paper I want to fix it - clean it up -organize it and add

Capturing User Generated Video Content in Online Social Networks

Clinton Daniel(&), Matthew Mullarkey, and Alan R. Hevner

Department of Information Systems Decision Sciences, University of South Florida, Tampa, FL 33620, USA {cedanie2,mmullarkey,ahevner}@usf.edu

Abstract. We build and evaluate an innovative artifact for the investigation of social content derived-platforms specifically to gain a unique understanding of the content shared and underlying behaviors of the contributors to these tech- nology platforms. The artifact‘s innovation is derived from the solution’s unique approach to converting and analyzing the multimedia – especially video - content to gain interesting insights into the social network connectivity of the actors on a given technology platform. The artifact directly addresses a practical need for industry practitioners to analyze social video network content using a rigorous and evidence-based DSR approach.

Keywords: User video content � Social network technology platforms DSR � Elaborated ADR

1 Introduction

Social networking continues to grow through the generation and posting of user- generated multimedia content on social network technology platforms. These users post content in order to network and connect with others sharing similar interests. As the capacity for multiple media formats and the capabilites of these technology platforms continue to increase the resultant volume and variety of content increases exponen- tially. This media content has expanded from text to include images, videos, and animations. In this research, we use an elaborated action design research (eADR) approach [23] within the design science research (DSR) paradigm [17] to diagnose, design, implement, and evolve an innovative artifact for the investigation and mining of the multimedia spectrum of this growing content. The goodness of the digital artifact is evaluated in situ with practicing professionals and proposed as a unique IT solution for the class of problem.

Multimedia technology platforms, such as YouTube, exist to provide video content as the primary medium for sharing and connecting. Users post to these platforms as a means to communicate throughout a network with the understanding that the content can be consumed and easily accessed by other users with similar interests. Although the context of the medium, such as video, is easily viewed by other actors through either a visual or audible interpretation, it is not easily analyzed by those who want to gain an understanding of its semantic or latent meaning or by those seeking to understand the underlying social behaviors of the actors. The analysis of video content across a

© Springer International Publishing AG, part of Springer Nature 2018 S. Chatterjee et al. (Eds.): DESRIST 2018, LNCS 10844, pp. 333–347, 2018. https://doi.org/10.1007/978-3-319-91800-6_22

multimedia platform presents researchers and practitioners with a rich contextual source of information on social networks, interpersonal norms and behaviors online, and interactions between platform users. Unfortunately, significant technical and algorithmic challenges exist for the analysis of multimedia content and this has tended to limit the ability of researchers and practitioners to make sense of this rich social content in context.

Our eADR approach was used to derive an innovative artifact for the search and analysis of multimedia content in a social content derived platform (YouTube) to gain a unique understanding of the content shared and underlying behaviors of the contrib- utors to these technology platforms. The artifact’s innovation is derived from the solution’s unique approach to converting and analyzing the multimedia content to gain interesting insights into the behaviors of the actors active in the technology platform.

2 Motivation

Social networking data are analyzed and studied within academic research and industrial practice as a source for understanding social phenomena since its inception [25]. Typically the analysis takes the form of social network analysis of the connec- tivity between nodes and concentrates on edges, links, and ties that connects them. In some cases the nature of the interaction between nodes is analyzed by analyzing the text exchanged. In these cases, unstructured data generated by textual exchange in social networks involves the extraction of the text from a social networking source, transformation of the data into a workable data set, loading the data into a structure for further processing, constructing a text corpus, and finally using the results of the processed corpus to apply an algorithm that will execute the analytics [5].

The approach for text mining in these cases is generally well understood. Nonetheless, often in each of the phases throughout this process a developer is required to build a custom solution that will accommodate the corpus construction requirements attributable to a given social network, the nature of the content, and the type of connectivity involved. Inevitably, the data are then attributed to the network of nodes and act as a measure of their connectivity.

In our research, we find that online social networks create a collection of content (i.e. corpus) that result from the multimedia nature of many of these networks. In the metadata, these social networks use a variable set of media that can define the context of its communication. If we as researchers are to investigate more deeply than the graph structure of the network or the text exchanged between nodes, then we need a rich, variable, innovative digital design to analyze the composition of the multimedia information within any given social network.

This research proposes a methodical process for the artifact instantiations involving practicing experts from industry and academic researchers that combines the value of the corpus construction with the unique attributes defined by the multimedia metadata in a social video network. Due to the unique attributes that define the metadata gen- erated by social networks which use video as its primary medium for content sharing, YouTube is used as a social network data source to investigate and develop an inno- vative artifact for the investigation and analysis of this third component of many social

334 C. Daniel et al.

networks. We conduct our research using the paradigm of Design Science Research (DSR) [16] and focus on the abstraction of an innovtive artifact through a series of iterative interventions as detailed in the elaborated Action Design Research (eADR) [22, 23]. Specifically, this research will answer the question of “How do we construct a metadata-enhanced corpus from social video networks?” and “How might an instan- tiated artifact help define a class of IT solution for analyzing multimedia in online social networks?”.

3 Literature Review

The principle goal of our literature review is to identify a specific social video network platform that has been consistently used within practice for the exchange of rich non-textual information between multiple nodes in a network. For the purpose of our literature review, a social video network is defined as a social network that uses video as its primary medium for content sharing. We searched for all articles that described a multimedia investigation and found 19 research articles published between 2007 and 2017 (see Table 1). In each case the investigators highlighted YouTube as a principle data source.

The research survey identifies YouTube as an interesting social network where video media defines the content and links to the connectivity of nodes. These studies contribute to knowledege of social network behavior in a variety of ways. For example, a study published in 2017 by Malik and Tian [20] cites the importance for researchers to data mine YouTube videos and their related metadata to support new knowledge discovery in domains such as STEM education and the medical sciences.

Additionally, the literature review reveals a significant overlap in methods used to collect data from YouTube in order to conduct research activities such as social net- work analysis, descriptive statistics, data characterization, data property assessment, content analysis, market analysis, trend analysis, consumer impact analysis, and effectiveness of automatic captioning as shown in Table 1.

In general, most research studies require the acquisition of YouTube’s data to conduct some type of emperical analysis. In most cases, the researchers acquire YouTube’s metadata for analysis which includes attributes such as the title, comments, channel, video category, video id, timestamps, like counts, and other aggregated statistics. Researchers extract the YouTube data using data collection methods such as implementing the YouTube Data API, web scraping, web crawling, custom scripting, and manual processing. The data are then extracted to files and/or relational database management systems for further processing and analysis. For instance, an article by Ahmad et al. [2] demonstrates an entire framework that is motivated to analyzing the content generated by the YouTube platform. Such sophisticated artifacts collect comments and other metadata using the YouTube Data API. Other sophisticated arti- facts, such as the one architected in Chen et al. [9], acquire YouTube captions with the use of web crawling techniques.

In the collection of articles reviewed, we observe that no research reports the implementation of an artifact that is designed to collect both the YouTube captions and the metadata using the YouTube Data API. We found that there is a strong emphasis on

Capturing User Generated Video Content in Online Social Networks 335

Table 1. Summary of literature using content generated by YouTube as a research data source

Author(s) Research motivation Data collection methods Publication year

Cha et al. [6] User-generated content linked to videos

Web crawler 2007

Gill et al. [13] Traffic characterization YouTube Data API 2007 Santos et al. [26] Analyzed structural

properties and social relationships

Custom web crawler and extractor 2007

Cheng et al. [10] Video statistics Web scraping, web crawling, YouTube Data API

2008

Lange [19] Study how video sharing supports social networking

Manually viewed videos and reading of comments to recruit interviews

2008

Abhari et al. [1] Understand YouTube traffic

Web scraping 2009

Chatzopoulou et al. [8]

Fundamental properties of video popularity

YouTube Data API 2010

Siersdorfer et al. [28]

Commenting and comment rating behavior

YouTube Data API 2010

Yuan et al. [31] Concept-based video search

Manual collection using YouTube site

2011

Figueiredo et al. [12]

Characterized growth patterns of video popularity

Google Charts API 2011

Yoganarasimhan [30]

How network structure impacts diffusion of content

Custom Perl script to parse HTML, MySQL

2011

Miller [21] Content analysis Comments were manually coded 2015 Chen et al. [9] Trend analysis Web crawling, raw caption file 2015 Gupta et al. [14] Framework for

marketing/promotional strategies

Video content was manually coded 2016

Harrison et al. [15]

Evaluate the impact of consumer-targeted videos

Manual collection of video metadata

2016

Ahmad et al. [2] Content analysis YouTube Data API 2017 Smith et al. [29] Effectiveness of the

automatic captioning tool

Manually viewed video and automatic captions, then analyzed errors in spreadsheet

2017

Malik et al. [20] Data collection and discovery

YouTube Data API, MySQL, Java 2017

336 C. Daniel et al.

the importance of collecting YouTube data to answer research questions in a variety of different domains. However, although the literature supports the implementation of more sophisticated techniques for analyzing social video networks such as YouTube, they are limited by their ability to automate the collection of captions and metadata using the YouTube Data API.

We shared our observations with practicing professional experts in the problem class of mining social network content and received confirmation that no existing IT artifact – tool or technique – exists to adequately mine the multimedia captions and metadata in context in social network sites for user generated content like YouTube. We identify this gap as the opportunity for the creation of an innovative artifact for the extraction and analysis of the full, rich YouTube multimedia content to mine the content in an automated, scalable manner that sheds new insights on the nature of the social behavior in multimedia social network platforms.

4 Method

The host platforms for social video networks, such as YouTube, offer complex application programming interfaces (APIs) to access their data structures for analysis. These APIs are complicated but offer unique access to metadata that is otherwise difficult or inaccessible to a consumer of the social content. In order to use these APIs for data analysis, an application programmer must develop a custom solution to con- struct a document corpus that will meet the requirements and address a specific research or business question within a specific context. Additionally, the programmer may require a technical solution that is capable of responding to changes with the question as intelligence is acquired through the supplement of metadata. This presents researchers and practitioners with what is called a wicked problem [16]. Design Science Research (DSR) offers a unique framework that is designed to address wicked prob- lems and expose research entry points which can directly address a technical problem domain [16, 18].

One product of DSR can be an innovative IT artifact that can be abstracted and implemented as a generalized utility within a problem domain. The Design Science Research Methodology (DSRM) developed by Peffers et al. [24] can be used for the production and presentation of a DSR artifact that addresses a given problem and can be generalized to an abstracted innovative artifact that addresses a class of problems with a practice-inspired, research-ingrained instantiated system.

Each iteration of the DSR process offers an opportunity to inform research and practice through the build and evaluation of an artifact. Sein et al. [27] proposes the Action Design Research (ADR) methodology which can be used for the creation of an ensembled DSR artifact through building, intervention, and evaluation activities that occur in whole or in part during interventions with practitioners. The DSR artifact can then be instantiated within the context of practice-inspired requirements that have been influenced by an iterative and theory-ingrained process. The elaborated Action Design Research (eADR) methodology proposed by Mullarkey and Hevner [22, 23] offers both researcher and practioner with an ADR process continuum that provides an expanded framework for iterative design cycles with multiple project entry points.

Capturing User Generated Video Content in Online Social Networks 337

We use the eADR methodological approach to iteratively diagnosis, design and implement an innovative artifact to analyze social video network content (see Fig. 1). Each of these stages are informed by a continuum of ADR cycles that iterate until a state is reached which indicates a contribution toward one or more innovative artifacts.

For instance, as described in the Literature Review section above, several iterations of eADR cycles are implemented to inform the “Diagnosis” stage where we worked with experts at our host companies and the technical literature to gain a precise understanding of the existing solutions in this class of IT problems. Practical discovery of the problem domain was inspired by expert interviews from firms such as LeapDoctor.com and DIYCaptions.com. Upon further evaluation of the data collected by the expert interviews, the literature review was initiated to further refine an understanding of the class of IT solutions available and determine the state of the solution domain. Each time an iteration of the ADR cycles within a stage was satisfied, the process was advanced to another stage.

To re-enforce the “practice-inspired” principle of the eADR methodology, design attributes of the artifact were initially influenced by a practical problem exposed during an expert interview with LeapDoctor.com. This intervention had a direct impact in the initial design of the research artifact. Additional inspiration for the DSR artifact was derived from an expert evaluation and intervention with DIYCaptions.com of the initial version of the video media analysis artifact. These expert interventions acquired through multiple eADR iterations directly contributed to the overall definition of the “goodness”, where “goodness” represents the measure to satisfy the problem domain, and practical utility of the desired ensemble artifact. The eADR Cycles make explicit the iterative, cyclical intervention activity at each stage that abstracts one or more artifacts for evaluation. This iterative approach helped us develop multiple innovative artifacts at each stage in the methodical movement towards an implementable social video content analysis solution.

Fig. 1. Illustrates the nature of the interventions at each stage in the eADR method. Adapted from [23]. Note: in the eADR cycle: P = Problem Formulation/Planning, A = Artifact Creation, E = Evaluation, R = Reflection, and L = Learning.

338 C. Daniel et al.

The next section describes the continuation of the eADR approach as the team of practitioners and researchers moved into artifact design, implementation and evolution.

5 Artifact Description

The intial version of the artifact was evaluated by researchers and practioners for the purpose of informing design decisions. Once design decisions are made, the eADR cycles iterate to further refine the design of the artifact. For instance, during one of the first iterations within the “Design” stage, a topic model analysis is performed to analyze the YouTube transcript document corpus that includes all of the target video captions. Specifically, Latent Dirichlet Allocation (LDA) [4] is used to generate a probable distribution of topics that include a probable distribution of terms which are distributed throughout the document corpus. This algorithm is chosen to perform analysis on the social video network because it generates a generalized probable distribution of topics that could describe the conversations occurring within the corpus of captioned text documents.

During a subsequent Design iteration, we plan, abstract, and evaluate a summative task required to measure the topics generated by LDA. We follow this iteration with a topic intrusion evaluation task. Chang et al. [7] developed topic intrusion evaluation tasks as a means of identifying topics that are generated by the LDA algorithm which may not fit well with the others. In other words, topic intrusion tasks help to identify an intruder topic among a distribution of topics generated by the LDA algorithm. This iterative evaluation of possible solutions lead to the generation of two potentially viable designs for consideration and evaluation prior to implementation.

5.1 DSR Artifact Version 1

The first version of the implementated DSR artifact consists of a five phase process to construct the text corpus required to peform social video network analysis. These five phases (as seen in Fig. 2) collectively define the artifact in the form of a process and can be summarized as follows:

1. Extract caption data from YouTube using API and Downsub.com. Webscrape YouTube metadata using Python. This phase of the design is motivated and informed by the results of the comprehensive literature review. The literature review reveals a constistent use of the YouTube Data API, web scraping, and custom scripting to create a collection of files (a corpus) for analysis.

2. Generate .TXT and .SRT files for store on local file system using Python. Because the initial phase in this artifact requires the extraction of data from a source, a collection of files must be stored in a file system for further processing at a later phase.

3. SQL Workflow loads and processes files into RDBMS. A Microsoft SQL Server Integration Services (SSIS) custom workflow is developed to extract, transform, and load the .TXT and .SRT files into a set of Microsoft SQL Server relational tables.

Capturing User Generated Video Content in Online Social Networks 339

These tables are constructed in schemas which support staging, processing, and permanent storage.

4. Export Caption data to text file corpus for analysis using SQL. Once the data has been structured for permanent storage in the relational database, a custom SQL script is written to extract the captions from the relational database to a text corpus for analysis. The SQL script includes a query of the collected metadata to produce a targeted set of text files.

5. Execute data analysis with Topic Modeling algorithms using R. Once the targeted text corpus has been constructed, social video network analysis is performed using R. The topic modeling algorithm, Latent Dirichlet Allocation (LDA), is used to generate a distribution of topics.

The first version of the DSR artifact is consistent with the work of other researchers as observed within the literature review. The new contribution of the artifact is the methodology in which it was constructed and the use of a third-party (Downsub.com) tool combined with the YouTube Data API to extract the actual captions of a YouTube video. Our research is motivated to acquire the YouTube captions for the purpose of using machine learning algorithms, such as LDA, to generate a distribution of topics. These topics are used to generalize the conversations that are occuring within the social video network based off of a search term such as “emerging technologies 2016”.

The use of Downsub.com is incoorporated into the version 1 design because we quickly realized that the YouTube Data API could only extract video captions using the “captions.download” method if third-party contributions are enabled for the specific caption. The API would return a 403 error on many of the captions we attempted to download using the “caption.download” method. The YouTube Data API documentation

Fig. 2. Illustrated summary of the two ensemble artifact designs after iteration through diagnosis and design stages of the eADR method.

340 C. Daniel et al.

reports this 403 (forbidden) error response to be: “The permissions associated with the request are not sufficient to download the caption track. The request might not be properly authorized, or the video order might not have enabled third-party contributions for this caption.”. Third party applications, such as Downsub.com, have the ability to download captions that are forbidden and otherwise unaccessible via the API. This led us to question how it is possible that third party applications are able to acquire these data based on the constraints of the API. Thus, we needed to find a new, innovative approach to achieve our goals of understanding video content on social networks.

5.2 Digital Innovation

In late 2017 we started investigating third party applications to determine how they are able to extract these data. At that time, we hypothesized that the lack of scalable access to the caption data may have been a major contributor to the limited use of these data found in current research. We discovered a third party web-based tool at diycaptions. com that would accept a YouTube videoID as a search term to generate automatic captions. The videoID is used by YouTube to identify a unique video posted to its platform and can be found within a YouTube URL. We tested the videoIDs that were forbidden for download by the API and they all returned captions for download. We contacted the site developer and asked him if he would be willing to evaluate our artifact and perhaps provide expert consultation on how we could improve upon our current design. His response to our e-mail was, “I’m using an API that Google uses internally but which isn’t documented for public use. I’d be happy to talk to you about it.” We then injected the helpful DIYCaptions.com developer into our eADR process so that he could evaluate and intervene with our design.

Working with the expert we identified that YouTube has an undocumented API method called “get_video_info” that can be called from a URL. An example of this URL is: https://youtube.com/get_video_info?&video_id={videoid}.

The {videoid} portion of the URL would be substituted with a specific videoid associated with the video we are targeting for a caption extraction. The URL then exposes a very long (can be up to 16 pages in a document) and complex encoded string. We had to decode the string to expose a JSON data structure buried within the file that contains important fields needed to extract the targeted caption. The two most important fields found within the encoded JSON string are baseURL and languageCode (see Table 2). We found that the baseURL field provides a direct link to an XML document which contains the caption data for the specific videoid. Additionally, the languageCode field identifies which language the captions are available for translation. For instance, in the example included in Table 2 there are two baseUrl entries in two different languages. The languageCode “nl” will take the user to a caption translated in Dutch while the languageCode “en” will take the user to a caption translated in English. Another valuable discovery we made with these URLs is that they will expire within a short period of time. So, it was important that we extract the XML data from the URL that includes the captions translated to English before the accessible URL expires.

Capturing User Generated Video Content in Online Social Networks 341

5.3 DSR Artifact Version 2

Capitalizing on the digital innovation that we achieved during the eADR cycle eval- uating the limits to the goodness of DSR artifact version 1, we were able to proceed with improvements and construct version 2. Version 2 of the DSR artifact consists of four phases (see Fig. 2) to include:

1. Web App used to auto search YouTube with API. This feature executes stored procedures that use Python within the Relational Database Management System (RDBMS). SQL Server 2017 is used as the RDBMS to write Python embedded scripts within SQL statements. These scripts are then executed as stored procedures and permanently stored in a database residing within the SQL 2017 instance. The stored procedures would then receive parameters that are passed as arguments through the web front end application in the form of a search term. The search term received from the web app would then be passed through to the YouTube Data API and retrieve a list of relevant video ids. A partial screen shot of the web app, we call “TUBE TOPIC”, is illustrated below in Fig. 3.

Table 2. Example of encoded URLs found within the JSON string fields baseURL and languageCode. This JSON string is encoded within the get_video_info data.

baseUrl languageCode

https://www.youtube.com/api/timedtext?caps=asr&hl=en_US&sparams= asr_langs,caps,v,xorp,expire&expire=1516448002&xorp=True&signature= C136764856DB2D4997A522236FBA65079B44D5BF.9044FB3AB0E434 90946A9C4E450F6789A4832591&key=yttt1&asr_langs=nl,pt,es,ru,en,ja, fr,de,ko,it&v=6MBaFL7sCb8&lang=nl

nl

https://www.youtube.com/api/timedtext?caps=asr&hl=en_US&sparams= asr_langs,caps,v,xorp,expire&expire=1516448002&xorp=True&signature= C136764856DB2D4997A522236FBA65079B44D5BF.9044FB3AB0E434 90946A9C4E450F6789A4832591&key=yttt1&asr_langs=nl,pt,es,ru%2Cen, ja,fr,de,ko,it&v=6MBaFL7sCb8&lang=en

en

Fig. 3. Illustration of the DSR artifact version 2 web app called “TUBE TOPIC”

342 C. Daniel et al.

2. Extract captioning and metadata from YouTube using API and DIYCaptions.com. Execution embedded Python and SQL within RDBMS. Some of the YouTube captions are extracted using the YouTube Data API. However, after collaboration with the developer of DIYCaptions.com we decided to leverage their technology and extract some of the captions using their web application. Collectively, we are able to store all necessary captions and metadata within the SQL Server 2017 instance that were required to perform social video network analysis.

3. Export captions and metadata to corpus for analysis using SQL. With a scalable and data rich version 2 of the DSR artifact, we are able to extract the captions from the SQL server instance, similar to version 1, and construct the text corpus for social video network analysis. However, in version 2 we are able to construct each text file to include both the caption data and the metadata into a single file. This is what we term a metadata-enhanced corpus.

4. Execute data analysis with Topic Modeling algorithms using embedded R within RDBMS. The metadata-enhanced corpus is then used by topic modeling algorithms, LDA, to generate a distribution of topics for analysis. The advancement of this process in version 2 is that it includes scripts within SQL Server 2017 that allow R code to be embedded within the SQL code. This allows the system to execute a single script in a single service environment without the movement of files from one location to another.

6 Evaluation

As seen in Fig. 1, evaluation of each abstraction of a DSR artifact occurs within each iteration of the eADR cycles. The first opportunity for evaluation occured within the eADR cycles as they informed the Diagnose stage. An expert interview was conducted with LeapDoctor.com that resulted in the practical inspiration to develop the initial version of the DSR artifact. For instance, during the interview a question was asked about how LeapDoctor.com would make a decision once they recognized they had a technical problem. Part of the response was: “If we are going to start from scratch – we have a lot of options. There are a 1000 ways to do the same thing. So we asked ourselves, what kind of technology do we want to use? Who had the technical expertise to work on the new technology? If we were to use the old technology, what kind of sense does this make if we are not using new technology?”

We recognized that there is a need for a practical utility that any firm could use to easily discover what people are “talking about” within the context of a domain of interest. In early 2017, the development of version 1 of the DSR artifact was initiated using YouTube’s social video network data to meet the practical requirements of the artifact inspired by LeapDoctor.com. Evaluation of the version 1 artifact was performed via user reviews [11]. Additionally, an expert interview and evaluation of the version 1 artifact was conducted with the lead developer of DIYCaptions.com in late 2017. The evaluation and intervention tasks conducted within each eADR cycle were critical to establishing the requirements for version 2 of the artifact.

Capturing User Generated Video Content in Online Social Networks 343

Early on in the development of the version 2 artifact, a Reflection phase of an eADR cycle revealed that we needed an algorithm that would generalize a conversation that a user could use to evaluate the output of the artifact. A topic modeling algorithm, LDA, was selected because it resulted in a distribution of topics that consist of a distribution of words which could be used provide a probabilistic implicit view into a group of conversations discovered within YouTube. Topic intrusion tasks, as described by Chang et al. [7] and reproduced within the medical domain by Arnold et al. [3], were conducted to evaluate the results of the topics generated by the DSR artifact (see Fig. 4).

Ultimately, each iterative cycle of artifact planning, abstraction, evaluation, and reflection leads to learnings that inspire the next cyle within or between a stage in the eADR method. Once we had implemented an ensemble artifact, our evaluation of its goodness in situ with practitioners demonstrated a number of limitations on its utility. A second iteration through the Design stage yielded a breakthrough in our options for solving the problem. The resultant digital innovation has been implemented and its evaluation, in situ, resulted in a consensus among the practitioners and researchers that a unique ensemble artifact has been constructed to meet the unique needs of this problem class.

7 Discussion and Conclusion

The DSR artifact constructed in this research contributes toward a novel artifact that could add value to industry practicioners and scholarly researchers. The artifact directly addresses a practical need for industry practioners to analyze social video network content using a rigorous and evidence based approach. Additionally, the artifact pre- sents researchers and practitioners with a technical and algorithmic solution to detect topics of information being shared between the social video network nodes.

Fig. 4. Topic Intrusion task example. Example of topics generated by the artifact using LDA on a text corpus. The probability distribution of topics generated by the LDA algorithm that are associated with the text corpus are as follows: Topic 1: .382, Topic 4: .025, and Topic 10: .189 In this case, the highest probable intruder is Topic 4.

344 C. Daniel et al.

This research proposes an analytical process design that combines the importance of the corpus construction requirement with the unique attributes defined by the metadata of a social video network. Specifically, this research addresses the questions of “How do we construct a metadata-enhanced corpus from social video networks?” and “How might an instantiated artifact help define a class of IT solution for analyzing multimedia in online social networks?”. To accomplish this, a methodical, iterative eADR method in situ is deployed. The iterations within the Diagnosis, Design, Implementation, and Evolution eADR stages lead to the construction of the novel DSR artifact. The eADR methodology consistently informs the various iterations. Each iteration offers an opportunity to abstract, evaluate, reflect, and learn as the researchers and practitioners plan the next iteration. Overall, the methodology approaches used in this research contributes to the practical utility and rigorous quality of the DSR artifact.

This research will contribute to the existing body of knowledge for the IT solution class of social video network analysis with the implementation of an artifact that is designed to collect both the YouTube captions and the metadata using the YouTube Data API. Future research directions will apply the artifact in data analytic studies involving YouTube. A clear limitation of the current research is the exclusive focus of the artifact on YouTube video content. Moving beyond YouTube, we will adapt the artifact for use in other social network platforms where video content is prevalent. We plan to observe how the artifact can be used to analyze social video network behavior through utilities such as network search activities.

References

1. Abhari, A., Soraya, M.: Workload generation for YouTube. Multimed. Tools Appl. 46, 91– 118 (2009). https://doi.org/10.1007/s11042-009-0309-5

2. Ahmad, U., Zahid, A., Shoaib, M., AlAmri, A.: HarVis: an integrated social media content analysis framework for YouTube platform. Inf. Syst. 69, 25–39 (2017)

3. Arnold, C.W., Oh, A., Chen, S., Speier, W.: Evaluating topic model interpretability from a primary care physician perspective. Comput. Methods Program. Biomed. 24, 67–75 (2016)

4. Blei, D., Ng, A., Jordan, M.: Latent dirichlet allocation. J. Mach. Learn. Res. 3, 993–1022 (2003)

5. Borgatti, S.P., Everett, M.G., Johnson, J.C.: Analyzing Social Networks. SAGE, Los Angeles (2013)

6. Cha, M., Kwak, H., Rodriguez, P., Moon, S.: I tube, you tube, everybody tubes: analyzing the world’s largest user generated content video system. In: Proceedings of the 7th ACM SIGCOMM Conference on Internet Measurement, pp. 1–14. ACM, San Diego (2007)

7. Chang, J., Boyd-Graber, J., Gerrish, S., Wang, C., Blei, D.M.: Reading tea leaves: how humans interpret topic models. In: Proceedings of the 22nd International Conference on Neural Information Processing Systems, NIPS 2009, pp. 288–296. Proceedings of the 2009 Conference, Vancouver (2009)

8. Chatzopoulou, G., Sheng, C., Faloutsos, M.: A first step towards understanding popularity in YouTube. In: INFOCOMM IEEE Conference on Computer Communications Workshops, pp. 1–6. IEEE, San Diego (2010). https://doi.org/10.1109/infcomw.2010.5466701

Capturing User Generated Video Content in Online Social Networks 345

9. Chen, L.-C., Tesng, H.-H., Liao, I.-E.: Information and communication technology trend analysis using. In: Recent Researches in Applied Informatics: Proceedings of the 6th International Conference on Applied Informatics and Computing Theory (AICT 2015), pp. 158–166. WSEAS Press, Salerno (2015)

10. Cheng, X., Dale, C., Liu, J.: Statistics and social network of youtube videos. In: 16th International Workshop, IQQos 2008, pp. 229–238. IEEE, Enschede (2008)

11. Daniel, C., Dutta, K.: Automated generation of latent topics on emerging technologies from YouTube Video content. In: Proceedings of the 51st Hawaii International Conference on System Sciences 2018, pp. 1762–1770 (2018). http://hdl.handle.net/10125/50109

12. Figueiredo, F., Benevenuto, F., Almeida, J.: The tube over time: characterizing popularity growth of youtube videos. Retrieved from Fabricio Benevenuto, Computer Science Department, Federal University of Minas Gerais (2011). http://homepages.dcc.ufmg.br/ *fabricio/download/wsdm11.pdf

13. Gill, P., Arlitt, M., Li, Z., Mahanti, A.: Youtube traffic characterization: a view from the edge. In: Proceedings of the 7th ACM SIGCOMM Conference on Internet Measurement, pp. 15–28. ACM, Dan Diego (2007)

14. Gupta, H., Singh, S., Sinha, P.: Multimedia tool as a predictor for social media advertising- a YouTube way. Multimed. Tools Appl. 76(18), 18557–18568 (2017)

15. Harrison, D., Wilding, J., Bowman, A., Fuller, A., Nicholls, S.G., Pound, C.M., Sampson, M.: Using YouTube to disseminate effective vaccination pain treatment for babies. PLoS ONE 11(10), 1–10 (2016)

16. Hevner, A.R.: Design science research. In: Tucker, A., Topi, H. (eds.) Computing Handbook, 3rd edn, pp. 22-1–22-23. Chapman and Hall/CRC, New York (2014)

17. Hevner, A.R., March, S.T., Park, J., Ram, S.: Design science in information systems. MIS Q. 28(1), 75–105 (2004)

18. Hevner, A., Chatterjee, S.: Design research in information systems. Springer, New York (2010). https://doi.org/10.1007/978-1-4419-5653-8. Eds. by R. Sharda and S. Vob

19. Lange, P.G.: Publicly private and privately public: social networking on YouTube. J. Comput. Mediat. Commun. 13, 361–380 (2008)

20. Malik, H., Tian, Z.: A Framework for collecting YouTube meta-data. In: The 8th International Conference on Emerging Ubiquitous Systems and Pervasive Networks (EUSPN 2017), vol. 113, pp. 194–201. Procedia Computer Science (2017)

21. Miller, E.D.: Content analysis of select YouTube postings: comparisons of reactions to the sandy hook and aurora shootings and hurricant sandy. Cyberpsychol. Behav. Soc. Netw. 18 (11), 635–640 (2015)

22. Mullarkey, M.T., Hevner, A.R.: Entering action design research. In: Donnellan, B., Helfert, M., Kenneally, J., VanderMeer, D., Rothenberger, M., Winter, R. (eds.) DESRIST 2015. LNCS, vol. 9073, pp. 121–134. Springer, Cham (2015). https://doi.org/10.1007/978-3-319- 18714-3_8

23. Mullarkey, M.T., Hevner, A.R.: An elaborated action design research process model. Eur. J. Inf. Syst. (2018). https://doi.org/10.1080/0960085X.2018.1451811

24. Peffers, K., Tuunanen, T., Rothenberger, M., Chatterjee, S.: A design science research methodology for information systems research. J. MIS 24(3), 45–77 (2008)

25. Robins, G.: Doing Social Network Research. SAGE, Los Angeles (2015) 26. Santos, R., Rocha, B., Rezende, C.G., Loureiro, A.: Characterizing the YouTube

video-sharing community. Retrieved from Rodrygo Santos: Department of Computer Science, Federal University of Minas Gerais (2007). http://homepages.dcc.ufmg.br/ *rodrygo/wp-content/papercite-data/pdf/santos2007report.pdf

27. Sein, M.K., Henfridsson, O., Purao, S., Rossi, M., Lindgren, R.: Action design research. MIS Q. 35(1), 37–56 (2011)

346 C. Daniel et al.

28. Siersdorfer, S., Nejdl, W., Chelaru, S., San Pedro, J.: How useful are your comments?: analyzing and predicting youtube comments and comment ratings. In: Proceedings of the 19th International Conference on World Wide Web, WWW 2010, pp. 891–900. ACM, Raleigh (2010)

29. Smith, C., Allman, T., Crocker, S.: Reading between the lines: accessing information via YouTube’s automatic captioning. Online Learn. 21(1), 115–131 (2017)

30. Yoganarasimhan, H.: Impact of social network structure on content propagation: a study using YouTube data. Quant. Mark. Econ. 10, 111–150 (2009). https://doi.org/10.1007/ s11129-011-9105-4

31. Yuan, J., Zheng-Jun, Z., Zheng, Y.-T., Wang, M., Zhou, X., Chua, T.-S.: Utilizing related samples to enhance interactive concept-based video search. IEEE Trans. Multimed. 13(6), 1343–1355 (2011)

Capturing User Generated Video Content in Online Social Networks 347