Data-driven operations and supply chain management: established research clusters from 2000 to early 2020

ABSTRACT Despite the long-recognised importance of data-driven operations and supply chain management (OSCM) scholarship and practice, and the impressive development of big data analytics (BDA), research finds that firms struggle with BDA adoption, which suggests the existence of gaps in the literature. Therefore, we conduct this systematic literature review of journal articles on data-driven OSCM from 2000 to early 2020 to ascertain established research clusters and literature lacunae. Using co-citation analysis software and double-checking the results with factor analysis and multidimensional-scaling-based k-means clustering, we find six clusters of studies on data-driven OSCM, whose primary topics are identified by keyword co-occurrence analysis. Five of these clusters relate directly to manufacturing, which, in line with the existing literature, indicates the crucial role of production in OSCM. We highlight the evolution of these research clusters and propose how the literature on data-driven OSCM can support BDA in OSCM. We synthesise what has been studied in the literature as points of reference for practitioners and researchers and identify what necessitates further exploration. In addition to the insights contributed to the literature, our study is amongst the first efforts to deploy multiple clustering techniques to undertake a rigorous data-driven systematic literature review (SLR) of data-driven OSCM.


Introduction
Through statistics, optimisation (Tiwari, Wee, and Daryanto 2018) and other supply chain analytics tools and techniques (Chae, Olson, and Sheu 2014), data have long been exploited in operations and supply chain management (OSCM), where production and logistics play key roles. Indeed, since 2000, the cusp of the new millennium and a period of vast enterprise resource planning (ERP) adoption where companies sought solutions to the Y2K (Year 2000) problem (Irani, Themistocleous, and O'Keefe 2001), data-driven decision-making has received increasing attention from production researchers (Kuo and Kusiak 2019) as such technological advances increase the possibility to collect and, later on, leverage data. Empirical research has shown positive correlation between the use of data-based tools and OSCM efficacy in multiple countries and industries (see Chae, Olson, and Sheu 2014;Chavez et al. 2017;Song et al. 2018).
In addition to empirical findings, the literature on OSCM has discussed frameworks and models to further develop data-driven supply chains (SCs). For instance, in an attempt to allay overreliance on expert yet CONTACT Duy Tan  possibly subjective judgment, Cheng et al. (2020) proposed a data-driven technique based on support vector machine (SVM) for supplier evaluation and used simulation based on a big firm's dataset for model assessment. Aimed at dealing with external uncertainty, Medina-González et al.'s (2020) multi-objective datadriven model exploits machine learning (ML), robust optimisation and meta-multiparametric programming to handle the stochasticity related to raw materials, demand, and environmental and social impact parameters to optimally manage a bio-energy production SC. In response to intermittent demand, shortening product lifecycle, long lead time, etc., in the semiconductor industry (Uzsoy, Fowler, and Mönch 2018), Fu and Chien (2019) proposed a UNISON-based analytics model, which integrates ML and adopts temporal aggregationdisaggregation mechanisms for demand forecasting, and validated their framework with a global electronics distributor.
Data from end-to-end supply chain management (SCM) processes are now exponentially increasing, heightening the need to analyse big data (BD) ). Thus, scholars and companies are striving to develop big data analytics (BDA) capabilities (Tiwari, Wee, and Daryanto 2018). Nevertheless, despite the growth in academic publications on BDA in OSCM (e.g. Dubey et al. 2019;Meriton et al. 2020;Maheshwari, Gautam, and Jaggi 2021), many firms are struggling with BDA adoption Viet et al. 2021). Only a few companies have succeeded in BDA in SCM (Wang et al. 2016a;Meriton et al. 2020) whilst others have not been able to make judicious use of the BD available (Lamba and Singh 2017;Mishra et al. 2018). This calls for more research in BDA adoption.  used two constructs, namely IT resources and data assimilation, to operationalise datadriven capabilities. In particular, the former include IT infrastructure and database, whereas the latter denotes if data are utilised for order management, forecasting or planning. This accords with empirical SCM research predicated on organisational information processing theory (OIPT) (Williams et al. 2013;Chen, Preston, and Swink 2015;Srinivasan and Swink 2018) where information processing capabilities (IPCs) moderate/mediate the correlation between organisational performance/competitiveness and SC visibility enabled by the data gained from IT utilisation. Williams et al. (2013) and Srinivasan and Swink (2018) broadly defined IPC as the ability to process and leverage data for a specific purpose whilst Chen, Preston, and Swink (2015) particularly modelled BDA as the unique IPC mediating the relationship conceptualised. In effect, several studies defined or referred to data-driven SCs as those that leverage BDA to improve SC competitiveness , sustainability  or performance (Yu et al. 2019;Gawankar, Gunasekaran, and Kamble 2020). This might indicate a shift of focus to BDA in research on data-driven OSCM. In fact, BDA has recently grown fast in SCM (Chehbi-Gamoura et al. 2020).
Recognising those issues and paradigm shifts, we conduct a systematic literature review (hereinafter referred to as SLR) to identify the knowledge structure of research on data-driven OSCM since 2000. The aims include determining clusters of studies on data-driven OSCM and ascertaining subfields and lacunae in emergent research topics, e.g. ML and BDA, so that directions and avenues for further research could be elicited. In other words, the aim of this SLR is to answer the following research questions: (i) What is the knowledge structure of research on datadriven OSCM from 2000 to early 2020 (published from 2000 to 2019 or accepted for publication before 2020)?
(ii) Are there any topics on data-driven OSCM which have recently been heeded by both scholars and practitioners or integrated into an established subfield?
Our SLR is in line with the need for systematic, transparent and rigorous synthesis of a comprehensive body of literature to provide a high-quality evidence base to inform practice, policy-making and research, and avoid loss of knowledge from earlier studies (Tranfield, Denyer, and Smart 2003;Rousseau, Manning, and Denyer 2008;Meriton et al. 2020). By reviewing the germane studies accumulated in the first two decades of this millennium, which saw the rise of data-driven decision-making and BDA, our work can offer interesting insight into how research on data-driven OSCM developed longitudinally, whereby an overview of the current theoretical foundations can be provided and research avenues identified.
This SLR will contribute directly to the advancement of production research because SCM has been amongst the major areas in production research literature (Kuo and Kusiak 2019;Silva, Ablanedo-Rosas, and Rossetto 2019). Indeed, according to Vrijhoef and Koskela (2000), SCM derived from and thrived in the manufacturing sector. Meanwhile, before incorporating other organisational functions, e.g. procurement and logistics, in its scope, OM (operations management) 'referred primarily to manufacturing production' (Bayraktar et al. 2007).
According to Uddin, Khan, and Baur (2015), the knowledge structure of a study domain, which can be ascertained by statistical and visualisation techniques, illustrates the evolution of its subfields. It is generally understood that the knowledge structure of a research area, based on which studies are developed, can be represented by the bibliographies of its publications over a given period as a means to identify its 'building blocks' (Samiee and Chabowski 2012). Therefore, in our paper, the knowledge structure of data-driven OSCM scholarship is regarded as a network of pertinent journal articles, whose clustering is based on co-citation.
We define data-driven or data-based OSCM broadly as the use of data for OSCM practices, e.g. forecasting, tracking and scheduling, in accordance with Williams et al. (2013), Srinivasan and Swink (2018) and . The selected papers from the Web of Science (WOS) and other databases, which were published after 1999 and accepted for publication before 2020, are then analysed, using co-citation analysis to identify research clusters. Factor analysis (FA) and multidimensional scaling (MDS) are carried out to validate the clustering results. Next, we read the papers retained from the analyses and use keyword co-occurrence analysis to determine each cluster's theme.
There have been SLRs on OSCM using data-driven approaches (Xu et al. 2018) or on a subfield of datadriven OSCM , but so far we have seen few SLRs on data-driven OSCM, which employ such data-driven approaches as co-citation analysis. Therefore, our study will be amongst the first attempts to apply this data-driven methodology for a literature review on data-driven OSCM. We emphasise that our SLR focusses on research into data-driven OSCM practices and applications, not on fundamental research which advances data-driven methodology without evident connection to OSCM practices. For instance, if a paper uses BDA to illuminate the characteristics of multinationals' SCs rather than address an OSCM issue, e.g. supply management and inventory planning (these OSCM issues will be enumerated in 2.2), we do not include that article in our sample. Likewise, an optimisation study might be excluded unless data play a role in its model formulation and a practical issue, e.g. demand uncertainty, is explicitly targeted. Conversely, our sample incorporates qualitative papers which, for example, discuss the benefits or performance of data-driven OSCM or the management and utilisation of data for OSCM. This paper is divided into five sections. Following this introduction, the next section presents our paper's Methodology, whereas the third one illustrates the clustering results. Implications for research and practice based on our findings are provided in section IV, and Conclusion is the last section, which also discusses this study's contributions and limitations.

Methodology
This paper is framed in line with Durach, Kembro, and Wieland's six-step SLR protocol (2017), which specifically addresses the characteristics of the SCM context and was built on the integration of the most cited and applied SLR guidelines, including Tranfield, Denyer, and Smart's (2003).
(i) First, we define the research questions and justify the timeliness, relevance and expected contribution of our SLR (see Introduction). (ii) Then, we determine the inclusion and exclusion criteria. (iii) Next, we retrieve a sample of potentially pertinent literature and specify the search procedure, databases and keywords. (iv) By applying the predefined inclusion and exclusion criteria, we choose the relevant papers. (v) Given the selected articles, we synthesise the literature, using co-citation analysis, FA, MDS and keyword co-occurrence analysis.
(vi) Finally, we present a descriptive summary of the selected publications and report the thematic findings.
In comparison to Tranfield, Denyer, and Smart (2003) and Rousseau, Manning, and Denyer (2008), Durach, Kembro, and Wieland (2017) suggested one additional step which requires specifying the search procedure, databases and keywords, including possible synonyms. We believe that the recency of their SLR paradigm, its focus on SCM scholarship and its heightened specificity will enhance our SLR transparency and replicability. Similar procedures can be found in other literature reviews, e.g. Seyedghorban et al. (2020), Badi and Murtagh (2019), Martins and Pato (2019) and Rebs, Brandenburg, and Seuring (2019). Nonetheless, each step can be modified to support the research. For instance, in the sampling step, some authors utilise cross-referencing or backward snowball search to find additional germane papers in the bibliographies of the selected papers (Hosseini, Ivanov, and Dolgui 2019;Rebs, Brandenburg, and Seuring 2019;Kamble, Gunasekaran, and Gawankar 2020). In forward snowball search, authors seek relevant studies that cite the selected articles (Karttunen 2018;Martins and Pato 2019). Another example is the use of bibliometric software to support co-citation-based clustering analysis (Feng, Zhu, and Lai 2017;Rebs, Brandenburg, and Seuring 2019;Seyedghorban et al. 2020). Since the procedure Durach, Kembro, and Wieland (2017) discussed is commonly-adopted, we frame our research in line with their guidelines to reinforce our SLR rigour. However, the validity and originality of our review are further enhanced by the deployment of multiple cocitation analysis methods in step (v) for data analysis triangulation, which Durach, Kembro, and Wieland (2017) did not explicitly point out. We elaborate on the applied procedure in the next subsections, beginning with step 2.

Inclusion and exclusion criteria
We adopt the inclusion criteria utilised in Glock et al.'s SLR (2019), which are in conformity with prior studies as follows: • Language: English. • Time span: from 2000 to 2019. Papers with publication time after 2019 which were accepted for publication and became available online before 2020 are also included in the sample. • Article type: Academic (peer-reviewed) journal article. In accordance with prior literature reviews (e.g. He et al. 2018;Xu et al. 2018), we only include research or review papers published in peer-reviewed journals in our study because such publications are considered verified knowledge (Ramos-Rodríguez and Ruíz-Navarro 2004). • At least one search keyword is mentioned in the title, abstract or keywords of the paper (cf. Badi and Murtagh 2019). Words with similar or related meaning are accepted, e.g. manufacturing and production, and purchasing and procurement.
Regarding the exclusion criteria, we eliminate the papers which include the keyword(s) but do not pertain directly to OSCM, e.g. research on trading and price prediction of financial asset. Since we follow the broad conceptualisation of data-driven OSCM (cf. Irfan and Wang 2019), the research papers selected can focus on data assimilation or data management resources (e.g. IT) but must explicitly mention the purpose(s) for which the data (recorded internally, obtained from external sources or shared with partners) are used in practice. In other words, the data in the selected articles must be employed to address a practical issue, e.g. supplier selection and inventory management, rather than test a conceptual model unless it is about the relationship between data-driven OSCM and organisational performance. We peruse the abstract and research question(s) of each paper to assess its relevance to our SLR focus.
The graphical summary for the inclusion and exclusion criteria is provided in Figure 1, which will be further elaborated in the next subsections.

Databases and keywords
To retrieve germane literature, we primarily use the WOS, which provides access to over '100 million references from 33,000 journals' (Martins and Pato 2019). This platform has been used by many scholars for literature review, e.g. Kamble, Gunasekaran, and Gawankar (2020), Seyedghorban et al. (2020) and Rebs, Brandenburg, and Seuring (2019). Nonetheless, other researchers, e.g. Badi and Murtagh (2019) and Martins and Pato (2019), also accessed other databases to ensure holistic literature retrieval. Indeed, in Liu et al.'s literature review (2017), other databases contained a nontrivial number of unique publications that were not included on the WOS. Therefore, we add Taylor & Francis, Emerald, INFORMS, ProQuest, Sage, Science Direct, Springer and Wiley to our search databases, and use the WOS template to format data.
With regard to the keywords, we use 'data-driven' or 'data-based' with those that have been utilised in previous literature reviews or deemed to be important OSCM themes (Table 1). Although these keywords cannot cover all OSCM topics, we believe that such generic keywords as SC and operations help find publications on uncommon subfields. To avoid overlooking important papers, we adopt cross-referencing to identify commonly-cited articles in our sample. Specifically, if a peer-reviewed journal paper is cited at least four times by our selected studies and satisfies the inclusion criteria aforesaid, we add that publication to our sample. The four-citation threshold was proposed by Feng, Zhu, and Lai (2017).

Literature synthesis
To determine clusters of research in the selected publications, we deploy co-citation analysis, which was put forth by Small (1973). When two papers are cited together in another study, we regard that as one co-citation, whose frequency indicates the similarity of the papers' research topics (Small 1973). Therefore, co-citation analysis is primarily to identify the main research topics in the literature (Feng, Zhu, and Lai 2017). We follow the four-citation threshold recommended by Feng, Zhu, and Lai (2017) to ensure adequacy and tractability of the citation data for analysis (Zupic and Čater 2015). To analyse the co-citation data, we use VOSviewer open-source software (version 1.6.10), which was developed by van Eck and Waltman (2010) and has been employed in co-citation-based literature reviews, e.g. Wang et al. (2016b) and Zhao, Zhang, and Kwon (2018).
As  explained, VOSviewer determines the locations of n items on a map based on their co-citation similarity and then clusters them based on their distances . Papers assigned to the same cluster by VOSviewer are those that are often co-cited. In their comparison study,  found that VOSviewer visualisation is more effective than that of MDS, which, by minimising a stress function, projects items into a low-dimensional space such that the distance between any two items can best reflect their dissimilarity. To check robustness of VOSviewer results, we utilise MDS (followed by k-means clustering) in scikitlearn (Pedregosa et al. 2011) and FA in STATA 15.1 for data analysis triangulation. With FA, we can find the high-order level factors that capture most of the correlation space in the co-citation matrix. Items captured by the same factor are highly-correlated with each other and associated with the same latent variable, or in other words, an overarching topic. These analyses were performed by Zhao, Zhang, and Kwon (2018) and Wang et al. (2016b), but our paper is amongst the first to use all three tools for a data-driven SLR of data-driven OSCM.
We use Bibexcel (Persson, Danell, and Schneider 2009) to adapt the input data for analysis in different software (cf. Feng, Zhu, and Lai 2017;Xu et al. 2018;Zhao, Zhang, and Kwon 2018;Hosseini, Ivanov, and Dolgui 2019). We then apply the three clustering techniques to the pre-processed data and retain the papers that are consistently clustered by all the methods. Next, we read the papers in each cluster to determine its theme with the support of VOSviewer keyword co-occurrence analysis (cf. Zupic and Čater 2015;Ikeziri et al. 2019). In addition to the author-assigned and WOS-provided keywords in the metadata, we employ Wang, Liu, and Wang's PageRank-based algorithm (2007) to extract important tokens in each paper's abstract and title. If these tokens are not included in the keyword sections of that article, we add them to the database before carrying out VOSviewer keyword co-occurrence analysis. With this, we can take advantage of the abstract content in the metadata and make additional contributions to our rigorous analysis. Details on our keyword extraction are given in the online companion.

Result report
The clustering results are reported for the papers published after 1999 and accepted for publication before 2020. Given that the WOS metadata give credit to first authors only, we focus our analysis on the papers rather than their researchers. As citation is commonly-used to measure a study's impact (Feng, Zhu, and Lai 2017), we use four indices, namely, global citation index, WOS citation index, in-sample citation index and PageRank, to identify influential papers in our sample. We utilise Google Scholar citation index as a proxy for our global citation index since Google Scholar covers several academic platforms including the WOS (Feng, Zhu, and Lai 2017). The in-sample citation index shows how many papers in our sample cited a given article. To obtain a comprehensive picture of research impact, we use PageRank which indicates the degree to which a paper is cited by highly-cited studies (Xu et al. 2018). Descriptive details of the selected papers can be found in the supplemental material.

Results and discussion
The preceding sections describe the first three steps in Durach, Kembro, and Wieland's procedure (2017). We discuss the remaining steps, namely literature retrieval and selection, literature summary and thematic report, in the next subsections.

Literature retrieval and selection
By setting the English language and academic/scholarly journal as search parameters, and using the aforementioned keywords, we retrieve a total of 2341 search results from all the databases. Then, we apply the inclusion and exclusion criteria and remove duplications to obtain a sample of 398 pertinent papers, 365 of which were published from 2011 onwards. This accords with prior findings that there has been mounting growth of interest in BD from both scholars and practitioners since 2011 (Nguyen et al. 2018;Tiwari, Wee, and Daryanto 2018).
As we focus on the WOS, papers which were found in both WOS and another database are credited to the WOS only (26 September 2020). In other words, the figures reported for each non-WOS database after duplication check in Figure 2 (those remarked with * ) refer to the number of unique publications retrieved from that platform only. As can be inferred from Figure 2, 94.22% of the shortlisted articles are included on the WOS. This ratio is 95.07% for publications after 2010 and 84.85% of those published before. The difference between these proportions is marginally significant at the 5% significance level, implying that the WOS publication coverage for datadriven OSCM can be deemed consistent between the two periods.
With an initially shortlisted sample of 398 papers, we conduct cross-referencing to identify highly-cited and relevant references in these publications. Journal papers with at least four in-sample citations that related to datadriven OSCM are added to the sample and we repeat the process until no new highly-cited reference can be found. In line with the conceptualisation of data-driven OSCM discussed, the papers considered relevant can focus on data assimilation or IT/data infrastructure but must explicitly mention how data are utilised for OSCM. After six iterations, we find another 182 relevant studies. In total, 580 papers are deemed relevant to data-driven OSCM and their bibliographies constitute the metadata for our clustering analysis. We note that this sample update has no statistically significant impact on the  The decrease in the WOS coverage proportion of papers from 2018 to 2019 in Figure 3 was because some studies had been accepted for publication and made available on the publisher's database but had not yet been updated on the WOS. Another reason is that with the growth in attention to data-driven OSCM from both scholars and practitioners, relevant studies are published in various journals, some of which are not included on the WOS. One interesting insight from Figure 3 is the high correlation between the research volume of data-driven OSCM and Google Trends for 'Big Data' and 'Industry 4.0.' Indeed, the observed upsurge in Google user search for 'Big Data' after 2010 was followed by the increase in popularity of 'Industry 4.0' search queries and the growth in studies on data-driven OSCM some years later. This partially reflects the attention of OSCM scholarship to practitioners' needs.
From the 580 journal papers selected for literature synthesis, those which were cited four times or more by other selected studies are considered in co-citation analysis as Feng, Zhu, and Lai (2017) recommended. The sample in clustering then includes 196 journal articles, but the metadata for PageRank computation and co-citation database are from the 580-paper pool. Summarised details of these papers are provided in the online appendix.
The next subsection synthesises those publications to attain an overall picture of the data-driven OSCM literature before cluster analysis is performed.

Literature summary
Overall, the 580 papers selected were published in 229 peer-reviewed journals, amongst which the journals with the largest numbers of selected publications are presented in Table 2. Of a particular note is that there is no huge difference between the numbers of papers selected from these publication outlets. In fact, no journal accounts for a majority of research articles on data-driven OSCM in our sample. This is consistent with the broadening interdisciplinary and integrative scope of OM (Manikas et al. 2020) and SCM (Swanson et al. 2018). In addition to journals specialising in transportation and production, two subfields of OSCM, we can see, in Table 2, journals in operations research and engineering, which also relate to OSCM. The diverse yet mostly high impact factors of these journals signify that there are many high-quality studies on data-driven OSCM.
Looking at some of the most highly-cited papers in our sample (see Table 3), we can observe a similar pattern between Google Scholar (proxy for global citation) and WOS citation indices, which is not surprising since Google Scholar citation index includes that of the WOS. Indeed, the correlation between global (Google Scholar)  and WOS citation indices is 0.9320 for the 580 papers selected. However, the correlation between in-sample citation and Google Scholar citation (WOS citation) is only 0.5689 (0.6021). As shown in Table 3, only one paper belongs to both groups of top-ten globally and in-sample cited research. According to Feng, Zhu, and Lai (2017), this disparity can be explicated by the varying degree of attention from scholars in different fields. This means that some globally highly-cited research is less heeded or deemed less pertinent by academics in data-driven OSCM and vice versa.
With respect to Google Scholar citation, the top-ten papers studied a variety of OSCM subfields, e.g. OSCM overall (Fosso Wamba et al. 2015), inventory and sales (Elmaghraby and Keskinocak 2003;Ben-Tal et al. 2004), supply management (Ho, Xu, and Dey 2010), healthcare operations (Raghupathi and Raghupathi 2014), production (Xu 2012) and demand/transportation forecasting (Hippert, Pedreira, and Souza 2001;Wu, Ho, and Lee 2004;Zhao and Magoulès 2012;Lv et al. 2015). Nonetheless, most of them are literature reviews (Hippert, Pedreira, and Souza 2001;Elmaghraby and Keskinocak 2003;Ho, Xu, and Dey 2010;Zhao and Magoulès 2012;Raghupathi and Raghupathi 2014). Modelling-based studies include those of Lv et al. (2015), Ben-Tal et al. (2004) and Wu, Ho, and Lee (2004), whereas Fosso Wamba et al.'s (2015) and Xu's (2012) articles are empirical and conceptual, respectively. It is interesting that although modelling (normative, descriptive and predictive) dominates research on data-driven OSCM in general, literature reviews account for the largest proportion of the ten globally most cited papers in our sample. Explicably, literature reviews give an overview of a certain research topic and thus can be commonly-cited.
As indicated in Table 3, none of the top-ten papers in terms of Google Scholar citation is amongst the ten most influential papers according to PageRank score computed on our selected research on data-driven OSCM. Out of the top in-sample cited research, only two belong to the group of ten most influential articles as per this index. The correlation between in-sample citation and PageRank index in our sample is 0.5886 whilst the figure is 0.5689 for global citation and PageRank index. This result is expected as highly-cited research is not necessarily influential because PageRank-based research impact is determined by the extent to which a study is cited by high-impact papers (Brin and Page 1998;Xu et al. 2018). Hence, the nominal value of citations may not fully reflect a paper's influence as indicated by PageRank index.
The next subsection will discuss the knowledge structure or, in other words, research clusters on data-driven OSCM studied since 2000.

Thematic report
We load the 580-publication metadata used for PageRank computation into VOSviewer, but only relevant references with at least four in-sample citations are considered. We then identify seven clusters of research on data-driven OSCM as depicted in Figure 4.
We check the robustness of VOSviewer results by performing FA in STATA and MDS in scikit-learn for data analysis triangulation.
For FA, we utilise the eigenvalues and factor rotation (Yong and Pearce 2013) to select the high-order level factors that capture most correlation in the co-citation space. From the bibliographies of the 580 selected studies, we compile co-citation matrix C |N|×|M| , where N and M are respectively the references and the 196 relevant journal publications cited by at least four articles in our 580-paper sample. C ij indicates the number of studies in our sample that cite both i and j. C ii = 1. Then, we run iterated principal factor analysis of C and factor rotation. Each factor here can be interpreted as a cluster of closely-related studies and each item in a factor is a paper in our retained sample. There are 26 factors with eigenvalues greater than 1, which altogether explain 100% of the variation, but only nine of them have more than three items with loadings greater than or equal to 0.7 each. To test the significance of the results, we run structural equation modelling to perform convergent validity analysis (Sethi and King 1994) and discriminant validity analysis (Fornell and Larcker 1981).
As illustrated in Table 4, the Cronbach's alpha of each factor exceeds the 0.7 threshold, indicating consistency amongst the items included therein (Dunn, Baguley, and Brunsden 2014). Whilst the correlations between factors are below the recommended 0.85 threshold , which indicates good discriminant validity, the    Average Variance Extracted (AVE) of each factor being higher than its squared correlations with other factors is another discriminant validity indicator (Song et al. 2018).
All factor loadings are of acceptable magnitude (greater than 0.7) and statistically significant, implying good convergent validity (Sethi and King 1994). The Composite Reliability and AVE are respectively above the thresholds of 0.7 and 0.5 (Fornell and Larcker 1981;Yu et al. 2018), further confirming the convergent validity. Thus, these nine factors can be considered robust. Table 5 demonstrates that most factors identified in FA fit entirely in the clusters appearing in VOSviewer result. Although the number of clusters/factors differs, the membership stays consistent. With these results, six clusters and 120 papers are retained.
According to , the similarity index used in MDS stress function can be the correlation between two items or their cosine. We run MDS with these similarity indices in scikit-learn (Pedregosa et al. 2011) and select the results of lowest dimensionality whose stress level is below the recommended 0.1 threshold (Kruskal 1964;Zhao, Zhang, and Kwon 2018). Afterwards, we perform the modified (Bagirov 2008) and fuzzy (Khan et al. 2020) k-means algorithms to cluster the cited papers. Despite changes in the number of clusters identified, the 120 papers retained in Table 5 fit neatly in the k-means clusters appearing in MDS outputs. Our implementation and detailed results of MDS and k-means clustering are provided in section A of the e-companion.
Overall, out of the seven VOSviewer-assigned clusters, only six survive all the three clustering methods and we have a final sample of 120 papers for thematic interpretation.
On a separate note, we notice that amongst the 196 papers, which are cited by at least four papers in this pool and thus selected for cluster analysis, around one third (63) have four in-sample citations. The figure is similar for the papers published since 2015 (12/40). However, amongst the 120 articles commonly retained by the three analysis techniques and hence included in thematic reporting, one fifth (24) are cited by four other in-sample studies. For the publications since 2015, the ratio is 5/24. We can see that papers with four in-sample citations are less likely to survive all the three clustering techniques deployed and that recent research, i.e. published since 2015, can also receive more than four citations. From these statistics, we believe that the threshold of four insample citations recommended by Feng, Zhu, and Lai (2017) is suitable because decreasing that threshold may not necessarily increase the number of studies retained for thematic reporting whilst raising it will reduce the contents covered.
In the next step, we read the 120 peer-reviewed journal articles retained from the analyses and use keyword co-occurrences to identify the themes. The lower right corner of each figure in the next subsections shows the average publication year of the papers mentioning the coloured keyword. It is to note that there are many papers in Clusters 1 and 2, so we will mainly mention the highly-cited articles on Google Scholar or those written by authors with multiple publications in that cluster as they are considered influential. Also, we will mention the papers whose topics we believe should be further explored though less common in their clusters.

Big data (data analytics) in OSCM
This cluster corresponds to Factor 2 with 25 papers retained. At Figure 5's centre are 'big data,' 'analytics' and 'supply chain,' the keywords co-occurring most often with other keywords in this cluster. Indeed, most papers therein discussed the application or benefits of BD/analytics in OSCM via literature reviews (Wang et al. 2016a;Hazen et al. 2018 2014) used a brief case study on jet engine remanufacturing to illuminate how the data quality problem could be addressed to enhance data-driven SCM. In another example, utilising data from Twitter related to companies in manufacturing, logistics, news and IT, Chae (2015) demonstrated his proposed analytical framework and the value of social media data in SCM. Schoenherr and Speier-Pero (2015) interviewed experts from several firms, including professional service providers (e.g. consultancies), and found that data analytics skills are desired for SCM professionals.
Some papers discussed the application or benefits of BDA in a specific OSCM subfield, e.g. Fosso Wamba et al. (2015) presented a case study on data-driven operations of the New South Wales State Emergency Service. Chong et al. (2016) and Cui et al. (2018) developed predictive models for retailing. Opresnik and Taisch (2015) proposed a conceptual framework for BD strategy in servitisation, a hybrid model of service provision and manufacturing. Nonetheless, the sector discussed most often by this cluster's papers is manufacturing. To leverage BD for fault detection in manufacturing, Kumar et al. (2016) proposed a MapReduce framework, which can handle imbalanced data. Carrying out empirical research on the manufacturing industry in India, Dutta and Bose (2015) and Dubey et al. (2016) investigated the influence of BD via a case study and survey, respectively. In another Asian country's context (China), Tan et al. (2015) and Zhao et al. (2017) developed a BD-based model and illustrated it in a case study, whereas Yu et al. (2018) used survey data from manufacturing firms to validate the relationship between data-driven SCs and firm performance. In addition to the empirical research mentioned in the preceding paragraph, other examples related to the manufacturing sector are the modelling-based papers of Huang and Van Mieghem (2014) and Wu et al. (2017). This reflects a finding in the extant literature that the manufacturing sector accounts for a large proportion of OSCM research.
Based on the publication time ( Figure 5's lower right corner), we can see that this is a recent cluster of research given the recent rise of BD research (Fosso Wamba et al. 2015;Nguyen et al. 2018;Tiwari, Wee, and Daryanto 2018;Chehbi-Gamoura et al. 2020), but three of the topten in-sample cited papers in our sample belong to this cluster and one of them also appears in the top-ten globally cited papers (Table 3). This implies the importance of this research cluster on data-driven OSCM. Hazen et al. (2014) are amongst the highly-cited authors with multiple publications in this cluster. Although modelling is a widely-used methodology in data-driven OSCM, empirical research dominates Cluster 1. Since empirical studies constitute a critical part of OM research (Gattiker and Parente 2007), the formation of a cluster of highly-cocited research dominated by this method can be expected. An interesting insight is that we did not use 'big data,' 'data analytics' or 'data science' as search keywords, but their appearance in our keyword co-occurrence analysis heralds the upward trend of (B)DA in data-driven OSCM. Figure 6 shows that the central keywords in this cluster are 'traffic' and 'model' or, more specifically, traffic-flow forecasting models, which are deemed, by all research in this cluster, vital in (operating) advanced/intelligent  transportation systems (ITS). This cluster's primary methodology is modelling, which includes both traditional (statistical) methods, e.g. auto-regressive integrated moving average (ARIMA) and econometric regression (Stathopoulos and Karlaftis 2003;Min and Wynter 2011), and ML schemes, e.g. k-nearest neighbours (Smith and Oswald 2003), SVM (Wu, Ho, and Lee 2004) and artificial neural networks (ANN) (Dia 2001;Vlahogianni, Karlaftis, and Golias 2005). Some authors, e.g. Chen et al. (2001) and Chan et al. (2012), combined both approaches in their models, but there are overall twice as many papers adopting ML algorithms as conventional method-based articles in this cluster. Nonetheless, the publication time shows the parallel development of these two research streams. Overviews and comparisons of both statistical and ML techniques can be found in the reviews of Vlahogianni, Golias, and Karlaftis (2004) and Karlaftis and Vlahogianni (2011), whereas Zhang, Wang, et al.'s survey (2011c) focussed exclusively on ML. Overall, there was an increase in complexity of the models studied, but Karlaftis and Vlahogianni (2011) claimed that simple and complex models can produce equally good results.

Transportation and traffic flow prediction
Revisiting Table 3, we find four articles in Cluster 2 amongst the top-ten in-sample cited papers, which partially reflects this cluster's popularity. Stathopoulos and Karlaftis (2003) and Vlahogianni, Karlaftis, and Golias (2005) are examples of authors with multiple publications in this cluster. Overall, this is an established cluster that contributes to the methodological landscape of data-driven OSCM, where modelling dominates.
Nearly 85% of the 46 papers in this cluster mentioned their data sources and around three quarters of them collected data from western economies, including Australia. Even in Asia, the research data were obtained from the countries with developed infrastructure, e.g. Taiwan, Singapore and South Korea. This is not surprising given the infrastructure necessitated for ITS, where data can be recorded in large volumes by detector stations (Dia 2001;Chan et al. 2012) or loop detectors (Stathopoulos and Karlaftis 2003;Min and Wynter 2011). With noisy or corrupt data recorded in such systems, Chen et al. (2001) and Quek, Pasquier, and Lim (2006) investigated the impact of missing data and noise tolerance, respectively, of ANNbased traffic-flow prediction, whereas others developed models to address this issue, which include data imputation (Qu et al. 2009;Chen et al. 2012) or denoising (Jiang and Adeli 2004). With more data-recording devices in ITS, voluminous traffic data have been collected (Zhang et al. 2011c), but we note that the retained papers did not explicitly discuss BD.

Demand forecasting
This VOSviewer-assigned cluster comprises two FA factors, one related to water demand forecasting (Factor 4) and the other to energy consumption prediction (Factor 9). We can infer from the three keywords at Figure 7's centre that this cluster's topic is 'demand forecasting.' Like Cluster 2, over 90% of this cluster's research deployed predictive modelling. The publication time (Figure 7) of this cluster's papers illustrates that this is an established research cluster, which is consistent with the vital role of forecasting in OSCM (Huang, Potter, and Eyers 2020), power system planning (Taylor 2003) and water distribution (Adamowski 2008).
Both traditional statistical methods, e.g. regression, exponential smoothing and ARIMA, and ML schemes, i.e. ANNs, were deployed almost equally in this cluster, but those published in the early 2000s focussed more on the former (e.g. Taylor and Majithia 2000;Zhou et al. 2000Zhou et al. , 2002Taylor 2003;Taylor and Buizza 2003). The predictive power of both modelling approaches were compared in more recent papers, of which typical examples include the research of Adamowski et al. (e.g. Adamowski 2008;Adamowski and Karapataki 2010;Adamowski et al. 2012). According to Donkor et al. (2014), ML algorithms are often used for short-term prediction, whereas traditional models, especially regression, are for long-term decision-making. This partly explains the parallel development of both approaches, but we predict that this cluster will see growth in publications combining both schemes like Cluster 2.
Of a particular note is that over 85% of this cluster's 16 studies used data from western countries, including Australia. Since those are industrialised and urbanised nations whose electricity and water infrastructures were built long ago and are now subject to ageing and deterioration, there is a compelling need for research into those distribution systems given that electricity and water are deemed vital in the economy and urban life (Adamowski et al. 2012;Hong and Fan 2016). This partly explains why we obtain an established research cluster on those topics. Another reason is the use of smart meters in the system, which help track demand more accurately (Adamowski et al. 2012;Hong and Fan 2016).

System integration in manufacturing
This is a recent cluster of research with an average publication year of 2011 ( Figure 8) and a standard deviation of 2.50. This cluster is in fact composed of Factors 5 and 6 in FA. Whilst Factor 5's articles studied manufacturing in the context of radio frequency identification (RFID) (e.g. Jiang 2007, 2008b;Zhang, Jiang, and Huang 2008;Huang et al. 2008a;Zhang et al. 2010;Zhang et al. 2011a;, their counterparts in Factor 6 dealt with the Internet of Things (IoT) context (e.g. Xu 2011;Xu, He, and Li 2014;Tao et al. 2014a;Tao et al. 2014b;Tao et al. 2014c). Except literature reviews (i.e. Xu 2011;Bi, Xu, and Wang 2014;Xu, He, and Li 2014), most studies in Cluster 4 were based on conceptual modelling, where conceptual frameworks/ architectures were proposed to apply the IoT or RFID to modern/wireless/cloud manufacturing.
We can notice that the concept behind these 15 papers is system integration, which is required to provide timely information for decision-making in manufacturing (Huang, Zhang, and Jiang 2008b). As can be inferred from Jun et al.'s (2009) discussion about RFID applications in product lifecycle (PLC) management, data need to be shared and analysed at each PLC phase to facilitate decision-making and augment efficiency, e.g. manufacturing, maintenance and reverse logistics. With regard to cloud manufacturing, a recent model where production resources can be shared and operated in the cloud amongst multiple enterprises (Tao et al. 2014a;Tao et al. 2014c), real-time data sharing and analysis are clearly entailed for system monitoring. Obviously, such wireless technologies as the IoT and RFID now play an integral role in collecting and synchronising information in manufacturing systems with multiple processes, levels and resources (Huang, Zhang, and Jiang 2007;Bi, Xu, and Wang 2014). Given the content commonality amongst this cluster's publications, the keyword 'system' is explicably located at Figure 8's centre, connecting all other vertices. Therefore, we name this cluster 'System integration in manufacturing.' Like Cluster 1, this cluster is not dominated by normative, predictive or descriptive modelling as the primary methodology, but is anticipated to see more highly-cited case studies, action research and prototyping-based papers examining the empirical performance of those frameworks and architectures (e.g. Fernández-Caramés et al. 2019). The publication time ( Figure 8) illustrates a research shift over time to the IoT, which leverages RFID at an advanced level (Xu, He, and Li 2014).

Data mining in manufacturing
As can be seen at Figure 9's centre, this cluster's seven articles studied data mining in manufacturing. Indeed, manufacturing systems record and accumulate a large volume of data, which can in turn provide meaningful insight for decision-making if properly-analysed (Chien, Wang, and Cheng 2007). Therefore, research on knowledge extraction from manufacturing data has long been carried out by scholars, amongst whom the highly-cited authors with multiple publictions in this cluster include Kusiak et al. (e.g. Kusiak 2000Kusiak and Kurasek 2001;Agard and Kusiak 2004).
An overview of data-mining applications in manufacturing, including design, production operations, quality control and maintenance, can be found in Harding et al.'s review (2005). For example, Kusiak (2001) proposed a rule-structuring algorithm based on rough set theory for data-driven knowledge discovery in semiconductor manufacturing. For yield enhancement in that industry, Chien, Wang, and Cheng (2007) proposed a fault diagnosis framework whilst Braha and Shmilovici (2002) experimented with data-mining methods to improve processes. Via case studies, Kusiak (2000) and Kusiak and Kurasek (2001) illustrated data-mining applications to fault detection in wafer manufacturing and electronics assembly, respectively. An example of data-mining applications to product design is the framework Agard and Kusiak (2004) proposed to handle product families. Overall, this is an established cluster with diverse research methods adopted, but the retained studies did not specifically target BDA. Yet, they indubitably laid foundations for recent research on BDA in OSCM (e.g. Dutta and Bose 2015).

Data-driven inventory management
The keyphrases linking all other vertices in Figure 10 include 'inventory,' 'lost sales,' 'newsvendor,' 'censored,' 'demand' and 'distribution.' In effect, this cluster's 11 articles developed models to address inventorymanagement problems where demand distribution or its parameters are unknown.
In particular, Burnetas and Smith (2000) employed a multiarmed bandit framework and a stochastic approximation procedure respectively to propose an adaptive pricing and ordering mechanism for perishable items, where the policy is updated as new information arrives. Meanwhile, Godfrey and Powell (2001) leveraged the concave adaptive value estimation algorithm to directly arrive at an optimal decision in response to a given level of remaining inventory and emphasised that they need not estimate demand distribution. Likewise, under the assumption of no information on demand distribution or its family, ordering decisions can be optimised by using only sales data with adaptive inventory policies developed based on stochastic gradient descent and online convex programming (Huh and Rusmevichientong 2009) or the Kaplan-Meier (KM) Estimator (Huh et al. 2011). In a different research vein, Liyanage and Shanthikumar (2005) proposed operational statistics, where demand estimation and inventory optimisation can be carried out together in one single step, directly estimating the optimal order quantity from the data under the assumption that demand distribution although unknown belongs to a given family. It is important to make full use of market data for high-quality inventorymanagement decisions (Wen, Choi, and Chung 2019). Yet, the data utilised for inventory management in practice are just samples of the true demand distribution, so Levi, Roundy, and Shmoys (2007) theoretically analysed the number of samples and sample size required to guarantee that the decision taken would result in a total cost within a predefined confidence level given that the SAA (sample average approximation) is deployed for the single-period problem and (shadow/approximate) dynamic programming for the multi-period problem. Later, Levi, Perakis, and Uichanco (2015) derived a tighter bound for the SAA applied to data-driven inventory management under censored demand. Overall, the impact of censored demand on inventory management can be found in the theoretical analysis of Besbes and Muharremoglu (2013), and Ding, Puterman, and Bisi (2002). This is an established research cluster on data-driven OSCM (publication time in Figure 10) with the primary methodology being modelling. We can see, in recent publications citing this cluster's research (e.g. Ban and Rudin 2019; Bertsimas and Kallus 2020), a rising trend of adopting ML and BDA to optimise inventory-management decisions with one-step algorithms instead of the traditional two-step approach where demand (distribution) must be estimated and then inputted into prescriptive analytics.

Result analysis
Since SCM which derived from and thrived in the manufacturing sector (Vrijhoef and Koskela 2000) has been amongst the main research areas in production scholarship (Kuo and Kusiak 2019;Silva, Ablanedo-Rosas, and Rossetto 2019) and OM originally 'referred primarily to manufacturing production' before incorporating other organisational functions, e.g. logistics and procurement (Bayraktar et al. 2007), our SLR of data-driven OSCM unsurprisingly ascertains findings closely-related to production research. Figure 11 depicts the publication and citation of the identified research clusters on data-driven OSCM.
Since 2000, data-driven decision-making has received increasing attention from production scholars in response to technological advances, which facilitate data management and storage (Kuo and Kusiak 2019). Thus, we can see clusters of studies where data were used to support production and its related fields, i.e. demand forecasting, inventory control and transportation.
Next, when data-driven decision-making became more popular in manufacturing, there was a dire need to collect and synchronise information in manufacturing systems with several levels and processes (Huang, Zhang, and Jiang 2007;Bi, Xu, and Wang 2014). Therefore, a research cluster on system integration, which is required to provide timely information for decision-making in manufacturing (Huang, Zhang, and Jiang 2008b), was formed.
Recently in SCM, BDA has grown fast (Chehbi-Gamoura et al. 2020). Therefore, the most recent research cluster in our SLR is big data (data analytics) in OSCM, where the manufacturing sector accounts for a large proportion.
We can see that the early clusters in our SLR studied IPCs, which, according to OIPT, can be reinforced by organisational integration and information systems (IS) (Trkman et al. 2010;Williams et al. 2013;Srinivasan and Figure 11. Publication and in-sample citation of research clusters on data-driven OSCM. Swink 2018; Irfan, Wang, and Akhtar 2019), a research theme established a few years later in data-driven OSCM. Thence, with IPCs supported by IS assimilation across the SC, firms can leverage voluminous data for OSCM (Irfan, Wang, and Akhtar 2019). This evolution of the highly co-cited literature on data-driven OSCM in our SLR leads us to hypothesise that companies planning to implement BDA should develop data analytics capabilities for their staff and SC partners based on shared data in a coordinated manner so that they can be accustomed to sharing and working on the same dataset before rolling out the new yet integrated IS. In other words, our hypothesis is that organisational data analysis capabilities can develop into BDA when moderated by adequate human and technology resources. This hypothesis accords with the content analysis of Sun et al. (2018), where having human resources with appropriate competencies for BDA adoption ranks higher than technology resources amongst the most vital factors in BDA adoption. Indeed, consultants recommend that feasibility analysis and staff education precede software/system rollout in organisational change processes (cf. Nguyen, Adulyasak, and Landry 2021).
We synthesise conceptual models/frameworks from the retained papers and determine other relevant elements in data-driven OSCM in addition to data assimilation and data management resources as operationalised in line with OIPT in our paper. The result is overall in accord with Sun et al. (2018), but we subsume technological and environmental factors under the external group, which also includes competitive pressures (Chen, Preston, and Swink 2015;Opresnik and Taisch 2015) and interorganisational cooperation (Opresnik and Taisch 2015;Giannakis and Louis 2016). Meanwhile, the internal variables comprise managerial support (Chen, Preston, and Swink 2015;Gunasekaran et al. 2017) and intrafirm collaboration (Dutta and Bose 2015;Giannakis and Louis 2016). We hypothesise that these elements support BDA implementation and thus should be taken in account when managers consider BDA adoption.
To help practitioners find relevant technical/ managerial details in the literature, we point out the retained papers discussing these factors in the supplementary document (Table IV).

Research avenues and implications
Our thematic discussion highlighted that manufacturing, demand prediction, inventory planning and transportation/traffic forecasting are commonly-studied in data-driven OSCM. Given the broad scope of OSCM (Swanson et al. 2018;Manikas et al. 2020), the small number of established research clusters on data-driven OSCM might partly account for the reasons why firms face difficulty adopting BDA. Indeed, ML, BDA and related technologies, e.g. IoT and RFID, were not explicitly discussed in several clusters. This opens many research directions.
First, despite the recognised importance of service operations in OSCM scholarship and practices (Heineke and Davis 2007), there has not been an established cluster for that subfield in research on data-driven OSCM. Even in the recently-published cluster (Cluster 1), most papers remain largely focussed on manufacturing, which Clusters 4 and 5 directly support. With servitisation being facilitated by today's BD (Opresnik and Taisch 2015), OSCM scholars might attend to data-driven service operations and servitisation to ensure practical relevance of their research endeavours. For instance, production researchers could examine which insight or model in data-driven production is transferable to data-driven service operations and vice versa, and what hinders knowledge transfer between the two sectors.
As regards studies on data-driven transportation, the main focus was on flow prediction and description in transportation systems. This does not mean that there are no journal-published papers on prescriptive models for data-driven transportation. Examples include the ML-based vehicle-routing research (Mao and Shen 2018;Tang et al. 2019;Lee et al. 2020). Nonetheless, with so few articles in the literature, there are obviously plentiful opportunities for further research, e.g. empirical evaluation of such data-driven vehicle-routing models. Future research could investigate how the improved traffic-flow forecast can benefit delivery planners and logistics managers, and which forecast horizon optimally facilitates their operations planning.
Turning next to demand-forecasting studies, we can see articles in the utilities sector (electricity and water). This implies that research on demand forecasting for other products and services has plenty of scope for further exploration, e.g. how to leverage data analytics for demand forecasting in e-commerce (see Ferreira, Lee, and Simchi-Levi 2016) and how to integrate traditional and ML approaches into a firm's system if their combination improves demand planning. Case studies and action research can boost practitioners' confidence in the model's efficacy and provide them with implementation guidelines, which should illuminate what and how to change, and how much to invest.
About data-driven inventory-management papers, their cluster is dominated by modelling. Thus, empirical studies on those models' real-life performance are essential additions to this research topic. Moreover, the datadriven inventory-management literature can be enriched by applying ML schemes and comparing or combining traditional optimisation approaches and ML algorithms (e.g. Ferreira, Lee, and Simchi-Levi 2016;Bertsimas and Kallus 2020). There are similar papers in our 580-paper pool (e.g. Sachs and Minner 2014; Ban and Rudin 2019), but their in-sample citations are currently below the inclusion threshold in our co-citation analysis. Fellow scholars might investigate how to select a suitable model (or a combination of models) to manage inventory, how to put it in place and which benefit to reap in reality.
Finally, there are several articles on other subfields of OSCM such as supply management (Cheng et al. 2020), warehousing (Fernández-Caramés et al. 2019), distribution planning (Chen et al. 2017), retail operations (Ozgormus and Smith 2020) and maintenance (Kumar, Shankar, and Thakur 2018), some of which are highlycited and included in our sample, e.g. supply management (Ho, Xu, and Dey 2010;Cheng et al. 2020), but have not established a distinct cluster in data-driven OSCM. Thus, the need for further research is undeniable.
Considering the academic vantage point, our paper demonstrates a concrete example of data-driven SLR where a well-established protocol is followed and data analysis triangulation is performed to enhance robustness and reduce subjectivity. Moreover, vis-à-vis other recent OSCM literature reviews, our paper searches more databases, uses more search keywords and covers a longer time span (cf. Brinch 2018; Mishra et al. 2018;Chehbi-Gamoura et al. 2020;Winkelhaus and Grosse 2020). Also, by using cross-referencing, we can retrieve highly-cited and relevant papers that are not indexed in the predetermined databases. Hence, our paper can uncover a broader overview of data-driven OSCM where BDA now features prominently.
From a theoretical perspective, we ascertain the literature evolution and knowledge structure of studies on data-driven OSCM which future research can reply on.
OIPT (Galbraith 1974), based on which we defined data-driven OSCM in this work, is one theoretical lens adopted by several papers in our selected sample, e.g. Trkman et al. (2010), Hazen et al. (2014), Chen, Preston, and Swink (2015), Srinivasan and Swink (2018) and Dubey et al. (2019). The reviews of Gupta, Modgil, and Gunasekaran (2020) and  also proposed other theoretical frameworks, but overall, organisational theories were less commonly considered in our selected literature on data-driven OSCM from 2000 to early 2020. Thus, we recommend drawing on diverse theoretical perspectives to enrich the theoretical bases of (big) data-driven OSCM scholarship. The following open research questions may suggest fruitful research agendas: • As a manufacturing system involves multiple processes (Huang, Zhang, and Jiang 2007;Bi, Xu, and Wang 2014), e.g. demand forecasting, distribution management, capacity planning, inventory control and supply management, how should manufacturers implement BDA in these processes (simultaneously)? Which department will play a leading role? What is a guiding theory for that rollout? OIPT provides an important theoretical framework, but the other theories discussed by Gupta, Modgil, and Gunasekaran (2020) and  are also worth considering. • Interorganisational collaboration is an influencing factor in BDA adoption (Opresnik and Taisch 2015;Giannakis and Louis 2016). However, Nguyen, Adulyasak, and Landry (2021) showed that the benefits of reduced stockout, inventory and bullwhip effect via information sharing were least salient at retail outlets in a fully-integrated distribution resource planning system. How should manufacturing firms engage their retailers in BDA implementation then? (Supply) uncertainty reduction in line with OIPT might be a good candidate to justify cooperation, but we do not exclude the potential of other theoretical lenses for this question. • In addition to retail partners, other SC players in a manufacturing system, e.g. suppliers and third-party logistics providers, also need to be involved. It is hence open to question what SC configuration is optimal to put BDA in place. What intermediaries may become redundant then? Upon BDA adoption, what benefits should be measured? Will those benefits be enjoyed fairly amongst SC members or should some redistribution mechanisms be in place to encourage ongoing BDA in that interconnected system? The answers likely lie in the information needs or core processes of the focal company (OIPT), but other theoretical perspectives, e.g. those on competitive pressures (Chen, Preston, and Swink 2015;Opresnik and Taisch 2015), might equally come into play. • BDA can be deemed a disruptive technology (Brinch 2018;Fu and Chien 2019), so adopting it may entail a new perspective. What theory can provide such a revelatory perspective on BDA in OSCM? We refer interested readers to the theories listed by Gupta, Modgil, and Gunasekaran (2020) and  for potential starting points.

Practical implications
For practitioners, our paper synthesises highly-cited papers that can inform their decision-making, especially in manufacturing, demand forecasting, inventory management and transportation. Several highly-cited models and frameworks have been developed to utilise data, which can be voluminous, noisy or incomplete (as in censored demand). Practitioners in production can refer to these established clusters to guide their decisionmaking but need to consider if the research context and assumptions are relevant to their enterprises. It should be noted that simple and complex models can perform equally (Karlaftis and Vlahogianni 2011), so managers should assess the fit between a model/framework and their system/practice and carefully plan the change process before deciding on adoption.

Conclusion
In this paper, we used the keywords and databases commonly-used in the OSCM literature to seek pertinent publications and applied co-citation analysis to determine the knowledge structure of data-driven OSCM since 2000, which is also when data-driven decisionmaking started receiving attention from production researchers (Kuo and Kusiak 2019). There are prior reviews on data-driven OSCM and co-citation-based SLRs, but our paper retrieved a larger literature sample thanks to cross-referencing and it is amongst the first endeavours to conduct a data-driven SLR of data-driven OSCM with multiple clustering tools, i.e. VOSviewer, FA in STATA and scikit-learn MDS-based enhanced kmeans clustering, as data analysis triangulation. This is the originality of our study. There are six clusters of research appearing in our analysis results, namely Big data (data analytics) in OSCM, Transportation and traffic flow prediction, Demand forecasting, System integration in manufacturing, Data mining in manufacturing and Data-driven inventory management, five of which closely relate to production. This is unsurprising because OSCM is closely associated with production research as previously discussed. Traditional statistical and econometric approaches remain commonly-deployed in forecasting, but ML programs and BDA are becoming popular. Indeed, ML and BDA have been widely-undertaken in the literature on SCM, production, traffic prediction and demand forecasting. We synthesised these clusters of studies as crucial points of reference for production research and practice based on ML and BDA. The evolution of the identified clusters also suggested a procedure for BDA adoption in production, where staff's data analysis capabilities must be prioritised and developed with the support of proper technology resources so that BDA can be rolled out successfully, which is in line with OIPT. Also, competitive rivalry, inter/intra-organisational collaboration and managerial support are important factors to consider in BDA adoption.
Given the interdisciplinary nature of OSCM (Swanson et al. 2018;Manikas et al. 2020) and manufacturing systems involving multiple processes, levels and resources (Huang, Zhang, and Jiang 2007;Bi, Xu, and Wang 2014), we expect to see more clusters of highly-cited research on ML and BDA for inventory control, supply management, distribution planning, etc., to improve the knowledge base that production scholars and practitioners can directly leverage.
Albeit we carried out a systematic literature review, which allowed finding answers to our research questions, our study has limitations. First, the clusters identified may have been confined to the keywords utilised for database search. However, since generic keywords such as SCM, OM and logistics were used and research on other OSCM subfields was also retrieved, the clusters identified here may well reflect the current knowledge state of the data-driven OSCM literature published since 2000. Second, one inherent drawback of co-citation analysis is its inability to identify emerging research areas (Fahimnia et al. 2019) as lately published articles are less likely to be included given their insufficient time to accumulate citations. Nonetheless, this indirectly confirms the proposed research opportunities.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
This work was supported by the Natural Sciences and Engineering Research Council of Canada: [grant numbers CGS D3-535738-2019, 2021-03264].

Notes on contributors
Duy Tan Nguyen is a PhD candidate in Logistics and Operations Management at HEC Montréal and a student member of the Group for Research in Decision Analysis (GERAD), Montréal, Canada. His research focus is on inventory control, distribution planning and supply chain management.
Yossiri Adulyasak holds the Canada Research Chair in Supply Chain Analytics and currently serves as an associate professor at HEC Montréal. His research has been published in top-tier journals and conferences in operations research (OR), machine learning (ML) and artificial intelligence (AI). He leads research and industrial initiatives which integrate predictive and prescriptive analytics in various supply chain applications in manufacturing, retail, omni-channel and transportation.

Jean-François Cordeau is a professor of
Operations Management at HEC Montréal, where he also holds the Chair in Logistics and Transportation. He has authored or coauthored more than 150 scientific articles in combinatorial optimisation and mathematical programming, mostly in the fields of vehicle routing, network design and transportation terminal management.