Data Science Project Barriers—A Systematic Review

Abstract

This study aims to identify and categorize barriers to the success of Data Science (DS) projects through a systematic literature review combined with quantitative methods of analysis. PRISMA is used to conduct a literature review to identify the barriers in the existing literature. With techniques from bibliometrics and network science, the barriers are hierarchically clustered using the Jaccard distance as a measure of dissimilarity. The review identified 27 barriers to the success of DS projects from 26 studies. These barriers were grouped into six thematic clusters: people, data and technology, management, economic, project, and external barriers. The barrier “insufficient skills” is the most frequently cited in the literature and the most frequently considered critical. From the quantitative analysis, the barriers “insufficient skills”, “poor data quality”, “data privacy and security”, “lack of support from top management”, “insufficient funding”, “insufficient ROI or justification”, “government policies and regulation”, and “inadequate, immature or inconsistent methodology” were identified as the most central in their cluster.

1. Introduction

The world economy has proven capable of generating more and more resources each year, despite climate concerns. During the decade that began in 2010, the world produced an average global GDP growth above 2% every year—except for 2020, the year of the coronavirus outbreak [1]. However, there seems to be nothing in the economy for which growth compares to that witnessed in data generation: in 2020, 64 zettabytes of data (64 × 10²¹ bytes) were generated, equivalent to approximately 32 times what was generated 10 years before, at the beginning of the decade [2].

This exponential growth in data generation has led to a host of data storage and processing challenges, but also to many opportunities. One of the first people to realize this was Clive Humby, who in 2006 coined the famous phrase “data is the new oil” [3], referring not only to the intrinsic value of data but also to its need for refinement. At least from an economic perspective, his prediction proved accurate, as the so-called “FAAMG” (Facebook, Apple, Amazon, Microsoft, and Google), technology giants that generate and process enormous amounts of data, were among the top 10 companies in the world by 2021 [4]. Reinforcing this idea, a report commissioned by BusinessWire [5] predicts a $243 billion market by 2027 for so-called Big Data (BD)—data of great volume, velocity, and variety.

To deal with the world of Big Data, a new area of knowledge has emerged and consolidated—Data Science (DS). The term in its current definition was coined in 2008 [6]. LinkedIn [7] says one hundred and fifty million new jobs related to digital transformation are expected in the next 5 years—among them are artificial intelligence (AI) specialists and data scientists. These professionals are responding to a growing demand from industry. An annual survey by NewVantage [8] that interviews blue chip executives in the US market shows that 99% of these companies are investing in Big Data and AI.

Despite the investment, not all companies are successful when it comes to implementation. Only 13% of companies surveyed by NewVantage [8] had fully deployed Big Data in their operations. From the perspective of deployment initiatives, the surveys reveal a low rate of success: estimates range from 27% [9] and 20% [10] to 8% [11].

On one hand, Big Data and AI present huge potential. On the other hand, despite high investment, the success rate of initiatives is low. However, the field of Data Science goes beyond Big Data or AI—despite the common use of those terms as synonyms—and the current literature focuses on the adoption of Big Data by organizations, with a reduced analysis of barriers that are related to the context of a DS project. Moreover, when categorizing barriers, the results differ from study to study. Thus, this study aims to answer the following research questions:

RQ1: What are the barriers to the success of Data Science projects?

RQ2: How should these barriers be categorized? Additionally, which barriers are the most relevant to their categories (or clusters)?

2. Literature Review

2.1. Data Science

Data Science encompasses “a set of principles, problem definitions, algorithms, and processes for extracting non-obvious patterns from large data sets” and aims to “improve decision-making by basing decisions on knowledge extracted from data” [12].

There are, perhaps, two arguments against this definition. The first is that “large” in “large data sets” is subjective. An answer to that is the definition of large as “too large or complex for a human being to analyse”. While we are capable of understanding patterns in dozens of records with two or three attributes, data science is generally applied in contexts where we seek to find patterns among tens, hundreds, or perhaps even millions of attributes among an even larger number of records. The second argument is that there is already a branch of science that deals with data collection and analysis: statistics. This argument is also valid; however, the fields of statistics and Data Science have been differentiated to such an extent that it is no longer possible to give the same name to different fields. The first and most important differentiation is that data scientists use both statistical tools and computational algorithms, such as machine learning (ML) algorithms. The second differentiation is that data scientists are also concerned with how to extract the data, process it, and then implement the model they built. Finally, the third differentiation is perhaps cultural. Breiman [13] published an article entitled “Statistical Modelling: The Two Cultures.” In short, the traditional view is that the goal of data analysis is explanatory, in the sense of bringing about an understanding of how the data was generated. In contrast, the new view, which better defines Data Science, sees the goal of data analysis as that of developing algorithms with the highest prediction capacity possible.

Data Science, therefore, has become sufficiently differentiated from statistics and other fields to earn a name of its own. DS combines knowledge from statistics, computer science, and ML to solve current problems by analyzing large datasets. The discipline is recent because only recently has humanity begun to produce and be able to analyze large amounts of data. In addition to the emergence of Big Data, the economy of scale provided by the data market and the evolution of processing equipment contributed to the emergence of Data Science [12], such that even today, despite the emergence of several higher-education courses in the discipline, the demand for these professionals continues to increase [7].

When defining Data Science, it is also important to characterize some similar terms that are commonly mixed: specifically, Big Data, data analytics, and machine learning. Big Data is traditionally defined based on the 3Vs: data in large volume, variety, and velocity. Volume means the amount of data, variety refers to the types of data, and velocity is the speed with which data is generated. More recently, the 3Vs have been expanded into the 5Vs, to include veracity and value, and even the 6Vs (see, for example, Rehman et al. [14]). Data analytics, or data analysis, is simply the process of inspecting, cleaning, transforming, and modeling data for the purpose of supporting decision making [15]. ML, on the other hand, is an advanced technique of data analysis and can be seen as a component of data analytics [16].

The terms Big Data and data analytics are often used together under the new term Big Data Analytics (BDA). The latter is sometimes used as a synonym for DS (see, for example, Krasteva and Ilieva [17]). However, some authors choose to differentiate between the terms—considering BDA as the process of analyzing Big Data, and DS the broader field of knowledge. Moreover, DS does not apply only to Big Data. Having lots of data helps, but having the right data is more important [12].

2.2. Barriers for Data Science Projects

From PMI [18], a project is “a temporary effort undertaken to create a unique product, service, or result. The temporary nature of projects indicates a beginning and an end to the project work or a phase of the project work.” DS deals, naturally, more with projects than with processes—whether developing a clustering algorithm for customer segmentation [19,20], detecting anomalies such as fraud [21,22], or building a predictive model to identify whether a taxpayer is evading their taxes [23,24]—data scientists usually work in unique, temporary efforts.

These efforts, however, do not always pay off. As noted previously, up to 92% of projects fail [11]. Hindering the success of DS projects are multiple barriers—such as insufficient skills [25], low data quality [26], and lack of collaboration [27].

To deliver, a standard DS project goes through the stages of understanding the business problem, understanding the data, preparing the data, modeling, evaluating, and deploying [12]. Different methodologies have been developed to address three of the fundamental aspects of a DS project—project management, team management, and data management, with Cross-Industry Standard Process for Data Mining (CRISP-DM) historically being the most popular [28]. However, DS methodologies lack integrity, including, notably, CRISP-DM [29]. Moreover, as noted by Morlock and Boßlau [30], a fourth aspect—change management—is ignored.

3. Methods

To fulfil the research objective of identifying and categorizing barriers to the success of DS projects, this study is subdivided into two stages (similar to Rameezdeen [31]). In the first, the PRISMA methodology [32] is used to identify the barriers in the literature. In the second stage, the thematic clustering of the barriers is performed using quantitative methods with techniques from bibliometrics and network science. As suggested by Xiao and Watson [33], grouping factors into themes is a way to deepen the understanding of the research problem.

3.1. Systematic Literature Review

To identify the barriers to DS projects in the existing literature, a systematic literature review (SLR) was conducted based on the PRISMA methodology [32]. The methodology consists of 4 steps: identification, screening, selection, and inclusion of studies. After searching and selecting the primary studies with the search terms, an expansion of the search was performed through the snowball technique (or citation searching), in which new studies are sought from the citations of the initial studies.

The selection process included the following criteria: (a) studies that list DS barriers; (b) research methodology includes some variant of multi-criteria decision-making (MCDM) methods, since these methods have a higher depth of factor analysis, including impact estimation and/or cause-and-effect relationships; (c) studies from journals included in Scopus and Web of Science (WoS) databases; and (d) studies in English. No temporal criteria have been included. The result of this research showed that there are no MCDM analyses specifically for DS projects. The result was not totally satisfactory in the sense that all 14 studies found are focused on the adoption or deployment of Big Data or Big Data Analytics by organizations in general, albeit in varying contexts. Although these studies are still helpful in answering the research questions, given Big Data is a similar field to DS, and thus were included in the SLR, they focus on barriers faced by organizations during an adoption process and do not fully include barriers related to DS projects.

Thus, to supplement the first search strategy, a second strategy was designed, broadening the scope to include methodologies beyond MCDM analysis and including conference studies, but restricting the scope to DS projects, to exclude studies that refer broadly to technology adoption or deployment in organizations. In this search, the selection process included the following generic criteria: (a) studies that list barriers for DS projects; (b) studies from journals or conferences included in Scopus and WoS; and (c) studies in English.

Table 1. Search strategies and search terms.

Search Strategy	Topic	Search Terms *	Result	Search Date
Strategy 1	Data Science	“data science” OR “big data” OR “data analytics” OR “machine learning” AND	213	20 October 2022
	Barrier	“barrier*” OR “obstacle*” OR “challenge*” OR “hindrance*” AND
	MCDM	“multi-criteria decision-making” OR “MCDM” OR “AHP” OR “TOPSIS” OR “VIKOR” OR “ANP” OR “DEMATEL” OR “PROMETHEE” OR “ELECTRE” OR “ISM” OR “TISM” OR “MICMAC”
Strategy 2	Data Science Project	“data science project” OR “big data project” OR “data analytics project” OR “machine learning project” AND	239	21 October 2022
	Barrier or project failure	“barrier*” OR “obstacle*” OR “challenge*” OR “hindrance*” OR “fail*”

* Terms such as “data mining” and “artificial intelligence” were initially considered and tested but ultimately excluded due to high retrieval of unrelated or overly broad studies, particularly outside the scope of data initiatives.

details the two search strategies and terms for identification. The sum of strategies 1 and 2 is equivalent to the total number of studies identified in databases (452), before removal of duplicates (137), as shown in Figure 1. After identification, the screening process included title and abstract reading.

After retrieval, the studies were further assessed for eligibility with full paper reading and excluded for (a) being out of scope, when the study did not support in answering the research questions (e.g., [35]); (b) lack of clarity, when the barriers are not clearly defined (e.g., [36]); (c) being excessively circumstantial, when the barriers were context-specific and could not be generalized (e.g., [37]); and (d) having inconsistent methodology. The data collection process involved gathering the barriers from each of the studies included in the review in a single list, marking each of the studies that mentioned them. This list is presented in the Results section.

Researchers (1, 2, and 3) defined the identification, screening, inclusion, and data collection process, which was then carried out by researcher (1). We assessed the methodological quality of included studies by considering key aspects such as clarity of aims, study design, transparency of data collection and analysis, and adherence to the stated methods. Given that the review focused on identifying barriers rather than measuring effect sizes, the overall risk of bias was considered low.

By using a structured approach to the SLR, following PRISMA guidelines, and searching two major research databases with two distinct search strategies, we aimed to reduce the risk of reporting bias due to missing results.

The authors declare adherence to the PRISMA Statement [32]. However, the SLR was not registered, and the review protocol was not registered a priori in PROSPERO or OSF.

3.2. Quantitative Methods for Clustering Barriers

Of the selected studies, 12 categorized barriers into thematic groups, i.e., classified barriers as organizational, technological, economic, etc. (e.g., [38]). Each study used its own categorization criteria—including the choice of the number of groups, the definition of the group, and how the barriers were classified. This presents the literature review with the challenge of how to categorize the barriers, given the multitude of criteria used in the different studies. This challenge was overcome by choosing a quantitative method of thematic grouping of the barriers that considers the categorization of all 12 studies. Rosenthal and DiMatteo [39] state that by numerically aggregating the results of multiple studies, quantitative methods help solve the problem of multiple answers to a research question.

Each study categorized the barriers into 2 to 8 thematic groups. In total, 52 distinct groups were formed. Since the barriers were grouped based on their relation to a theme, it is implied that the barriers in a group are correlated (even if only thematically). Since a barrier–barrier relationship is defined as the common presence of two barriers in a group, the number of relationships formed in a group grows proportionally to the number of barriers in that group, following:

$$R={\displaystyle \frac{N\ast (N-1)}{2}},$$

(1)

where $R$ is the number of relationships in a group and $N>1$ is the number of barriers in a group.

From the relationships formed by the 52 groups, a relation network is established, with each barrier as a vertex, and each relationship a connection between two vertices. This network can be interpreted from the perspective of network science, whose goal is to develop approaches and techniques for understanding networks [40].

For the categorization of the barriers, this study used hierarchical clustering with the average linkage method, which is one of the most used in clustering and tends to generate better results [41]. This method considers not only the relationships between vertices, but also the strength of those relationships—or, more accurately, the lack thereof—and is therefore suitable for this study. If a barrier–barrier pair was categorized in the same group, this means that there is a relationship between these barriers. The higher the number of common groups two barriers have with each other, the greater the strength of the relationship.

The proximity measure that best characterizes this conjuncture is the Jaccard index. Named after the scientist of the same name [42], the measure was chosen thanks to its normalized range of values (from 0 to 1) and the characteristics of the network. Its complement, Jaccard distance, is commonly used by clustering algorithms [41], and is defined by the formula [43]:

$$J_{\delta }\left(B_n,B_m\right)=1-J\left(B_n,B_m\right)=1-{\displaystyle \frac{\left|B_n\cap B_m\right|}{\left|B_n\cup B_m\right|}},$$

(2)

where $J_{\delta }$ is the Jaccard distance, $J$ is the Jaccard index, and $B_n$ and $B_m$ are sets. In this case, each set represents a barrier and contains the different groups (from 1 to 52) into which the barriers are categorized.

To determine the number of clusters, a combination of the elbow method and the silhouette score was used. The elbow method involves the relationship between the average within-cluster distance ($\overline{∆}$) and the number of clusters ($k$). The idea is that increasing $k$ from a certain point provides diminishing returns in terms of decreasing $\overline{∆}$. That value is the “right” number of clusters [44]. The average within-cluster distance for a number of clusters can be calculated with:

$${\overline{∆}}_k={\displaystyle \frac{{\sum }_{i,j}d(b_i,b_j)}{{\sum }_k{R}_k}},$$

(3)

where $d$ is the Jaccard distance between two barriers $b_i$ and $b_j$, which belong to the same cluster, and $R_K$ is the number of relationships within a cluster.

The silhouette statistic is calculated for each vertex and provides how well it fits its assigned cluster. The statistic is normalized: a value near 1 means an almost perfect cluster allocation and a value near −1 means the opposite. For a given vertex ($b_n$), the silhouette ($s$) is the difference between the mean nearest-cluster distance ($c$) and the mean intra-cluster distance ($a$) divided by the biggest between the two [45]:

$$s_{b_n}={\displaystyle \frac{c_{b_n}-a_{b_n}}{max(c_{b_n},a_{b_n})}},$$

(4)

where the silhouette score is the average value for all silhouettes.

To complement the analysis, two centrality measures were calculated: degree centrality and closeness centrality. The degree centrality $C_D$ represents the number of vertices adjacent (that is, related or connected) to a vertex ($b_n$) whose degree is being calculated [46]:

$$C_D\left(b_n\right)=degree\left(b_n\right)$$

(5)

The second measure is closeness centrality, formally defined as the inverse of the sum of the distances from one vertex to the others. This measure computes the relative proximity of a vertex to the others [40]. Here, the measure is calculated within-group and is used to quantify the centrality of a vertex relative to its cluster. Closeness centrality ($C_c$) in its normalized form is defined by:

$$C_c\left(b_n\right)={\displaystyle \frac{N-1}{\sum _{b_m}d\left(b_n,b_m\right)}},$$

(6)

where $N$ is the number of vertices—or barriers—in a cluster, $d$ is the Jaccard distance between two vertices, $b_n$ is the vertex whose centrality is to be obtained, and $b_m$ represents the other vertices.

We enhanced robustness and confidence by excluding studies with inconsistent methodology and by including only barriers that were reported by more than one study. This approach helped reduce the influence of outliers. Additionally, by clustering them as described, we ensure that the thematic categories reflect patterns observed across multiple sources.

4. Results

4.1. Systematic Literature Review

The SLR resulted in a total of 26 studies, whose context, methods, and keypoints are presented in

Table 2. Summary of the studies of Data Science barriers.

#	Author	Context	Methods	Keypoints
A	Kavre et al. (2022) [47]	Small and medium-sized enterprises (SME), India	Literature review, expert opinion, ISM, DEMATEL	Barriers to BDA adoption. Analyzes the impact and relationships among the barriers.
B	Sharma et al. (2022) [48]	Manufacturing, public sector, India	DEMATEL	Barriers to machine learning adoption. Analyzes the impact and relationships between the barriers.
C	Gangadhari et al. (2022) [49]	Oil industry, India	Literature review (PRISMA), DEMATEL, COPRAS, MOORA, Delphi (n = 10)	Barriers to machine learning adoption. Analyzes the impact and relationships between the barriers.
D	Nayal et al. (2021) [27]	Agriculture, COVID-19, India	Delphi, ISM, Fuzzy MICMAC, ANP	Barriers to machine learning adoption. Analyzes the impact and relationships between the barriers.
E	Raut, Yadav, et al. (2021) [25]	Manufacturing, India	ISM, Delphi, Fuzzy MICMAC, DEMATEL, expert opinion (n = 47)	Barriers to BDA adoption. Analyzes the impact and relationships between the barriers.
F	Park & Kim (2021) [26]	No specific industry, Korea	AHP (n = 50), regression analysis (n = 226)	Barriers to BD adoption. Categorizes and analyzes the impact of the barriers.
G	Gupta & Goyal (2021) [50]	Manufacturing, India	Literature review, survey, ISM, MICMAC, Fuzzy AHP, expert opinion (n = 16)	Barriers to BDA adoption. Categorizes and analyzes the impact and relationships among the barriers.
H	Bahrami et al. (2021) [51]	Startups, Iran	Interviews, survey, Fuzzy AHP	Barriers faced by BD startups. Categorizes and analyzes the impact of the barriers.
I	Raut, Narwane, et al. (2021) [52]	Supply chain, India	Literature review, survey, DEMATEL, ANP, expert opinion (n = 15)	Barriers to BDA adoption. Analyzes the impact and relationships among the barriers.
J	Bag et al. (2020) [53]	Third sector, supply chain, Africa	Literature review, Fuzzy TISM (n = 5), survey (n = 108), SEM	Barriers to BDA adoption. Analyzes the impact and relationships among the barriers.
K	Alalawneh & Alkhatib (2020) [38]	Financial, industrial, services, public and supply chain sectors, Jordan	Literature review, Semi-structured interviews, survey (n = 23), AHP, TOPSIS	Barriers to BD adoption. Categorizes and analyzes the impact and relationships among the barriers.
L	Zhang & Lam (2019) [54]	Maritime industry	Fuzzy-Delphi (n = 6), Fuzzy AHP (n = 20), TOPSIS	Barriers to BDA adoption. Categorizes and analyzes the impact of the barriers.
M	Moktadir et al. (2019) [55]	Supply chain, Bangladesh	Literature review, Delphi (n = 15), AHP, expert opinion	Barriers to BDA adoption. Categorizes and analyzes the impact of the barriers.
N	Shukla & Mattar (2019) [56]	Agriculture	Literature review, ISM, MICMAC, expert opinion	Barriers to BDA adoption. Analyzes the impact and relationships between the barriers.
O	Kastouni & Lahcen (2022) [57]	Telecommunications	Case Study (n = 1)	Barriers for the studied BDA project. Categorizes the barriers.
P	Martinez et al. (2021) [29]	No specific context	Literature review	DS project management methodologies. Proposes a framework for framing the methodologies. Raises challenges for DS projects.
Q	Aho et al. (2021) [58]	No specific context	Survey (n = 50)	Barriers and other issues for DS projects.
R	Escobar et al. (2021) [59]	Manufacturing	Literature review	Barriers to DS projects.
S	Wang & Wang (2020) [60]	Small and medium-sized enterprises	Literature review, case studies (n = 8)	Proposes a knowledge management model for BD. Raises barriers for BD projects.
T	Saltz & Shamshurin (2019) [61]	Undisclosed	Case study (n = 1)	Agile methodologies for DS projects. Raises barriers for DS projects.
U	Jensen et al. (2019) [62]	Energy	Case study (n = 1)	Barriers for the BD project studied.
V	Kim et al. (2018) [63]	Technology	Survey (n = 793)	Barriers and other issues faced by data scientists in projects. Categorizes the barriers.
W	Barham & Daim (2018) [64]	Smart cities	Literature review	Barriers for BD projects. Categorizes the barriers.
X	Barham (2017) [65]	No specific context	Literature review	Barriers and other issues for DS projects.
Y	Becker (2017) [66]	No specific context	Survey (n = 19), system modelling and simulation	Model for predicting outcomes of BD projects. Raises causes of failure for projects of this type.
Z	Chipidza et al. (2016) [67]	Supply chain	Case study (n = 1)	Barriers for the BD project studied.

. In the keypoints column, we summarize the main objective of the paper, what the studied barrier is hindering (e.g., BDA or ML adoption), and indicate whether the study analyzes the impact of each barrier, the relationships among them, and whether it provides a classification for the barriers. These last aspects are generally determined by the study’s methods—DEMATEL, for instance, calculates a relationship matrix.

Of the 26 studies included in the SLR, 12 categorized the barriers into themes, such as organizational or technological barriers. The barriers were grouped into two to eight categories, depending on the study. The method used for categorization was based on existing literature, expert opinion, or the authors’ own judgment. In addition, 11 studies performed an impact analysis capable of informing which are the most critical or highest impact barriers. In all these cases, some MCDM method was used for this purpose.

Figure 2 details the number of articles by research context. It is possible to notice a wide range of industries and sectors that were studied, with some focal points in manufacturing and supply chains.

Differences in context and methodology may help explain the heterogeneity in reported barriers across studies. In the following section, we list each barrier identified in individual studies and then group them into broader thematic categories. This approach helps capture underlying issues faced by data science projects across diverse settings.

Barriers Identified

The studies included in the SLR identified, on average, 12 barriers to the adoption of Big Data or to the success of DS projects, with a high degree of repetition. As presented in the methods section, one of the exclusion criteria for studies from the systematic review was a lack of clarity in exposing the barriers—that is, the study needs to establish, in objective and clear terms, which barriers were found in the investigation. After that, the procedure for identifying the barriers was a simple extraction of the barriers as written in the studies.

In addition to the study exclusion criteria, barriers were also discarded if (a) they were too specific to the research context, such as the industry or the case study in question; (b) they were not categorized into any group; or (c) they were not confirmed by more than one study.

The compilation procedure was based on the names and descriptions of the barriers as given by the studies. Thus, in response to RQ1, the review identified 27 unique barriers, described in

Table 3. Description of the barriers.

#	Barrier	Description
B1	Insufficient skills	The organization does not have individuals with the necessary skills to execute the project.
B2	Poor data quality	The data provided is of poor quality (i.e., lacking accuracy, relevancy, and representation)
B3	Insufficient IT infrastructure	The organization’s IT infrastructure does not support the project’s needs.
B4	Data privacy and security	Privacy and security risks are not managed properly.
B5	IT illiteracy	Employees involved with the project do not have basic technology knowledge in order to be able to cooperate or use the tools developed.
B6	Lack of support from top management	Top management does not provide the necessary support for the development of the project.
B7	Complexity of data or technology	Data or technology characteristics are too complex.
B8	Lack of an integrated data environment	There is a lack of an integrated data environment from which the project can load and write information.
B9	Insufficient funding	There is not enough funding for the project’s needs.
B10	Immature technology and lack of appropriate tools	The technology is immature in the sense of well-established responsibilities, processes, and tools.
B11	Strategy mismatch	There is a mismatch between company strategy and project goals.
B12	Inadequate data sharing policy	The organization’s data sharing policy hinders the team’s access to needed data.
B13	Scalability issues	The established infrastructure does not have sufficient scalability capacity to support data processing and storage over time.
B14	Resistance to change and other cultural barriers	There is unwillingness on the part of the organization to adapt to the new processes. This includes other cultural barriers such as fear of technology and behavioral problems.
B15	Insufficient ROI or business case	The proposed project does not have sufficient ROI or business case (justification) to gain the support of the decision-makers.
B16	Inadequate training programs and facilities	The organization’s training programs and/or facilities are inadequate.
B17	High investment and maintenance cost	Project implementation requires a high investment cost and also leads to high maintenance costs.
B18	Government policies and regulation	Government regulation and/or policies are insufficient, counterproductive, or unclear.
B19	Poor data management and architecture	The organization’s data management and architecture are inadequate, making it difficult to acquire data and integrate the new processes into the existing architecture.
B20	Lack of coordination, collaboration, and communication	There is a lack of coordination, collaboration, or communication between the parties involved in the project.
B21	Data availability	The necessary data is not available, either due to a lack of access, or other issues, such as data loss, lack of record keeping, and physical (non-electronic) storage.
B22	Inadequate or inconsistent methodology	The methodology used is inadequate, inconsistent, or immature.
B23	External sources of data	External sources of data are not available to the organization, either due to legal issues, high costs, etc.
B24	Scope, objectives, and expected results unclear	Scope, objectives and expected results were poorly defined at the beginning of the project.
B25	Uncertainty about benefits	The organization is uncertain about the benefits of the project. There is a lack of support to start the project or to develop the activities.
B26	Deployment and sustainability issues	Deployment and/or sustainability are inadequate leading to low usage of project deliveries, possibly due to a lack of an implementation plan to manage the changes in the business processes.
B27	Associated risks	Risks associated with the project are high or overestimated.

and identified in

Table 4. Barriers identified in the literature. (∙) indicates a reference to the barrier, (x) indicates a reference that classified the barrier as critical, Freq. indicates the number of references for each barrier, and F. (x) indicates the number of references that identify each barrier as critical.

			Search strategies
			Strategy 1														Strategy 2
#	Freq.	F. (x)	A	B	C	D	E	F	G	H	I	J	K	L	M	N	O	P	Q	R	S	T	U	V	W	X	Y	Z
B1	20	5	∙		x	∙	x	x	∙	∙	∙	∙	∙	x	∙	x	∙	∙	∙						∙	∙	∙	∙
B2	18	2		x	∙	∙	∙	x	∙		∙	∙		∙	∙	∙	∙	∙	∙			∙	∙	∙	∙
B3	15	4		∙	x	∙		∙	x		x	∙	∙		x	∙	∙	∙	∙		∙						∙
B4	15	4		x	x	x	∙	∙	∙	∙	∙		∙		x	∙		∙	∙		∙				∙
B5	15	2	∙	x	∙	∙			∙	∙	∙		∙	x					∙	∙	∙		∙		∙		∙
B6	14	4	x	∙	∙		x	∙	x	∙	∙		∙	x		∙		∙							∙		∙
B7	11	2	∙					∙	∙			∙	∙		x	x								∙	∙	∙	∙
B8	11	2					∙	x	∙		∙	∙			x		∙							∙	∙	∙	∙
B9	10	3			∙		x	x	x	∙	∙	∙	∙		∙	∙
B10	9	3				∙			x				∙		x	x		∙	∙		∙			∙
B11	9	2				∙			x		x			∙							∙		∙		∙		∙	∙
B12	9	1			x				∙	∙			∙		∙	∙	∙							∙	∙
B13	9	0				∙	∙		∙				∙		∙	∙	∙	∙						∙
B14	9	0	∙		∙		∙				∙	∙	∙			∙									∙		∙
B15	9	0				∙	∙				∙		∙	∙		∙		∙					∙				∙
B16	8	3	x			∙		x	x	∙		∙	∙		∙
B17	8	2		x	∙	∙		x	∙				∙		∙	∙
B18	8	2				x		∙		∙	∙	∙	∙			x									∙
B19	8	1		x			∙			∙		∙		∙			∙				∙						∙
B20	8	1	x	∙		∙												∙	∙			∙	∙				∙
B21	8	0			∙	∙												∙	∙			∙		∙	∙			∙
B22	6	1					x											∙	∙	∙			∙				∙
B23	6	0		∙	∙			∙					∙												∙	∙
B24	6	0																∙	∙	∙			∙		∙		∙
B25	5	2						x			x			∙						∙				∙
B26	4	0			∙													∙		∙			∙
B27	4	0	∙			∙				∙						∙

.

The 27 barriers identified and described above are presented in Table 4, cross-referenced with each study. Evaluating the SLR results, it is notable that some barriers are more frequently cited in the literature. Barriers B1 through B6 appeared at least 14 times in the literature (>50%), displaying higher confidence that data science projects face these barriers independent of context. Of these, B1, B3, B4, and B6 are also the most cited as critical. Barrier B1 (Insufficient skills) is the most frequent, both in the number of citations (20) and the number of times the barrier was identified as critical (5).

4.2. Clustering the Barriers

The input for the hierarchical clustering is the distance matrix, which determines how dissimilar two barriers are. This is based on the categorizations made by the studies included in the SLR: a relation between two barriers is formed when they belong to the same group. The greater the number of common groups to which a pair of barriers belongs, the greater the strength of the relationship, or similarity, that these barriers have. Conversely, the smaller the number of groups in common, the greater the distance, or dissimilarity, between the barriers. Jaccard distance, presented in the methods section, was the measure of dissimilarity used.

The distance matrix is shown in

Table 5. Distance matrix. Values represent Jaccard distance between one barrier and another, where 1 = no similarity (no relation) and 0 = maximum similarity. Note: the “B” is hidden starting from B10.

	B1	B2	B3	B4	B5	B6	B7	B8	B9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26
B2	1.0
B3	1.0	0.7
B4	1.0	0.6	0.9
B5	0.6	1.0	1.0	1.0
B6	0.8	1.0	1.0	0.9	0.8
B7	1.0	0.6	0.8	0.7	1.0	1.0
B8	1.0	0.3	0.8	0.7	1.0	1.0	0.4
B9	0.9	1.0	1.0	1.0	1.0	0.9	1.0	1.0
B10	1.0	0.9	0.6	0.9	1.0	1.0	0.8	1.0	1.0
B11	0.9	1.0	1.0	1.0	0.9	0.6	1.0	1.0	1.0	1.0
B12	0.9	0.9	1.0	0.9	1.0	0.8	1.0	0.9	1.0	1.0	0.9
B13	1.0	0.6	0.7	0.7	1.0	1.0	0.6	0.7	1.0	0.6	1.0	1.0
B14	1.0	1.0	1.0	0.9	1.0	0.8	1.0	1.0	1.0	1.0	0.8	0.7	1.0
B15	0.9	1.0	1.0	1.0	0.9	0.9	1.0	1.0	0.9	1.0	0.9	1.0	1.0	1.0
B16	0.7	1.0	0.9	1.0	0.9	0.9	1.0	1.0	0.9	0.9	1.0	0.9	1.0	0.9	1.0
B17	0.8	1.0	1.0	1.0	1.0	0.9	1.0	1.0	0.3	1.0	1.0	1.0	1.0	1.0	0.9	0.9
B18	1.0	1.0	1.0	0.9	1.0	1.0	1.0	1.0	0.9	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0
B19	1.0	0.8	0.9	0.9	1.0	1.0	0.9	0.8	1.0	1.0	1.0	0.9	0.9	1.0	1.0	1.0	1.0	1.0
B20	0.9	1.0	1.0	1.0	1.0	0.9	1.0	1.0	1.0	1.0	0.8	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0
B21	1.0	0.6	0.9	0.9	1.0	1.0	0.9	0.7	1.0	0.8	1.0	0.9	0.9	1.0	1.0	1.0	1.0	1.0	1.0	1.0
B22	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	0.6	1.0	1.0	1.0	1.0	1.0	1.0
B23	1.0	1.0	1.0	0.9	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	0.3	1.0	1.0	1.0	1.0
B24	1.0	1.0	1.0	1.0	1.0	0.9	1.0	1.0	1.0	1.0	0.8	0.9	1.0	0.8	0.8	1.0	1.0	1.0	1.0	1.0	1.0	0.8	1.0
B25	0.9	0.9	0.9	0.9	0.8	0.9	0.9	0.9	1.0	1.0	0.8	1.0	1.0	1.0	0.8	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0
B26	0.9	1.0	1.0	1.0	0.9	0.9	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	0.9	1.0	1.0	1.0	1.0	0.8	1.0	0.6	1.0	0.8	1.0
B27	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	0.8	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0

, and the network of relationships created from it can be seen in Figure 3. Visually, it is possible to see that the barriers cluster naturally into similar groups.

The hierarchical clustering of the barriers was accomplished with the average linkage method, using Jaccard distance as a measure of dissimilarity. The average linkage method tends to generate better results [41].

The number of clusters was determined using a combination of the elbow method and the silhouette score. The elbow method uses the average within-cluster distance to look for an elbow, that is, the point at which a further increase in the number of clusters has a diminishing decrease in the average distance. The silhouette score compares between dissimilarity to within dissimilarity and is a normalized value between −1 and 1. The number of clusters chosen should maximize the silhouette score [45]. Looking at Figure 4, the elbow is formed at six clusters, and the silhouette is maximized at six and eight clusters. Therefore, six was the chosen number of clusters; that is, within the range for the studies (between two and eight clusters). The clustering can be seen in the dendrogram in Figure 5.

In

Table 6. Clusters and centrality measures.

#	Cluster	Closeness	Degree
	People
B1	Insufficient skills	1.38	11
B5	IT illiteracy	1.28	7
B16	Inadequate training programs and facilities	1.16	9
B25	Uncertainty about benefits	1.10	10
	Data and technology
B2	Poor data quality	1.57	10
B8	Lack of an integrated data environment	1.52	9
B7	Complexity of data or technology	1.42	9
B13	Scalability issues	1.42	8
B3	Insufficient IT infrastructure	1.29	10
B4	Data privacy and security	1.27	14
B21	Data availability	1.22	8
B10	Immature technology and lack of appropriate tools	1.20	7
B19	Poor data management and architecture	1.10	7
	Management
B6	Lack of support from top management	1.31	14
B11	Strategy mismatch	1.27	9
B14	Resistance to change and other cultural barriers	1.22	6
B12	Inadequate data sharing policy	1.17	11
B20	Lack of coordination, collaboration, and communication	1.09	4
	Economic
B9	Insufficient funding	1.72	7
B17	High investment and maintenance cost	1.50	5
B27	Associated risks	1.09	1
	Project
B22	Inadequate or inconsistent methodology	1.54	3
B15	Insufficient ROI or business case	1.33	10
B26	Deployment and maintenance issues	1.33	7
B24	Scope, objectives, or expected results unclear	1.28	7
	External
B18	Government policies and regulation	4.00	3
B23	External sources of data	4.00	2

, the six clusters are presented, as well as the centrality measures for each of the barriers. The clusters were named based on the common theme between the barriers in the group, especially those considered most central to the cluster.

5. Discussion

Analyzing the Jaccard distances, some highly correlated pairs stand out, for example: barriers B2 (Poor data quality) and B8 (Lack of an integrated data environment) are frequently grouped in the same category, as well as B9 (Insufficient funding) and B17 (High investment and maintenance cost). The themes of the pairs are similar, which was the expected result.

Barriers B1 (Insufficient skills), B2 (Poor data quality), B6 (Lack of support from top management), B9 (Insufficient funding), and B22 (Inadequate or inconsistent methodology) were the barriers with the highest closeness centrality values of their respective clusters and can therefore be considered as the geometric center of each cluster. Again, barriers B1, B6, and B9, and additionally, barriers B4 (Data privacy and security), B15 (Insufficient ROI or business case), and B18 (Government policies and regulation), were the barriers with the highest degree in each cluster. Degree measures the number of vertices adjacent to a vertex, and thus these barriers have a greater number of relationships, internal or external to the cluster. Both measures denote greater relevance to these barriers in relation to the others, although from different points of view, and could be considered as potentially having a greater overall importance to the success of DS projects. Compared to the frequency of citations in the SLR, 4 these barriers are also among the most cited (B1, B2, B4, and B6).

The vertex of greatest closeness is not always the vertex of greatest degree—while B2 (Poor data quality) represents the geometric center of the cluster of data and technology barriers, B4 (Data privacy and security) is the highest degree in the cluster, as it has more connections to barriers from different clusters, such as barrier B18 (Government policies and regulation), for example. This suggests that data privacy and security in a DS project are related to government policies and regulations. In fact, the attention given to the topic of data privacy is greatly influenced by local regulations. While Europe has the extensive and restrictive GDPR (EU General Data Protection Regulation), in the US, data privacy is regulated by a multitude of federal regulations and even state legislation, such as California’s CCPA (California Consumer Privacy Act) [68].

At the top of people barriers is insufficient skills. DS is a skill-intensive field, requiring knowledge in statistics, ML, data wrangling, privacy regulations, domain expertise, and others [12], and therefore requires professionals with those specific skills.

For the management barriers, the most central is a lack of support from top management. This is likely to hinder success because a project requires continuous resource investment and inter-organizational collaboration [26]. Similarly, if there is a mismatch between strategy and project goals, it is unlikely to maintain investment and collaboration.

On the side of data and technology, data quality comes as the most frequently cited and as the most central to the cluster. Throughout the data lifecycle—from generation to processing, storage, and consumption—many factors can impact data quality, such as collection errors, lack of validation, mismanagement, etc. Data quality is commonly understood as the degree to which it is fit for use, that is, it must be accurate, relevant, and easy to interpret [69]. Without those qualities, it is hard to expect success in a project with such a high reliance on data. Besides managing poor data quality, the lack of an integrated environment can be very time-consuming [63] since this adds the complexity of having to merge and normalize data from many different sources.

Economic factors also play an important role. Big Data applications require investing in costly IT infrastructure, human resources, and tools [38]. Without financial support, higher-cost options such as leveraging cloud computing and increasing data generation can be vetoed, negatively impacting project outcomes [25].

Most central to project barriers is inadequate or inconsistent methodology. Besides the issue of not following standard practices [58], current methodologies lack integrity—which, according to Martinez et al. [29], is failing to properly address project management, team management, or data management. Moreover, a fourth aspect—change management—is notably ignored [30], which could lead to sustainability issues.

Lastly, organizations running DS projects face external barriers, which include government policies and difficulties in acquiring external data sources for usage. As stated, governments can impact data access and usage through privacy regulations, but also through other means, such as the availability of public data, public IT infrastructure, and collaboration with the private sector [26].

The centrality rankings are broadly consistent with the findings of previous reviews where impact or criticality was measured, which have frequently identified the lack of top management support, insufficient funding, and poor data quality as critical. The clusters also align broadly both in number of clusters—between two and eight for the studies included—and themes, where technology, data, people, organization, and cost-related are the most common. The differences between all studies, including this paper, mostly arise in how individual barriers are clustered. While B4 (Data privacy and security) is usually categorized in data-related themes, Barham and Daim [64] classify it as legal-related, reflecting the different perspectives that can be taken into account. Beyond the comprehensive list of barriers, the main contribution of this study is quantitatively derived categories, considering the classifications from all included studies, each with a central barrier identified through network centrality.

Organizations undertaking DS projects should expect to encounter one or many of the barriers listed here. Proactively addressing these issues, such as establishing data collection processes to avoid data quality issues, ensuring collaborators have the needed skills, and offering support from top management, is essential to improving project prospects. However, no amount of preparation can fully avoid all problems, and managers must act decisively. The six barrier categories proposed in this study provide a structured framework to guide managerial thinking, helping identify where issues are likely to arise and what areas require attention. When resources are limited, the centrality measures can serve as a useful tool for prioritization.

Limitations

The studies included in this review varied in focus, context, and terminology. Most of the evidence was qualitative or descriptive in nature, which is appropriate for identifying barriers but may affect the comparability of findings. Additionally, as many findings relied on participants’ perspectives, they reflect subjective experiences that are often context-specific, limiting generalizability. The methodological rigor of MCDM studies, and the number of studies from multiple contexts, as seen in Figure 2, help mitigate these limitations, but not entirely.

While the review followed established PRISMA guidelines, it relies on manual processes for screening and data extraction, which, while thorough, carry a small risk of oversight. Despite efforts to conduct a comprehensive search, it is possible that some relevant studies were missed.

6. Conclusions

The SLR resulted in the identification of 27 barriers to DS project success referenced by multiple studies in the literature. These barriers were grouped into six thematic clusters: people, data and technology, management, economic, project, and external barriers.

Barriers B1 (Insufficient skills), B2 (Poor data quality), B3 (Insufficient IT infrastructure), B4 (Data privacy and security), B5 (IT illiteracy), and B6 (Lack of support from top management) were the most frequently identified in the literature, being cited by 14 or more studies (>50%). Of these, B1, B3, B4, and B6 were also the most frequently cited as critical.

B1 (Insufficient skills), B2 (Poor data quality), B4 (Data privacy and security), B6 (Lack of support from top management), B9 (Insufficient funding), B15 (Insufficient ROI or business case), B18 (Government policies and regulation), and B22 (Inadequate or inconsistent methodology) were identified as being the most central, or of greater importance, based on measures of network centrality. The intersection between the barriers most frequently cited in the literature and those are barriers B1, B2, B4, and B6.

The distance matrix constructed from the categorizations of the studies included in the SLR and used for clustering the barriers also provides a proxy for the correlations between the barriers, which can serve as a basis for building hypotheses of interrelationship and causality—that is, how the barriers relate to each other and whether one set of barriers causes others. If the lack of an integrated data environment (B8) causes problems in data quality (B2), then solving the former will also solve the latter. This type of analysis is important for understanding failures in DS projects and could be explored in future studies.

We hope the results of this study will serve as an extensive yet specific survey of the barriers that hinder DS projects. Furthermore, this study used a quantitative method for clustering the barriers, helping solve the challenge of multiple answers to the research question. The categorization helps with the understanding of the barriers to the success of DS projects, and the centrality measures provide information about the relative importance of each.