Review of information systems research for media industry–recent advances, challenges, and introduction of information systems research in the media industry

Abstract

Global media reports clearly show a tremendous increase in spending on Information Technology (IT) and Information Systems (IS) in the media sector. Two main trends are currently visible. First, as stated in McKinsey & Company’s Global Media Report, consumers continuously shift from spending on traditional media products towards rapidly-increasing spending on digital services and media products – consumer patterns have rapidly changed. Second, as stated in Gartner’s reports about the media industry, spending on IT services in the media industry increases gradually, and as a whole, the media industry is the third-largest spender on IT, after banking/finance and manufacturing. Third, as reported by both authors in their works, several facets of the media industry have undergone extreme changes, including business models, declining revenue, content models, management, economics and public funding. The aim of this study is to investigate research work related to IS in the media industry, in particular in the management and economic areas. To conduct this study, we investigated a large corpus of studies that have been contributed to IS research within the Association for Information Systems (AIS) within the past decades. We utilised advanced text mining methods to identify contributions and thematic areas. Our results clearly indicate that there is a significant downward trend of research works dealing with media industry aspects. This trend was a surprise to us, as it contradicts the emergence of new digital technologies which became key drivers in the media industry after 2009. We conclude this article by giving research directions, illustrating the opportunities and importance of investigating media industries within the context of IS, and introducing the research field of Information Systems in the eMedia Industry.

Appendix: topic classification

The full (and simplified) analysis process is illustrated in Fig. 5, and involved the principal steps of pre-processing, training and scoring, and post-processing and is based on prior work (Lugmayr and Zotto 2013) and has been applied in e.g. financial news sentiment analysis (see (Lugmayr and Gossen 2012) and (Lugmayr 2013f)). The main tool for text processing was R. The approach described below is based on various literature resources as e.g. (Blei et al. 2003a), (LDA n.d.), (Jockers 2014), and e.g. (A gentle introduction to topic modeling using R n.d.)

Fig. 5

Schematic and Simplified Overview of the Processing Steps

Pre-processing

The first step in the analysis process was to perform pre-processing on the corpus C, which included stop-words processing, lemmatization, parts-of-speech processing, and lexicalization to obtain a cleaned corpus.

Lexicalisation of all documents and creation of a vocabulary list

The given corpus C consists of M documents D _1 .. M and associated metadata F _1 .. M (T, Y, A, C, S, G) such as title, year, author, category, source, and abstract. Each document D _j consists of N _j unique terms, where the total sum of N terms can be calculated through $N={\displaystyle {\sum}_{j=1}^M{N}_j}$. The total term list of the corpus is given through V _1 .. N, and the number of a particular term V _n in document D _j is annotated as ${N}_{V_{n^j}}$.

Calculation of the identity matrix in all documents

As described in various R manuals and in (Jockers 2014), we began by the calculation of the term frequency document matrix ${TF}_{j,t}=\frac{N_{V_{t^j}}}{{\displaystyle {\sum}_{f=1}^{N_j}{N}_{V_f}}}$ to calculate the number of times a term is in a one document. To gain insights about into the importance of single terms, we calculate the logarithmic scaled inverse fraction (or inverse document frequency) ${IDF}_{j,t}= \log \frac{M}{1+\left|\left\{d\in D:{V}_t\in d\right\}\right|}$, resulting into the term-frequency/inverse document frequency through TFIDF _j , t  = TF _j , t  × IDF _j , t , in short annotated as W.

Training and scoring

To train topics and perform scoring, we utilized the Gibbs Sampler LDA (Blei et al. 2003a) (LDA n.d.). a and β are the pre-known (assumed) parameters of the Dirichlet distribution, as e.g. described in (Wikipedia n.d.). K is the number of assumed topics in the corpus C, and Z _n , j contains the topic of the nth word in document D _j . ${N}_{V_{n^j}}$ is the number of terms in one particular document. Topic(s) including topic word(s) were obtained through $P\left(W,Z,\theta, \varphi, a,\beta \right)={\displaystyle {\prod}_{k=1}^KP\left({\varphi}_k;\beta \right)}{\displaystyle {\prod}_{j=1}^DP\left({\theta}_j;a\right){\displaystyle {\prod}_{t=1}^WP\left({Z}_{j,t}\left|{\theta}_j\right.\right)P\left({W}_{j,t}\left|\varphi {z}_{j,t}\right.\right)}}$.

Post-processing

Post-processing included parameter fine-tuning (e.g. appropriate selection for pre-given parameters of the Dirichlet distribution and selection of topics), followed by some statistical processing of topics and document metadata, as well as the visualization of results. In addition, we performed a keyword in context analysis to understand the topics and argumentation of each contribution on a semantic level.