Information systems are backed up by data and rely on them. The quality of information systems depend on a proper and accurate data management: the design of data management cycles should consequently be supported by sound quality analyses which may identify risks and qualify the phases in terms of quality dimensions. In this paper we present an analysis framework for the quality of data management cycles. Quality and efectiveness of information systems depend on a sound, consistent, and well posed management of the data sources on which they rely. Being those data lakes, big data repositories, relational or non-relational databases or other sources, a proper design, operation, and evolutionary maintenance of the data processing system is the fundamental basis to support the applications composing the information system. Complexity derives not only from data volumes, source and nature heterogeneity and technological articulation, but from non-functional specifications either of the data processing system or of the applications. A structured approach to the choice of the right data management cycle should include a correct articulation in phases and a proper evaluation of the characteristics of each phase which can support the compliance with the desired non-functional requirements in each epoch of its lifetime, from the bootstrap of the system to its decommissioning. The adoption of well-known data management cycle may provide a good reference for designer and administrators to dominate complexity. Literature ofers several diferent Data Lifecycle Models (DLM) for data management [ 1, 2], but does not specifically focus on a structured approach to the choice. We proposed in [ 3] an approach for the selection of the most appropriate DLM given an overall qualitative evaluation of the relevant desired characteristics by leveraging a panel of experts: in this paper we focus on a framework for the qualification of each DLM phase to provide a finer grain support for DLM design.
A data lifecycle model (DLM) is a conceptual framework that describes the stages through which data passes during its existence, from initial creation or capture to its eventual deletion or archival [4]. The model typically includes phases such as data generation, collection, storage, processing, analysis, sharing, and disposal. It helps organizations manage data systematically to ensure data quality, compliance, security, and efective usage. A typical DLM is shown in Figure 1.
In scientific and technical literature there are several data lifecycle models designed to deal with diferent data management issues or particular case studies. Some DLMs, such as USGS [ 6], DATAOne [7] or IBM DLM [8], have been widely used by a number of research group in their activity of data management in various diferent disciplines.
The availability of many best practices that could be used as Standard for the data lifecycle model could improve significantly the aspects related to data quality. Using that DLMs could have the following positive aspects: • Clarity: literature models provide a clear, structured, and often visualized framework that helps clarify the stages of data management, making it easier for teams to understand and communicate processes [9]; • Best Practices and Compliance: these models are typically built on established best practices and are designed to support regulatory compliance, data quality, and security standards [10]; • Eficiency and Consistency: standard models reduce ineficiencies and risks associated with ad hoc approaches, supporting optimized data retention, cost savings, and improved decisionmaking; • Broad Applicability: literature models are often designed to be widely applicable, providing a high-level framework that can be adapted across diferent domains and organizations [11]; • Support and Documentation: here is often extensive documentation, community support, and existing tools aligned with standard models, making implementation and troubleshooting easier. On the other side, there is also to consider some drawbacks: • Lack of Customization: literature models may not address specific organizational needs or unique workflows, potentially leading to gaps in coverage or relevance for specialized projects [11], [9]; • Oversimplification: standard models may mask the real-world complexity of data processes, presenting them as linear or closed systems, which can be misleading in dynamic or multidirectional research environments [9] [12]; • Rigidity: adhering strictly to a standard model can limit flexibility and innovation, especially when dealing with novel data types or emerging business requirements.
AHP is a structured technique for organizing and analyzing complex decisions proposed by Saaty in 1987 [13] [14] [15]. It works by breaking down a problem into a hierarchy of criteria and alternatives, then making pairwise comparisons to establish priorities.
Firstly, it is necessary to structure the decision problem as a hierarchy, so it is decomposed into three primary levels: 1. Goal: the overall objective; 2. Criteria: the factors contributing to the goal, {1, 2, . . . , }; 3. Alternatives: the choices to be evaluated, {1, 2, . . . , }
Goal →− { 1, 2, ..., } →− { 1, 2, ..., }
For each element in a level, a pairwise comparison matrix is constructed. The entries are based
on the Saaty’s 1-9 scale [14] of relative importance.
(
The matrix has two key properties: i) Reciprocity, = So, the matrix can be represented as:
The priority vector w, which represents the relative weights of the elements being compared, is found by solving the eigenvalue problem:
w = maxw where max is the principal (largest) eigenvalue of matrix . and w is its corresponding right eigenvector, normalized so that its elements sum to 1.
⎡11 ⎢21 = ⎢⎢⎣ ...
12 · · · 1 ⎤ 22 · · · 2 ⎥
... . . . ... ⎥⎥⎦ 1 2 · · · ⎡ 1 ⎢1/12 = ⎢⎢⎣ ...
1/1 12 1 . .
. 1/2 · · · · · · 1⎤ · · · 2⎥ . . . ... ⎥⎥⎦
1 ⎛1 ⎞ ⎜2 ⎟ w = ⎜⎜⎝ ... ⎟⎟⎠ with
∑︁ = 1 =1
In practice, the eigenvector is often approximated using the geometric mean method, which is simpler and yields very similar results. The steps are: • Calculate the geometric mean ˜ for each row : • Normalize the geometric means to get the priority vector w:
⎛ ˜ = ⎝∏︁ ⎠ =1
⎞1/ ˜ = ∑︀
=1 ˜ • This gives the approximated priority vector w = (1, 2, . . . , ) .
AHP measures the consistency of the decision-maker’s judgments using the Consistency Ratio (CR). To do that, firstly Consistency Index have to been computed using the following formula: (8) (9) (10) (11) CI = max −
− 1 It is possible to calculate the Consistency Ratio (CR)
CR =
CI
RI
In which the Random Index (RI) is extracted from the known table for the matrix size as described in [14]. A CR value of 0.10 or less is considered acceptable, while if CR > 0.10, the judgments may be too random and should be revised.
The final step is to aggregate the weights from all levels of the hierarchy. Obtain the priority vector for alternatives under each criterion (walt for criterion ) and the priority vector for the criteria themselves (wcrit). For each alternative , its final global priority is calculated by summing the products of its local priority under each criterion and the weight of that criterion.
= ∑︁ (local weight of under ) × (weight of )
=1 Or, in vector form for all alternatives:
G = local · w crit where the columns of local are the local priority vectors for each criterion.
The alternative with the highest global priority is the preferred choice.
AHP-Express is a variant of the well-known AHP method proposed in [16]. The core diference between the two methods lies in how they handle the evaluation of alternatives, which leads to significant diferences in the number of required judgments and the underlying mathematical procedure.
The AHP-Express method simplifies the original AHP by replacing the matrices for alternatives with direct rating. For each criterion , the decision-maker directly assigns a local priority to each alternative ; the scale may be the same Saaty 1-9.
These ratings are then normalized per criterion to create the local priority vector for each criterion. The most common normalization is the additive normalization:
local = ∑︀=1 for = 1, 2, ..., (12) where is the rating of alternative under criterion . The other steps are identical to the original AHP method.
Pointed out that using standard DLMs could significantly boost data quality in a research project, it is crucial to determine how to choose the best suited Data Lifecycle Model. first of all, it is necessary to detect the various issues related to data quality and to refer in which phase of the DLM are they related. From an accurate analysis of scientific literature, has been populated the following Table 1 [ 17], [18], [19], [20].
- Incomplete data - Biased sampling - Format inconsistencies - Data corruption - Unauthorized access - Redundancy - Transformation errors - Loss of data fidelity - Inconsistent aggregations - Misinterpretation - Outdated data - Sampling bias - Data degradation - Compliance violations - Unintended retention
In the Table below several data quality risks - and the related mitigation strategies - have been highlighted. A possible strategy devoted to select the best alternative standard DLM for the given problem is to compare the various DLM to find if there are the right mitigation strategies in every examined phase.
The comparison between DLMs could be evaluated using widely used Multi-criteria decision-making (MCDM) methods, such as AHP or TOPSIS.
There is also to underline that qualitative issues in DLM apply on a vertical and on a horizontal direction: the vertical direction concerns a single phase of the DLM and afects all applications which involve that phase; the horizontal direction concerns all phases involved by a single application. Moreover, depending on how the data lifecycle works in the real case (e.g. sporadic queries; continuous data flow; data burst, etc.) there are diferent paths in DLM phases.
In other words, there are two diferent series of parameters to be applied in the comparison: 1. set of parameters based on how the specific phase of the specific DLM performs in the requirements listed in Table 1; 2. set of parameters based on how the specific phase of the specific DLM is crucial in the contest of the whole lifecycle (relative importance of the single phase).
A well-known MCDM method that could be used in this case is the Analytic Hierarchy Process (AHP) by Saaty [14]. In this method, are to be populated the Decision Matrix A[i,j] and the Weight Vector w(k); the Decision Matrix has to be populated using the first set of parameters while the second set of weights will converge in Weight Vector.
To validate what discussed so far, we applied it to two case studies (Section 5) and the evaluated alternatives using the AHP process. The decision framework is based on the AHP-Express variant, which reduces elicitation by fixing a reference criterion and comparing each remaining criterion against it according to Saaty’s 1–9 scale. The tool requires as input a CSV/Excel file with DLMs as rows and criteria (factors) as columns, computes criterion priorities and DLM scores, and returns the final ranking together with bar and radar charts for visual and quick highlights. For a detailed discussion of the tool, please refer to the following paper [3].
In order to practically illustrate the approach, two appropriate case studies has been analyzed. Firtly, it is necessary to select a series of DLMs to be used as alternatives in AHP method as described in Section 3.
The DLMs included in out assessment are: • DCC Curation Lifecycle: DCC is a data life cycle model that allows data to be managed and preserved efectively, from its creation or receipt to its final preservation. It involves a planning phase in which strategies are implemented to preserve digital material throughout its life cycle, for example through the use of standards and technologies. In fact, it includes plans for the management and administration of all actions in the preservation life cycle.[21] • DDI Lifecycle: this DLM consists of eight phases (concept, collection, processing, archiving, distribution, discovery, repurposing) in which attention is paid to reducing conceptual inconsistency in data through the use of ontologies in collection and the importance of archiving them in order to enable data reuse by the user. This dlm was created to address the lack of work on ontological dynamics in a manner consistent with the data lifecycle. Ontologies enable the promotion of data management, standardisation and integration.[22] • CIGREF Lifecycle: It is a data management-oriented DLM. It includes phases such as collection, consolidation and structuring, followed by distribution and exploitation of data. In our opinion, it pays particular attention to the governance, compliance and security/privacy requirements typical of corporate contexts. [23]. • CRUD Lifecycle: This is a DLM that involves the adoption, for each of its five phases, of strategies aimed at reducing data security risks. The phases of creation, Storaging and Destruction are mandatory, while Use and Share and Archive phases are optional. [24] • DataONE Lifecycle:Although this DLM comprises numerous stages, it cannot be defined as complex. In each stage, diferent people can interact directly with the data, and it is unlikely that a single person will interact with the data in all stages. The data life cycle is useful because it can be used to identify data flows and work processes for scientists, librarians, or others associated with the scientific data process. It is not suitable in the context of Big Data. [7]
The criteria has been chosen as in [3], and represents the diferent phases of DLMs.
Subsequently, the values for all criteria have been assigned by a team of experts using Saaty scale (1 to 9) according to the the comprehensiveness of the quality requirement as described in Table 1. The assigned scores are shown in Table 5.
We designed two case studies that stands as complementary viewpoints in data governance and quality, and are related to the management of water data. The only diference across cases is the reference criterion used during pairwise elicitation, which encodes the perspective of the decision maker.
In this first case study we took the perspective of an authority responsible of certifying the quality of water resources. Because early lifecycle failures spread and are expensive to fix, upstream governance—procedures, sampling plans, legal compliance, and stakeholder coordination—is essential in such a situation. In order to align the comparisons with a pre-analytical quality emphasis, we thus choose Starting as the AHP-Express reference criterion. What we expected from this first trial was that models that delay quality controls to later phases are expected to receive lower relative priority in this perspective, which tends to favor DLMs that explicitly address planning, provenance capture, protocol standardization, and intake controls when using Starting as the reference. The obtained results are presented in Sec. 5.1.1
In Table 3 are reported the criterion weights where is possibile to notice that most of the importance is concentrated on Starting and Administration (each ≈ 0.332 ), followed by Computation and Security (≈ 0.111 ), while Assessment and End-of-life are comparatively less emphasized. The resulting ranking (Table 4) places DCC first ( 7.174), followed by CIGREF (6.424), CRUD (6.383), DDI (6.313), and DataOne (2.968). A visual of the final rank, if presented in Fig. 3.
The radar plot in Fig. 4 compares the normalized profiles of DLMs over phases. Here, DLMs with robust profiles over phases tends to prevails over the others. Specifically, DCC has good benefits in Starting and Administration, while CIGREF is very strong in Computation/Administration and presents null or lowers scores in Security and End-of-life, which penalizes the overall evaluation over such phase. CRUD excels in Security and End-of-life, but these two have lower weights than the two main ones; DataOne is ranked as last.
The radar aligns between the models profile and the weight structure visible in the final ranking bar plot.
Here, we assume a generic data user/company who uses data that has already been certified by third parties. Transformation quality, computational reproducibility, and processing pipeline performance/traceability become the primary risk factors. Thus, we chose Computation as the reference criterion for AHP-Express.
DLMs that specify processing standards, validation during transformations, versioning, reproducibility, and auditability of computational steps should be given preference in this viewpoint. Strong intake models with less information about processing controls might be ranked lower overall. The obtained results are discussed in Section 5.2.1
The criterion weights are reported in Table 5. here, most of the importance is concentrated on Computation (≈ 0.299 ), Starting(≈ 0.256 ), and Security (≈ 0.224 ), while Assessment (≈ 0.107 ), Administration (≈ 0.064 ), and Endof-life (≈ 0.050 ) are comparatively less emphasized. The resulting ranking (Table 6) places DCC first ( 7.740), followed by CRUD (6.313), CIGREF (6.021), DDI (5.836), and DataOne (4.868). A visual of the final rank is presented in Figure 5.
From a visual inspection of the radar plot in Figure 6 we can see that (paying focus to the reference criteria Computation), models that are strong on Computation and Security tend to prevail, provided they do not show structural gaps on the remaining axes. DCC exhibits a balanced and wide polygon, combining high Computation with strong Starting/Administration. CRUD excels on Security and End-of-life—axes that carry non-negligible weight here, so it overtakes CIGREF, which is very strong on Computation/Administration but shows null values on Security/End-of-life. DDI, strong on Starting/Assessment but only moderate on Computation. DataOne improves, compared to the previous case study, thanks to Computation, yet it remains last because of gaps on Security/End-of-life. The radar, also here aligns with the profile of the models with the weight structure and explains the ordering observed in the final scores bar plot.
In this paper a possible framework to be used to improve data quality has been presented, using AHP method to select the best alternative DLM to be used, having in mind the data quality improvement. Future work will involve the implementation of a series of test cases to assess the relevance of the results.
By comparing the results obtained in the two presented case studies, is possible to note that there is a shit in the final rank of the DLMs, where CRUD and CIGREF basically switches. This behavior is due to the fact that who produces the data gives more importance to phases weighted diferently in the two models, i.e. Starting/Administration, whereas who uses the final data may give more importance to phases like Computation.
The author(s) have not employed any Generative AI tools. [7] S. Allard, Dataone: Facilitating escience through collaboration, Journal of eScience Librarianship (2012). [8] IBM Software, Wrangling big data: Fundamentals of data lifecycle management, Technical Report,
IBM, 2013. [9] A. M. Cox, W. W. T. Tam, A critical analysis of lifecycle models of the research process and research data management, Aslib Journal of Information Management 70 (2018) 142–157. [10] J. Zhang, J. Symons, P. Agapow, J. T. Teo, C. A. Paxton, J. Abdi, H. Mattie, C. Davie, A. Z. Torres, A. Folarin, et al., Best practices in the real-world data life cycle, PLOS digital health 1 (2022) e0000003. [11] A. Sinaeepourfard, X. Masip-Bruin, J. Garcia, E. Marín-Tordera, A survey on data lifecycle models:
Discussions toward the 6vs challenges, Technical Report (UPC-DAC-RR-2015–18) (2015). [12] T. Weber, D. Kranzlmuller, Methods to evaluate lifecycle models for research data management,
Bibliothek Forschung und Praxis 43 (2019) 75–81. [13] R. W. Saaty, The analytic hierarchy process—what it is and how it is used, Mathematical modelling 9 (1987) 161–176. [14] T. L. Saaty, Decision making with the analytic hierarchy process, International journal of services sciences 1 (2008) 83–98. [15] M. Tavana, M. Soltanifar, F. J. Santos-Arteaga, Analytical hierarchy process: Revolution and evolution, Annals of operations research 326 (2023) 879–907. [16] J. E. Leal, AHP-express: A simplified version of the analytical hierarchy process method, MethodsX 7 (2020) 100748. [17] C. Batini, M. Scannapieco, DATA AND INFORMATION QUALITY: Dimensions, Principles and
Techniques, SPRINGER, 2018. [18] M. Ghasemaghaei, G. Calic, Assessing the impact of big data on firm innovation performance: Big data is not always better data, Journal of Business Research 108 (2020) 147–162. URL: https://www. sciencedirect.com/science/article/pii/S0148296319305740. doi:https://doi.org/10.1016/j. jbusres.2019.09.062. [19] K. Lee, N. Weiskopf, J. Pathak, A framework for data quality assessment in clinical research datasets, in: AMIA Annual Symposium Proceedings, volume 2017, 2018, p. 1080. [20] L. L. Pipino, Y. W. Lee, R. Y. Wang, Data quality assessment, Communications of the ACM 45 (2002) 211–218. [21] S. Higgins, The dcc curation lifecycle model, International Journal of Digital Curation (2008). [22] X. Ma, P. Fox, E. Rozell, P. West, S. Zednik, Ontology dynamics in a data life cycle: challenges and recommendations from a geoscience perspective, Journal of Earth Science 25 (2014) 407–412. [23] S.BOUTEILLER, Business data challenges. How to manage company data to create value?, Technical
Report, CIGREF, 2014. [24] X. Yu, Q. Wen, A view about cloud data security from data life cycle, in: 2010 International Conference on Computational Intelligence and Software Engineering, 2010, pp. 1–4. doi:10.1109/