BIBLIOMETRIC ANALYSIS OF NEURAL BASIS EXPANSION ANALYSIS FOR INTERPRETABLE TIME SERIES (N-BEATS) FOR RESEARCH TREND MAPPING

Article History: Bibliometrics is the statistical analysis of articles, books, and other forms of publication. The bibliometrics analysis is performed with data on the number and authorship of scientific publications and articles, and citations to measure the work of individuals or groups of researchers, organizations, and countries to identify national and international networks and map developments in new multidisciplinary fields of science and technology. In addition, bibliometrics assesses and maps the research, organization


INTRODUCTION
Bibliometrics comes from the words Biblio or bibliography and metric. Biblio means book or bibliography, and metric is related to measurement. Bibliometrics is a quantitative method used to measure, track, and analyze scientific literature [1]. According to [2], bibliometrics is divided into two groups, namely groups that study publication distribution and that discuss citation analysis. The first group analyzes documents quantitatively so that there are 3 (three) bibliometric laws, which are described in the following paragraphs.
The first is Lotka's law (1926) [3], to calculates the distribution of results of different authors. It contains a mathematical relationship between the number of works produced and the number of authors. Lotka studied this phenomenon in 1926, but it was not until 1949 that Lotka's law of author productivity was declared. Lotka's law is one of the main research subjects in bibliometrics. According to Lotka, if a hundred people make one work, then a quarter of every 100 people will make two works, one in nine out of 100 will make three works, and there will be a third. Sixteen out of 100 people do the work. Four works, etc. This is Lotka's inverse square law of productivity.
The second is Zipf's law (1933) [4] [5], to classify words and their frequency of occurrence in literature, namely the frequency of occurrence of words in the document (words appear or frequency) turns out to have a specific pattern. It can be used as a parameter during indexing. Studies like this have yet to be developed in Indonesia, but the research is known as Zipf's Law abroad.
The third is Bradford's law (1934), which estimates the exponential decline in search results for references in scientific journals. One formula stated that if the number of articles sorted journals in a field into three groups, each containing about a third of all articles, the number of journals in each group would be equal to 1:n:n² [6] [7], where n is the "Bradford multiplier" and depends on the particular journal collection. There are several formulations related to this principle. In many disciplines, this pattern is referred to as the Pareto distribution.
According to [8], bibliometrics has three components: bibliometrics for bibliometrics; bibliometrics for science (information science); and bibliometrics for science policy and management. The Bibliometrics analysis (BA) is used to quantitatively analyze and measure various specific indicators in the published literature in a particular field to obtain knowledge maps based on large databases [9]. It allows researchers to summarize publication information regarding article distribution by author, institution, year, journal, and specialty, author-organization collaboration, and split word analysis [10]. The BA is based on relevant information about scientific publications/documents. The necessary information that can be used is to identify the source (journal/document, volume, page), author's name, institution/institutional address, references, type of document, title, keyword, abstract and subject. Based on some publication/documentary information, the researcher used the author's name of the document type of the journal in Conciencia as the basic unit in the BA.
The current research results are the results of research conducted and published in journals, proceedings, seminars, books, and the like. Bibliometrics is widely used to identify subject trends, such as; identifying key journals and library usage patterns. Bibliometrics is also used to build research models of scientific communication. Most of these models are tested and used mainly at the institutional level to explain scientific productivity; explain publication growth, identify critical journals; and document screening.
This article is a bibliometric analysis from the direct science page, which discusses Neural Basis Expansion Analysis for Interpretable Time Series (N-BEATS) using VOSviewer. N-BEATS is a deep learning architecture that connects backward and forwards residuals to a stack of connected classes commonly used for univariate time series prediction [11]. The N-BEATS deep learning architecture has the property of being independent of any particular specification of the time series or input rate [12]. The advantage of using N-BEATS architecture over other deep learning architectures is interpretability and does not depend on feature engineering or input scaling in a particular time series [13]. Using the N-BEATS architecture, it is possible to process time series data without preprocessing or decomposition [11].
N-BEATS architecture is divided into two basic block parts. First, the basic block has four fully connected (FC) layers with nonlinear ReLu. This part produces two outputs, a forward predictor or "forecast" ( ) and a backward predictor or backcast ( ). Second, the basic block has a fully connected (FC) layer without nonlinear ReLu and base layers and . The base layer is a linear projection layer. In the block ℓ,there is an input ℓ , and two output vectors l , and l . l is the block's best estimate of ℓ or backcast, and l , is a forward forecast on the block of horizon length ( ),which has a range of values from 2H to 7H. The N-BEATS architecture uses two configurations: generic architecture (N-BEATS-G) and interpretable architecture (N BEATS-I). Some Scopus-indexed N-BEATS research in the last ten years needs to be mapped to obtain the latest research trends. Therefore, this article examines N-BEATS for research trend mapping with bibliometric analysis.

RESEARCH METHODS
This research is a literature review by summarizing and evaluating the N-BEATS dataset using bibliometric methods based on Co-word analysis. Co-word analysis is based on the assumption that the keywords of a paper or article are an adequate description or can make a representation of the content of its content. The research was conducted by screening articles from the Scopus database site from 2008 to 2022 with the keyword N-BEATS [14]. The data obtained is then called metadata. Metadata in the form of the number of publications per year, journals that contain articles in the field of N-BEATS, authors, author origin, and subject.  Based on Figure 1, the steps taken are defining search keywords, initial search research, refinement of the search, compiling the initial data statistics, and data analysis. Refinement of the search is done by adding filters and restrictions so that the data obtained will be more specific and specific. Compiling the initial data statistics is done by looking at the distribution of each research using various graphical visualizations such as histograms and line graphs.

Development of N-BEATS Publications in Scopus Metadata
Scopus is the world's most extensive collection of literature abstracts, with citations that provide summaries of a wide range of peer-reviewed scientific publications and research. Scopus can help researchers effectively track, analyze and visualize research. More than 22,000 high-quality abstracts published by 5,000 publishers worldwide are provided in the Scopus database from science, technology, medicine, social sciences, arts, and literature. Scopus has 55 million records since 1823, 84% of which are from Elsevier reference citations [15].   Figure 2, the articles processed in this study included 647 data and showed an upward trend except in 2010 and 2013, which showed a decrease. Figure 3 shows the number of studies by country.

Bibliometrics with VosViewer
The earliest known bibliometric research dates back to the early 20th century. In the early stages of its development, the bibliometric method was widely applied in the medical field [16]. Around the 1940s, the development of bibliometrics had strong foundations from the formation and maturity of Brad Ford's Law,

Zipf's Law, and Lotka's Law [3] [17] [18]. Since then, this method has been widely used in natural sciences [19], subjects such as mathematics [20], physics [21], and chemistry [22].
Visualization of Similarities Viewer (VOSviewer) is a computer program with free access to explore bibliometric knowledge network maps [23]. The algorithm used in VOSviewer is similar to the Multi-Dimensional Scaling (MDS) algorithm. VOSviewer is capable of generating clusters that automatically display different colors on the map. The clustering algorithm works with a parameter (ɣ) that can be modified to get more or fewer clusters. Still, according to [23], cluster density and color can be visualized using VOSviewer. The advantage of VOSviewer over other analysis applications is that it uses text mining to identify mapping-related noun phrase combinations and built-in clustering methods to test the data [24] [25]. Although there are various programs that analyze the similarity of text units and matrices, VOSviewer has a strength that lies in its intuition [26]. VOSviewer also has various interactive features and options that make it easy for users to access and explore bibliographical data, such as the co-occurrence relationships between key terms and concepts and the number of citations therein [24] [26].
On the metadata, bibliometric analysis was carried out using the VOSviewer application, with the analysis carried out including Co-occurrence analysis, used to reveal research topics/variables statistically, and co-authorship analysis used to find the relationship between several researchers based on research documents produced by researchers. The bibliographic analysis explores research fields that change yearly due to researchers' findings and attention to different research problems.
Bibliometric analysis is done by visualizing networks, overlays, and densities to identify bibliometric networks between articles or online publications from downloaded metadata. A bibliometric network consists of vertices in the form of circles that represent keywords, while the edges or nodes of the network represent the relationship between pairs of vertices. Mapping and clustering in bibliometric analysis through VOSviewer software are complementary. Mapping can be used to obtain a detailed picture of the structure of the bibliometric network [23]. In addition, clustering provides an overview or preview of bibliometric clustering. The approaches used in the bibliometric analysis are citation analysis to see one article cited by other articles and co-citation analysis to find two or more articles cited by one article.

Bibliometrics Study Coverage
Scientific concepts in a document can be seen through the words (co-words) used. Co-occurrence analysis is the basis of collingual analysis. Keywords that originate from two or more different documents serve as indices for these documents [27]. Co-words analysis aims to analyze various documents' content, patterns, and trends by measuring the terms' strength [28] [29] [30] [31]. Synonym analysis calculates the number of research document keywords that co-occur in the articles under study. Previously, these keywords were set by the author. The relationship between documents is stronger when many keywords exist in a particular document group.
The co-words analysis map is based on the appearance of the same keyword in the article, the occurrence of unique or important terms in the article, and the title or abstract. These keywords and terms represent a concept from the subject analysis. Using denormalized keywords can lead to inconsistent terms, and normalization requires using a thesaurus. A thesaurus is a list of terms that cover one particular field so that the terms used are more specific. A thesaurus differs from a list of subject headings, which are usually general and cover all fields of science. Indexing uses descriptors that represent each concept. Standardizing keywords with a thesaurus aim to make the words used consistent so that one term is used for concepts presented in different articles and has the same meaning.
Furthermore, the metadata was analyzed using the VOSviewer application by selecting the data option 'create a map based on text data,' which aims to create a term network or term relationship (term) based on text data. Figure 5 shows the results of this visualization through extraction in VOSviewer with title and abstract fields with complete counting and 10 number of term occurrences and obtained eight groups. describes what variables frequently appear associated with N-BEATS. The larger the circle shown for each variable indicates that the variable is used more frequently. The terms dominate this: approach, feature, heart rate, frequency, computer, mechanism, wave, and application. There are 7 clusters for coauthorship analysis with an author analysis unit and complete counting method with a restriction of a maximum number of authors ten and authors minimum of 2 documents. It can be seen that Zhang, y; Wang, h; Wang, y; li, y; Xu, x; Zhou, y; and li, j dominate each cluster, and Zhang, y contributes the most as an author. Figure 6 shows that some authors partnered with other authors to publish.

Figure 6. Visualization of Network Author and Co-authorship
Based on Figure 6, the larger the circle shown for each author's name, the more frequently the author publishes, such as Zhang. Next, there are 17 clusters for co-occurrence analysis type with keywords analysis unit and complete counting method with a minimum restriction of two keywords. Of the 1700 keywords obtained that meet the criteria, 172 documents result from their network co-occurrence with other keyword links, as shown in Figure 7.  Based on Figure 8, the N-BEATS algorithm has a close relationship with deep learning. Apart from that, other related algorithms are LSTM and CNN, as well as several fields of application such as temperature, heartbeat, and arrhythmia.

CONCLUSIONS
Based on the results and discussion, it can be concluded that the development of documents related to N-BEATS in 2008 -2022 indexed by Scopus was highest in 2022, reaching 310 publications (14.90%).
China's Ministry of Education is the institution that publishes the most N-BEATS research, and the United States is the most significant contributor. Zhang, Y is the most prolific author with the most subjects in Medicine and Engineering. The N-BEATS development map is based on co-words clusters into seven and co-authors clusters into 9 clusters. The direction of N-BEATS research is growing over time, supported by collaboration, given the rapid development of interpretation in N-BEATS. In addition, the bibliometric mapping identifies authors who take the theme of the N-BEATS field and have a collaborative author relationship with each other. The last mapping is the development of the field of N-BEATS science based on co-occurrence (keywords). This mapping identifies the relationship between scientific concepts and nine dominant groups. Thus, the direction of future research is to serve as a basis for further research on archives and see the variety of topics that can be researched, not only in the national scope but also in the international realm.