Data Summarization in Internet of Things

With recent advances in the field of Internet of Things (IoT), the quantity of data being generated by various sensors and Internet users has increased dramatically which in turn has skyrocketed the need for efficient data compression methods. Data summarization is an efficient and effective technique for data compression that can generate a brief and succinct summary from typically larger quantities of data in an intelligent and highly useful manner, which can be done at various levels of abstraction. The impact of using such a technique in large IoT networks can be significantly advantageous in terms of reduction in the processing time, overall computation, data storage-transmission requirements, energy consumption, and possible workload on IoT users. In this work, a review of existing methods for data summarization techniques at various levels of abstraction of typical IoT networks are discussed. The levels of categorization that are considered are Low-level and High-level. Under each abstraction level, various techniques are further classified while briefly describing their essential characters.


Introduction
There has been a paradigm shift in this age of information, with the advent of Internet of Things (IoT). The impact of IoT can be matched to that of major technological advances such as the industrial revolutions. IoT refers to a highly intertwined system of objects that are connected via Internet, which can transfer useful information with minimal or no human intervention. Along with the numerous potential opportunities IoT offers, there exist many challenges that need to be addressed.
With the increasing growth of the family of users, sensors, and other connected objects, the amount of data being generated is large. An IBM study estimated that more than 2.5 quintillion bytes of data is generated globally everyday [1]. Therefore, data management and compression become a necessity. Summarization of data can be of great help to users and can improve the efficiency of connected objects in the IoT cloud which can automate the laborious and the resource-hungry task of finding intriguing relevant information, from the large volume of generated data.
To this end, the combination of IoT and data summarization is important to be explored, as it provides innumerable advantages some of which are listed down here: • Saving time and effort of IoT users, by quickly acquiring target information. • Decrease in energy consumption, valuable in low power IoT environments. • Increase in tractability of storage, transmission, and computational requirements. • Increase in conciseness and quality of data. • Decrease in redundant information.
In this work, we explore this relatively less explored area of data summarization in IoT and try to consolidate the notable contributions that have been made. The integration of summarization technology into the IoT framework can be broadly viewed at two different levels of abstraction, based on the approach and type of data, as Low-level and Highlevel. Low-level data summarization is concerned with using suitable resource-curbed, efficient data summarization techniques in highly constrained IoT environments such as WSNs. On the other hand, High-level data summarization deals with generating high-quality summaries of documents at a higher abstraction level by utilizing popular algorithms and concepts from the fields of Artificial Intelligence and Big Data. The High-level data summarization algorithms need not necessarily be less resource-intensive.
The rest of this paper is organized as follows: in Sect. "Low-Level Data Summarization", we consider the topic of Low-level data summarization and describe various possible techniques that can be utilized. Next, High-level data summarization, its multiple classifications, and existing approaches are described in Sect. "High-Level Data Summarization". A set of open research issues is briefly discussed in Sect. "Open Research Issues". Finally, Sect. "Conclusion" concludes the paper.

Low-Level Data Summarization
In low power IoT environments with limited resources, such as Wireless Sensor Networks (WSNs), one of the primary concerns is to maximize the lifetime of the deployed sensors by minimizing the power consumption. Data summarization, as discussed earlier, is one of the most important avenues of research in this regard. Due to low storage and limited computational power of sensor nodes, usage of complex resource-hungry summarization algorithms is forbidden. The need for algorithms that deal with this problem at a lower abstraction level is clearly evident. Data summarization techniques that are most effective in such environments are as follows: • Sensor node clustering • Local data fusion

Sensor Node Clustering
Typically, a cluster is a conglomeration of similar objects, which in the case of IoT environments, can be a virtual grouping of sensors. Clustering for data aggregation has many advantages, such as increased resource utilization, reduced power consumption, optimized transmission, improved fault tolerance, and enhanced scalability.
A sensor cluster generally consists of a Cluster Head (CH), which handles intra-cluster data aggregation from the member nodes and transmits the low-level summarized data as a single packet to the Base Station (BS).
A clustering algorithm deals with the partitioning of groups of sensors into clusters. The typical steps involved in such an algorithm in each clustering interval or epoch are illustrated in Fig. 1. In any clustering algorithm, the process of clustering begins by initializing a fraction of the sensor nodes as CH either randomly or based on some algorithm-specific heuristics. The initial part of each clustering cycle or epoch involves the election of CH and assortment of various sensor nodes into different groups based on numerous considerations, such as energy consumption and transmission range. CHs can be periodically rotated in an intracluster manner to ensure load balancing. The CHs receive data packets from all the members of its cluster. Their responsibility is to reduce the overhead by removing redundant packet headers and summarizing A wide variety of clustering algorithms have been studied by various researchers, while analyzing each of their characteristics. Some of the different algorithms available are briefly described below [2,3]: (i) Voronoi fuzzy clustering-It is a merger of the Voronoi structure and Fuzzy C-Means algorithm [4]. The Voronoi structure is a well-known computational geometric structure consisting of cells, each holding a single sensor node. This algorithm is designed to reduce energy consumption in the sensor nodes [5]. (ii) Voronoi-based genetic clustering-Voronoi structure is used to estimate the sensing range of each sensor node. Subsequently, Genetic Algorithms are used for grouping the nodes optimally [93]. (iii) K-means data relay clustering-Arbitrary selection of "k" cluster heads and member assortment based on distance. This algorithm is experimentally verified to be efficient in terms of computational time and energy consumption [6]. (iv) Highest-degree clustering-Connectivity-based clustering where the sensor node with the highest number of neighbours (or) maximum degree gets elected as the cluster head. All the nodes that are within the transmission range of a CH form its cluster. (v) Lowest ID clustering-Identifier-based clustering in which a unique ID is assigned to each sensor node. CH is elected by choosing a node such that its ID is lower than that of its neighbours (cluster members).

Local Data Fusion
In the context of Low-level data summarization, local data fusion can be defined as the compilation of data received from a local group of similar sources (sensor nodes) or clusters into a more concentrated and efficient message, thereby increasing the quality of final data and decreasing the cost of operation. This process can involve several techniques [7], some of which are: • Collection-Gathering data from a cluster of sensor nodes. • Manipulation-Calculating relevant quantities, such as moving average and standard deviation, from the aggregated data. • Verification-Using the calculated quantities to check if a given data point satisfies certain criteria for summarization. These criteria can be specified by algorithms such as Bollinger Bands (see Fig. 2).
• Elimination-Discarding data points that do not satisfy the criteria imposed by the fusion algorithm under consideration. • Combination-Selectively storing relevant data points to a local database for transmitting later.
The process of local data fusion can typically take place in a CH (see Sect. "Sensor Node Clustering") where the data aggregated can be further summarized by performing simple operations, such as average, maximum, minimum, and total, thereby removing any inherent redundancies, before sending it to the BS. One of the examples of performing a Low-level local data fusion process is following a particular method called Data Summarization in the Node by Parameters (DSNP) [7]. This approach claims that the volume of messages produced in IoT systems such as WSNs can be reduced with a minimal loss.
Intra-cluster transferral of data from other sensor nodes to CH is illustrated in Fig. 3, which shows an example of a Low-Level data summarization system with various sensor clusters. The proposed architecture in [7] consists of an IoT server (e.g., Base Station in Fig. 3), which sends various parameters, such as sample size (n) and type of summarization, that the user or BS requires from the sensor nodes (e.g., cluster heads in Fig. 3). The type of summarization can include performing basic operations, such as Total, Avg, Max, and Min, on the collected low-level sample dataset {X 1 , X 2 , … , X n } . This sample dataset of size n, shown in Fig. 3, represents the aggregated data in CH obtained from its member sensor nodes.
The summarized value can be stored, in the database consisting of data to be transmitted to the BS, or discarded based on the Bollinger Bands algorithm. This algorithm, illustrated in Fig. 2, sends the summarized value for checking, which triggers the calculation of the moving average ( ) and the standard deviation ( ). These quantities are used for computing the upper and lower limits of the tolerance interval [ − , + ] . These limits are also known as Bollinger bands (see Fig. 2). If the summarized value fails to be within the interval defined by these Bollinger bands, it is recorded; otherwise, it is discarded. This algorithm, therefore, avoids unnecessary storage of redundant data points and records any value that indicates a sudden variation in the environment.
Performance studies were conducted in [7] to evaluate the performance of DSNP using an experimental setup based on two different configurations of sensors. The first setup used a normal, sequential mode of operation, and the second configuration transmitted summarized data after the application of DSNP. Each of these configurations consisted of a temperature and a luminosity sensor. These studies revealed that as opposed to the normal recordings, for luminosity sensors, DSNP achieved a significant Fig. 3 An example of a low-level data summarization system deployed using clustering and DSNP techniques, shown at a particular time instant. Note that, the cluster heads can be periodically rotated within the cluster to ensure load balancing SN Computer Science reduction in the number of total records by 97% and a reduction of 80% for temperature and humidity sensors [7].
In essence, data fusion, a popular choice in many fields, such as Robotics, AI, image processing, wireless sensor networks, and IoT, can be effectively combined with sensor node clustering. Furthermore, such techniques can be deployed in complex IoT environments requiring low-level data summarization, such as WSNs, to regulate the volume of data generated and improve data quality and energy efficiency.

High-Level Data Summarization
As discussed in Sect. "Introduction", data management plays a critical role in IoT networks. Many smart devices, such as smartphones, smartwatches, and smart surveillance systems, typically generate higher abstraction level documents, such as textual, audio, and video files, at an exponential rate. Document summarization is a technique used for bringing down the sizes of documents while preserving their informative outlines [8]. Big document summarization, which deals with processing and abridging data repositories of large size, thus becomes an essential aspect of data management in the IoT cloud. Some of the key aspects and technologies involved are [8,9]: • Big data analytics-Concerned with conversion of large amounts of unstructured or structured raw data, retrieved from several sources, into a useful data product for the end-users. • Semantic sensor networking-Deals with collection of data from dense networks of sensors setup to identify events of interests. • Efficient data processing and analysis-Involves processing data using computationally less-intensive architectures to minimize overall resource consumption. • Generation of informative summaries-Concerned with maintenance of quality of the final summaries generated in terms of various metrics including salience, memorability, non-redundancy, and coverage.
An illustrative classification of various subtasks involved in high-level data summarization is shown in Fig. 4. The classification is done based on various considerations such as: • Based on type of summary-Categorization of techniques based on the nature of the generated summary. • Based on security aspects-Methods are classified based on privacy preserving considerations. • Based on documents-Various approaches are categorized taking into account the structure of the inputs. • Based on type of data-Classification is done by considering the high-level nature of the input data. Some of the techniques and sub-classifications that fall under textual, audio, video, clinical, and multi-modal data summarization, respectively, are discussed in this section. Fig. 4 Classifications of high-level data summarization

Textual Data Summarization
As mentioned earlier, the increased use of the Internet has increased the need to access information from big textual data documents of sizes greater than 1 GB. Summarization of such large text documents can assist users in saving time not only in determining whether it is important or not but also in seeking specific information of interest without having to read the whole document in the case of a large database documents. Historically, many of the traditional Natural Language Processing (NLP) techniques involved methods, such as Latent Semantic Analysis (LSA) and its variants [10][11][12][13] using Term Frequency-Inverse Document Frequency (TF-IDF) [14] features. Typical tasks performed in these techniques include sentence clustering based on topic and extracting relevant sentences in a single or multi-document setting.
In one of the approaches [8], big data text documents of IoT cloud are processed, and semantic features such as semantic weight, which communicates how much the sentence reflects critical subjects of the big document data, are extracted. These semantic features are obtained by distributed parallel processing of the Non-negative Matrix Factorization (NMF)-based cloud technique over Hadoop [8]. Also, recent multi-text document summarization techniques are focussed on the use of Deep Learning techniques [15,16]. Figure 5 provides a classification of the same based on various categories. A comprehensive survey covering some of these sub-categories in more detail can be found in [17].

Audio Data Summarization
With the rapid development of affordable technology for digital audio storage and compression techniques, there has been an explosive growth in digital audio data collections in various formats. As a result, there exists a need for capturing the compendium of these audio files, using techniques designed to maximize the utility of the generated summary in an application-specific manner. Therefore, many applications have come into existence to perform various tasks related to the summarization of these audio files: • Automatic music summarization [18][19][20][21][22]-Approaches under this category deal with extraction of a concise summary representative of the salient theme present in the input musical audio. Generating such summaries can be in turn useful in a variety of music applications, such as audio fingerprinting for content-based retrieval, music indexing, genre classification, and web-based music distribution. • Acoustic event detection and summarization [23,24]-This class of techniques is concerned with detection and localization of perceptually important acoustic events of interest present in the audio input. • Personal audio recording summarization [25][26][27]-Methods under this category are concerned with highlevel segmentation of the single-party and multi-party audio recordings. Such segmented audio clips can be used for further research related to personality analysis and social network creation.
Most of the existing approaches for these applications involve extraction of manually designed audio features including Mel Frequency Cepstral Coefficients (MFCC) [19] and its variants, mean instantaneous amplitude and frequency [24], zero crossing rates, spectral flux, and Linear Prediction Cepstrum Coefficients (LPCC) [19]. The work in [23] studied various schemes for the intramodal fusion of extracted features to produce a mono-modal saliency curve. The extracted features are typically further processed using traditional machine learning techniques and related

Video Data Summarization
Video Summarization (VS) has become very important for generating short summaries of the giant volume of video footage generated by the innumerable optical sensors (cameras) of the IoT cloud. For example, these cameras may be present in smart cities to support certain application domains, including healthcare, security, and surveillance systems by detecting, recognizing and reporting significant events autonomously. Due to the huge success and endorsed strengths of Deep Neural Networks such as Convolutional Neural Networks (CNNs) in the field of computer vision, modern VS techniques utilize them for processing input video data streams as shown in Fig. 6. Here, there exists a compromise between keeping up the standards of the summaries produced and efficient data processing in deep networks within an acceptable period, in a resource-constrained IoT environment.
To maintain the informativeness and interestingness of the generated summaries consisting of keyframes (see Fig. 6), optimal feature selection becomes critical. Some of the typical used features in the field of VS are described below [9]: 1. Memorability-It is a property of the image that measures the probability of being remembered by human viewers after a single view. This feature has a high value for frames with people, striking objects, and significant events. Such images are considered more memorable.

Entropy-The amount of information present in an
image is measured by image entropy. The greater the value of this feature, the more information the frame contains and vice versa. 3. Aesthetic Score-It is a measure used to describe the complexity and vividness of a video frame under consideration. A higher aesthetic score corresponds to a colourful and complex frame.
In one of the VS approaches [9], a deep CNN framework is trained along with the hierarchically weighted fusion of Fig. 6 Flowgraph of a typical modern video summarization system in IoT the above-mentioned features is utilized to produce a costeffective solution. Monitoring metrics, such as computational complexity, mean opinion score, and F1 score, are essential while evaluating VS methods.

Clinical Data Summarization
Advance in medical technology has triggered a rapid increase in the use of wearable biosensors for keeping track of various vital signs of patients (see Fig. 7). Such medical applications in turn have opened opportunities for providing better and timely medical care, made possible by realtime monitoring of patient's vitals, and prediction of critical health disorders. However, large-scale use of various sensors in the Internet of Medical Things (IoMT) poses the following domainspecific challenges [94,95]: • Examining voluminous bio-sensor data-Physicians, who already handle many patients regularly, might be overwhelmed and prove to be ineffective for making therapeutic judgments based on the continuous flow of incoming time series data from multiple remote bio-sensors. • Resource constraints-Deployment of remote health monitoring systems in sparsely connected regions is restricted due to unavailability of sufficient data bandwidths. • Preserving clinical relevance of summaries-For the entire monitoring system to be utilitarian, it is necessary to generate intelligent summaries that can help the physician to identify and alert the onset of any critical conditions. • Need for personalization-In most of the cases, it can be useful to provide personalized summaries specific to a disease or patient, as suggested by the physician.
Clinical data summarization techniques generally utilize symbolization of time series bio-signal data obtained from remote bio-sensors [95][96][97]. One of such approaches [94,95], converts the raw sensor time-series data into a set of clinically relevant limited bandwidth symbols called "motifs".

Multi-Modal Data Summarization
Apart from textual, audio, and video summarization, there has been an increase in attention towards data-driven multimodal summarization (MMS) approaches [28][29][30][31] recently, owing to the rapid development of NLP, CV, and Automatic Speech Recognition (ASR) techniques. MMS deals with processing data from two or more streams of different highlevel structures. Based on the overall structure of the multi-modal data content, there exist two broad sub-classifications: • Asynchronous multimedia data-Textual descriptions or captions for images and subtitles for videos are absent. Summarizing data of this type is a more general and challenging task. • Synchronous multimedia data-Images are paired with textual descriptions and videos are paired with corresponding audio subtitles.
As illustrated in Fig. 8, data from different streams or modalities are preprocessed and corresponding features are extracted, either manually using traditional NLP, CV, and ASR methods or automatically using deep representations obtained from deep neural networks. These extracted features are typically fused semantically and are used to optimize an objective function with multiple goals. Some of the popular objectives considered for joint optimization are described below [29,30,32]: 3. Readability-The summary should avoid sentences that are ill-formed, which are occasionally introduced by ASR systems while transcribing speech data. . Coverage-The summary should be comprehensive and must capture the event highlights of the input documents. 5. Size of summary-The summary should be succinct and abridged. Constraints can be introduced into the objective function to fix the maximum size of the summary [30].
One of the MMS approaches [31] proposed a model, which merges electroencephalography (EEG) brain signal features with the audio-visual features extracted from some video clips to facilitate extraction of highly affective and personalized summaries. On the other hand, [32] compute a multimodal saliency curve by integrating features from aural, visual and text streams of videos, to obtain generic summaries independent of the video semantics, syntax, structure, or genre.

Open Research Issues
Despite the advantageous combination of data summarization and IoT, it is a relatively new area, and there exist many challenges that need to be addressed. Some of these challenges are identified in this section to stimulate further research in these directions: • Performance analysis of clustering algorithms-Sensor node clustering algorithms have been studied extensively independently [2,3]. However, their performance needs to be analyzed when they are part of a Low-level data summarization system in conjunction with local data fusion setups. Some of the aspects worth considering during this analysis are energy efficiency and optimal cluster formation capabilities. • Deep learning for audio summarization-Most of the existing high-level audio summarization techniques rely on computation of domain-knowledge intensive handcrafted features and then processing them using traditional machine learning techniques as discussed in Sect. "Audio Data Summarization". As an alternative, it can be useful to explore deep learning-based techniques which involve automatic feature extraction. • Effective and efficient neural architecture design-As discussed in Sect. "Video Data Summarization", there exists a tradeoff between resource-intensive data processing in deep neural networks and quality maintenance of generated summary, especially for video data. Therefore, more research needs to be done in developing computationally efficient architectures to obtain superior summarization performance despite limited resources. • Generating summaries for asynchronous multi-modal data-As far as high-level multi-modal data summariza-tion is concerned, there exist relatively many techniques [33][34][35][36] that discuss solutions for structured and synchronous multimedia data (see Sect. "Multi-Modal Data Summarization"). However, limited research has been done to generate summaries for asynchronous multimodal data, which needs to be explored further.

Conclusion
In this paper, we surveyed the relatively unexplored yet promising research area of data summarization in Internet of Things (IoT) by conducting a detailed survey on various approaches and techniques and examining them at two abstraction levels. The various advantages that integration of data summarization technology into the IoT framework can offer are discussed. Also, for increased clarity in understanding this integration, we classify it into Low-level data summarization and High-level data summarization and covered various algorithms and techniques that fall under them. Finally, we also discussed a few challenges that remain open and further stimulate research in this area.
Funding None.

Conflict of interest
The authors declare no conflict of interest.