different fields of knowledge. A general and widely applicable description is, however, apposite: "A computer program is said to learn from experience (E) with respect to some class of tasks (T) and performance measure (P), if its performance at tasks in T, as measured by P, improves with experience E" [77]. ML algorithms can be divided into three (supervised, unsupervised, and semi-supervised), based on the nature of the data involved, or into two (transductive and inductive learning), based on the concept learning perspective [30].

ML algorithms can be utilized to optimize the ability of traditional RSs to provide accurate recommendations to the user. In [97], the authors designed the Optimal State based Recommender (OSR) System by exploiting some machine learning algorithms, including Distributed Kalman Filters, Distributed Mini-Batch SGD (Stochastic Gradient Descent), Distributed Alternating Least Square (ALS) based classifier, and some ML platforms. It shifted conventional recommendations, based on user/item preferences only, into accurate recommendations that deal with real-time data. Some conventional recommendation techniques cannot be directly adapted for use in the design of RSs. Accordingly, some studies have focused on improving some of these techniques or functions. For example, the authors in [41] proposed a framework to build an e-commerce RS that exploited the multi-sources of information to produce accurate recommendations. The framework has three main components: data sources, recommendation evidence weight, and fusion decision. The data source component uses microformats to provide a unified representation and platform for information about the mobile user. Recommendation evidence weight uses an improved radial basis function (RBF) neural network to define the weights of recommendations. With the improved function, the weight of each piece of evidence is easy to evaluate. In the final component, the Dempster-Shafer (DS) evidence theory has also been improved to fuse information and power spectrum to provide accurate recommendations. The framework was adapted for the exploitation of multi-source information that assists women with online clothes shopping. The results showed that exploiting multi-source information had a significant impact. For example, the probability of recommending shoes increased from only 12% with the traditional method to 85% with the proposed method.

In [12], the authors exploited an ML classifier to build a recommendation engine that provides personalized wearable technologies recommendations for proactive monitoring. This approach consists of three main models: (1) the classifier model, which is responsible for predicting at-risk diseases and suggesting some measurements for each person; (2) the optimization model, which is based on the measurement results and gives suitable wearable technology based on the previous measurements; (3) the Monitoring Framework, which is responsible for monitoring the readings from all the recommended devices and sending these readings to the classifier model to update its measurements.

Moreover, the K-Means algorithm [8] is adapted in the UTrave RS application, which clusters user profiles to recommend points of interest for the user. The building of the UTrave application was based on Universal Profiling and Recommendation (UPR) that combine two main steps, which are creating the user profiles and filtering. It is evaluated by two steps: (1) a simulated user to test the accuracy of the clustering user profiles and (2) real users to verify the whole system. In another experiment, Rasch et al. [85] adapted unsupervised learning to build an RS for smart homes. The system learns user patterns and conducts recommendations based on user contexts. The authors divided the system into two main phases: a training phase, with a sequence of sensors events as input; and a recommendations phase where the input is the current user context. The proposed RS is evaluated by using two publicly available datasets that are considered as a single person. Meanwhile, a decision tree [124] is used to build a system which provides lifecare recommendations. The system combines the two main parts: a peer-to-peer dataset and adaptive decision feedback. In the first part, the collection of the data is based on the peer-to-peer networking environment, Open API and biosensors, after which a decision tree is used to defined and classify the data. The adaptive decision feedback part plays an important role in providing flexible results of the recommendations and facilitating the real-time lifecare services. In a more comprehensive study, Valtolina et al. [112] considered both the decision tree algorithm and social network to propose a multi-level RS. Others [87] adapted the Analytic Hierarchy Process (AHP) to build an RS for car parking recommendations. Simulation parking is used to test the system, where the result outperforms other modules using the same simulation. In another study, [14] ML algorithms were adapted to design an RS for music recommendations that was more accurate than traditional ones. The system used sensory data from wearable devices to detect the emotions of each user and then used this data as a source to conduct accurate recommendations. Three kinds of ML are used to classify the collected data into the target emotion, which are: random forest, kNN and decision tree.

Deep Learning (DL) is generally considered to be an extension of ML. It applies two main steps: adding multiple layers, which increase the depth of the model, and transforming the data by using diverse functions representations and abstractions of multiple levels [94]. DL includes several components, such as activation function, convolutions, and pooling. It relies on different architectural paradigms (i.e. Multilayer Perceptron (MLP), Autoencoder (AE), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and Restricted Boltzmann Machine (RBM)). The main advantage that enables DL to solve complex problems quickly and effectively is its ability to learn features. However, its limitations are a longer training time and the need for large datasets to describe the target problem. DL has been used in numerous applications that deal with continuous data, such as weather data. In [123], the DL technique is utilized to build an intelligent system for a fitness club. The main goal of the system is to recommend suitable exercise or detect the users’ fitness actions by monitoring the user. The authors exploited a 3D CNN to identify fitness actions which can extract time and space features from continuous frames. The public KTH dataset is used to test action recognition by using the 3D CNN, where the test accuracy reached 0.8865. Inaccurate recommendations have a negative impact on the user. To address this issue, a deep neural MLP has been adapted to make system recommendations more effective for a smart museum [47].

With the recent tremendous approaches to RS, reinforcement learning (RL) has received increasing attention. It meets two majority requirements for recommendations: 1) treating the interaction between the user and the agent as the main procedure for a recommendation; 2) learning the optimal policy to increase the cumulative reward without any predefined instructions. Some studies have started to adapt RL for their RS for the IoT.

Massimo et al. [71, 72] have proposed inverse reinforcement learning to model user behaviour as a way of improving the quality of recommendations. This technique learns about user decision behaviour by observing the user’s actions and groups observations to learn from a small number of samples. After the user behaviour is learned by using a linear function, known as a reward function, the recommendations are produced by matching the user’s predicted choices with a large group of observations of both the user and groups of similar users. The tourism domain applies the system in both outdoor and indoor environments. The initial results showed that IRL could learn user preferences, even when the datasets were small and noisy. Authors in [44] used a reinforcement learning algorithm (Upper Confidence Bounds algorithm (LinUCB)) to design a framework that provides context-aware recommendations in a smart city. Also, [83], the authors proposed a system, SML, to monitor the daily indoor activities of seniors with mild cognitive impairment. This aims to recommend the correct sequence of tasks for each activity so that the user can reach his goal.

Over the last few years, m-health has attracted the attention of researchers who have seen it as a potential way of combining IoT with RSs to provide long-term healthcare. As a consequence, there are now numerous applications. In [111], the data from mobile devices and other IoT devices surrounding them were exploited to provide personalized recommendations for dietary/fitness. The idea of this mobile application is that a higher-level context is calculated after the data is collected from smartphones and other sensors. Then, accurate recommendations are produced to match the user’s needs. For example, when the user has finished an exercise class, the system recommends a suitable restaurant that could provide a healthy meal. Yong et al. [123] designed an intelligent application to guide users in gyms (shown in figure 4). This application includes a part responsible for providing course reminder recommendations that are based on the users’ fitness. Another study [124] designed a lifecare recommendation mobile service which improves the quality of life, as shown in figure 5.

Car park recommendation is considered to be one of the crucial services in smart cities [53]. There is rapid growth in the design of applications for parking recommendations that exploit IoT data. In [48], one such application has been designed that exploits context services data owned by several providers in order to produce accurate recommendations for users. The basic idea of this application is that four context services are used as sources to produce accurate recommendations, which are described using CSDL. Also, an OBD II device is connected to the application using Bluetooth that provides some of the required sensory data, such as speed and fuel level. As can be seen in Figure 6, the application considers the context services before providing recommendations. For example, if there is bad weather, it considers the short walking distance in the car park recommendations (figure 6 (a)), but if the day is sunny, the car park recommendations will consider low-cost (figure 6 (b)). The last part of figure 6 refers to the application running on a mobile phone and OBD II device, which is exploited as a data source in this application.

In [121], which is another smart parking application focused on collecting contextualized data for each driver, IoT data and car park services data (see Figure 7), contextualization methodology is applied to the collected data to provide parking recommendations. The experiment’s results showed that the contextualization produced a three-fold reduction in query response time compared with other approaches.

Another study [28] focused on using a mobility RS where each agent was responsible for providing a list of recommendations. For example, one agent provided only direct car parking spaces based on the destination. In contrast, another agent had city policy regulations so the recommendations could be different from the previous one. The main feature of this application is that the user has several recommendations that consider his preferences from many perspectives. However, parking applications still face two main issues: using the nearest parking space, instead of parking that is recommended, and the recommended parking space being stolen. These two problems cause conflict parking. To address these problems, in [104], an application for parking conflict reduction was proposed that uses a periodical recommendation. Parking spaces that are available at different periods during the day are recommended to ensure the accuracy.

Some recent tourism literature indicates that providing tourists with a unique and differentiated service has a positive impact on marketing. When tourists explore an area, they require high-quality services that will lead to memorable experiences. However, the dramatic increase in options available to them means tourists have difficulty making decisions. E-tourism RSIoTs address this issue by providing accurate recommendations. There are several tourism RSs. For example, authors [20] designed a smart tourism application by exploiting two of the main sources to provide real-time recommendations: user mobility pattern and points of interest. The application sends recommendations that are suitable for each user and based on his or her location. Also, in [72], a tourism application is designed which recommends points of interest to visitors. User behaviour is used as a source to reduce potentially false information that could affect the quality of the recommendations made by tourism applications. The idea of this application is that there are several media in different spaces that the visitor can watch by using the NFC tag. Each user has a specific ID, and the media are shown when the NFC tag is passed through the NFC reader. By monitoring and filtering all the interactions between the media and the visitors, the system can recommend points to visit that match the user preferences. The authors in [50] proposed an RS application which provides a list of artworks recommendations that are based on calculating the social affinity between the user and user experience.

When a personal RS makes recommendations that resolve complex problems, the bridge between the user and objects is built. The design of this kind of system has become important, especially with IoT data. Accurate recommendations can be provided by exploiting massive amounts of IoT data and knowing when and what kinds of recommendation should be pushed. For example, an application (ProRec) [92] has been designed that supports the multitype context for recommendations. Three types are used: restaurants, gas stations, and attractions. When the user downloads the application, s/he can control it by using the options setting. The recommendations will be pushed to the user, based on how close they are to the restaurants, rating, and other options (such as whether the restaurants are open at lunchtime). Figure 8 displays different screenshots of the application. Acceptance or rejection of these recommendations will affect the results. For example, if the user does not accept the pushed recommendations, the rating of these recommendations will be reduced.

There are other applications that provide recommendations by focusing on exploiting the data of other applications. For example, the authors in [90] used a sample application scenario to show how social IoT provides benefits for IoT applications by using other applications’ data. Several IoT applications are used: skiing, arranging meetings with friends, vehicle-to-vehicle communications, traffic monitoring and demonstrations of wearable (see Figure 9). An interoperability layer is responsible for enabling the interoperability between these applications so that the IoT RS can exploit the data to produce recommendations to the user. One of the most important features in this scenario is that IoT data can be exploited by different applications at the same time and it can investigate scalability when large amounts of things are added to the network. However, some implementation challenges are identified in the paper.

The work in [55] used the data from an e-mail survey and traffic analysis to provide personalized recommendations to the user. The idea of this application is that the user has a chance to define their choices of interest, such as emotion, location, weather and time. The system then monitors and analyses the smart devices’ traffic to predict services recommendations which match their preferences. Another application [21] exploited IoT data, such as weather data, to provide several recommendations that are based on the weather. It uses sensory data to collect weather observation sequences and analyses them. Then it correlates current observations with previous ones to extract the trends that are predominant during this period and provides recommendations to the user.

Social recommender systems not only conduct item recommendations based on social network data but also help to create relationships between users based on their preferences [67]. Most applications for IoT are focused on exploiting social data to correctly define user preferences. For example, Amoretti et al. [8] built a UTrave application that recommends points of interest for visitors (shown in figure 10). This application consists of a mobile application side and a server side. The former collects personal data by using applications installed on the client’s mobile devices; the latter leverages the collected data. The main feature of the application is that it could deal with the cold start problem – when the user sends a request recommendation for a new POI to the server but it is not available, so the server starts to use the online data sources of the user (Demographic data) to recommend the nearest POIs from the user. Yuan et al. [125]exploited some of the IoT data, such as tweet data (time, user, words, Geo), to provide accurate recommendations.

In the last few years, smart home applications have become increasingly popular; consequently, RS applications for smart homes have become available that help the user to choose the option that matches her/his preferences. For example, a heating schedule management application [84] is designed that exploits open data from the city. These are gathered by using the IoT infrastructure to provide heating recommendations for a specific user. The application considers several factors while compiling the list of recommendations. The user, for instance, can choose an application that automatically turns the heating on during a specific period. For example, a resident with a flat of 60 $m^{2}$ in a specific location might want to apply economic heating for the next ten hours. Another application [117] has exploited three kinds of correlations to produce accurate home recommendations: the relationships between users, the correlations among things and the interaction between users and things. Also in [32], the authors exploited conventional recommendation approaches for AGILE projects. These help the user to choose several kinds of recommendations; for example, recommending a suitable application for an installed device, suggesting some devices that may be required for some applications, or recommending some specific protocols which match devices. A recommender system [73] was designed that provides indoor comfort products recommendations for users. The system is divided into two main parts: the first is the data collecting part, achieved by using a network of sensors and extracting user preferences, while the second part involves using previous information to make product recommendations. Figure 11 shows a screenshot of the web page of the system, which provides the product recommendation at the top of the screen and the evaluation and delete button at the bottom.

In recent years, online shopping RSs have been developed that help customers by providing various guidelines. Most of these systems use conventional recommender approaches to providing recommendations for large websites, such as Amazon and eBay. However, with IoT, the recommendations could be made more attractive by exploiting IoT data as sources. A shopping centre application [89] has been designed that recommends healthy products to the customer by exploiting sensory data. The customer can also pay automatically using a digital wallet. An RS [110] is proposed which makes DS more interactive. The system exploits the interaction between the user and the DS to make the recommendations.