A Reference Model for Product Data Proﬁling in Retail ERP Systems

Rolf Krieger and Christian Schorr

Institute for Software Systems, University of Applied Sciences Trier, Birkenfeld, Germany

Keywords:

Data Quality, Data Proﬁling, Exploratory Data Analysis, Product Data, Enterprise Resource Planning.

Abstract:

Due to the high volume of data and the increasing automation in retail, more and more companies are dealing

with procedures to improve the quality of product data. A promising approach is the use of machine learning

methods that support the user in master data management. The development of such procedures demands

error-free training data. This means that product data must be cleaned and labelled which requires extensive

data proﬁling. For typical retail company data bases with usually complex and convoluted structures this

exploration step can take a huge and expensive amount of time. In order to speed up this process we present a

reference model and best practices for the systematic and efﬁcient proﬁling and exploration of product data.

1 INTRODUCTION

Product data quality is an increasingly important topic

in the retail industry. The automation of purchasing,

sales, delivery and inventory processes can only suc-

ceed if the referenced product data are error-free. Es-

pecially for e-commerce where the decision to buy a

product can only rely on its virtual presentation in a

web shop, correct product descriptions are of utmost

importance. Moreover, highly automated storage cen-

tres where robots move products from the shelves to

the packaging area require precise data about physical

dimensions and weight in order to ensure a ﬂawless

delivery to the waiting customer. If the quality of the

product data is poor, delays in delivery, losses due to

incorrect prices, additional work etc. are the result.

Manual data quality control is nearly impossible

due to the vast amount of products. Current semi-

automated quality control systems based on data qual-

ity rules are prone to miss erroneous data due to

the vast amount of attributes, tables and their com-

plex relations. Machine learning provides a promis-

ing approach to support the user to ensure data qual-

ity. But major prerequisites for machine learning are

high quality test and training data (Domingos, 2012).

Consequently, at the beginning of a machine learning

project the quality of product data has to be estimated

and product data have to be cleaned. In addition do-

main understanding is of importance and can only be

gained by a careful and detailed exploration of the

data. Intensive data proﬁling is therefore necessary in

order to gather meta data and to get acquainted with

the product data.

To ease this central task of product data quality

management, we propose a reference model which

deﬁnes a reference process and a high-level architec-

ture model of a data proﬁling tool taking into account

the speciﬁc properties and requirements of product

data. The reference process should allow to identify

critical data records efﬁciently. Based on our results

the process of proﬁling product data should be accel-

erated signiﬁcantly.

The paper is structured as follows: Section 2 de-

scribes the goals, tasks and state of the art of data pro-

ﬁling while section 3 explains product data. In Sec-

tion 4, we introduce our reference process for proﬁl-

ing product data. In section 5, we present best prac-

tices to support the proﬁling steps and illustrate them

using examples. Finally, we summarize the main re-

sults and give an outlook on our further research.

2 STATE OF THE ART

Data proﬁling is the application of data analysis with

the purpose of determining information about the con-

tent, structure and quality of data. According to (Ol-

son, 2008), the input of the proﬁling process are data

and meta data about them. Meta data describe the

structure, content, history and quality of the data. In

practice meta data are usually incomplete. The output

of the data proﬁling process are information about er-

roneous data and complete meta data. Thus data pro-

ﬁling is carried out to gain business understanding and

Krieger, R. and Schorr, C.

A Reference Model for Product Data Proﬁling in Retail ERP Systems.

DOI: 10.5220/0007953303170324

In Proceedings of the 8th International Conference on Data Science, Technology and Applications (DATA 2019), pages 317-324

ISBN: 978-989-758-377-3

317

to provide information for data cleaning. Analysing

the structure of the data involves checking their con-

sistency and format. Computing descriptive statistical

information like minima, maxima and percentages as

well as determining data types and lengths falls un-

der this category. Proﬁling the data content identi-

ﬁes speciﬁc properties concerning missing values or

other errors. A crucial part of data proﬁling especially

with regard to subsequent feature engineering tasks in

machine learning projects is to discover how differ-

ent parts of the data are related to each other. Iden-

tifying embedded value dependencies and functional

dependencies between attributes or tables as well as

potential (foreign) keys are part of this data proﬁling

aspect. An excellent overview about the topic can be

found in (Abedjan et al., 2015). Data proﬁling is also

related to data exploration. A survey on data explo-

ration techniques can be found in (Guido et al., 2015;

Di Blas et al., 2014). Both commercial and open-

source software tools are available for data proﬁling

and cleaning. Often vendors of commercial ERP sys-

tems also have a corresponding data quality tools in

their portfolios. SAP sells its Information Steward,

Oracle its Enterprise Data Quality and IBM offers In-

foSphere Information Server. The company Informat-

ica specialized in business analytics sells Informatica

Data Proﬁling while SAS offers its DataFlux Manage-

ment Studio. Apart from their comprehensive scope,

the major advantage of these tools is their seamless

integration into the respective ERP systems of their

companies, which avoids migration and transforma-

tion problems. Their disadvantage is that they are

not optimized for the proﬁling of product data in gen-

eral. Beside these commercially available products a

plethora of open-source tools exists. A few mature

software packages are Metonome (Papenbrock et al.,

2015), developed and managed by the University of

Potsdam, Germany, Proﬁler (Kandel et al., 2012) and

Talend Open Studio. The drawbacks of free data pro-

ﬁling software usually lie in their narrow focus on

speciﬁc applications especially in case of more sci-

entiﬁc origins. In contrast to these tools we propose a

data proﬁling tool and reference process speciﬁcally

taking into account domain knowledge about product

data.

3 PRODUCT DATA

3.1 General Description

Product data are of utmost importance in retail com-

panies. For technical and business reasons, products

are classiﬁed into categories arranged in a multi-level

merchandise category hierarchy. Moreover, there are

a variety of different product types like simple prod-

uct items, grouped products (e.g. displays, sets etc.)

and generic products and their variants. Product at-

tributes are referenced by many processes in purchas-

ing, material requirements planning, inventory man-

agement and sales. Table 1 shows some attributes of

product data in a standard ERP software. We can dis-

tinguish between descriptive and process-related at-

tributes.

The values of descriptive attributes are the same

for all locations (e.g. stores, warehouses) of a com-

pany whereas process attributes may be maintained

differently. For example, the safety stock may be 10

pieces in one store and 15 pieces in another. Process-

related attributes control the processes in a company

and are decisive for process automation. Usually the

product attributes provided in a standard ERP soft-

ware do not meet all requirements of a company.

Therefore, they are supplemented by custom-speciﬁc

attributes, which can be either descriptive or process-

related.

Mostly, product attributes are also classiﬁed in the

ERP software thematically. Typical classes are basic

data, logistic data, purchasing data etc. Basic data

include the product description, packaging units, di-

mensions, volumes, gross and net weights, hazardous

substance codes etc. Logistic data include safety

stock, service level, delivery time, goods receipt pro-

cessing time etc. Purchasing and sales data include

prices, minimum order quantities, delivery periods

etc.

The attributes that have to be maintained for a spe-

ciﬁc product item depend on its category and type.

Consequently, the product categories and types have

to be taken into account in the data proﬁling process.

Table 1: High-level classiﬁcation of product attributes in a

ERP system.

descriptive process-related

standard

- product description

- category

- volume

- unit of measure

...

- reorder point

- safety stock

- goods receipt

processing time

...

custom-speciﬁc

- alcoholic strength

- description of

ingredients

...

- age limit for

sales

- ﬂags for logistic

processes

...

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

318

3.2 Case Study: SAP Retail

To get an idea of the amount and complexity of prod-

uct data managed in an ERP system, we conducted

a case study in two European full-range retail com-

panies. Both companies use SAP Retail as ERP sys-

tem to support their business processes. Table 2 gives

some information about the product items of the com-

panies.

Table 2: Rounded number of product categories, product

items and types of two retail companies.

Company 1 Company 2

Main categories 12 90

Sub categories 200 3.200

Items 750.000 ≥ 1.400.000

Product types 7 9

As in many other ERP systems, in SAP Retail the

attributes of products are divided into data segments,

e.g. basic data, purchasing data, sales data and logis-

tic data and are managed in more than 10 tables. Table

3 shows the number of attributes of the basic data ta-

ble. For instance, the basic data table of company 1

has 290 attributes. 80 attributes are in use. 42% of

them are custom-speciﬁc.

Table 3: Number of attributes of the basic data table in SAP

ERP of two companies.

Attributes Company 1 Company 2

Total 290 245

standard 206 209

custom 84 36

standard (used) 46 43

custom (used) 34 20

3.3 Technical Challenges

Due to the results of our case study, proﬁling and

cleaning of product data are difﬁcult tasks. We iden-

tiﬁed the following challenges:

3.3.1 High Number of Product Items

Big retail companies have more than 1 million items

in their ERP systems. Some of the attributes, e.g. lo-

gistic attributes, have to be maintained for each lo-

cation. In case of a company operating 50 stores 50

million logistic data records have to be maintained.

3.3.2 High Number of Product Categories

Some larger retail companies have more than 3.000

product categories. The category to which a product

item belongs often deﬁnes the attributes that have to

be maintained. Mandatory ﬁelds in particular are de-

ﬁned by the product categories. The large number of

product categories makes it difﬁcult to identify erro-

neous product data records.

3.3.3 Different Product Types

There are a variety of different product types. Similar

to the product categories they deﬁne the attributes that

have to be maintained for each item. Consequently,

the product items have to be partitioned by product

types for data proﬁling.

3.3.4 Age of the Data

ERP systems are often in use for years. Due to the

complex IT system landscape in a retail company,

product data are archived rarely. Hence, a lot of out-

dated items are still kept in the ERP system.

3.3.5 Business Processes Optimization

For business, technical and organizational reasons,

business processes are continuously optimized or

adapted in retail. Changing a business process often

requires the adjustment of process related attributes

of the product data. Mostly, outdated items are not

updated accordingly and do not fulﬁl current business

and technical requirements.

3.3.6 Complex Data Model

Product data are managed in many data tables in a

ERP system. In our case studies we identiﬁed more

than 10 tables. Some of these tables have a lot of at-

tributes (e.g. 80) of different types (e.g. numerical,

character, dates etc.). In case of some attributes spe-

ciﬁc knowledge of application experts is required to

understand their semantic and usage.

3.3.7 Changes to the Data Model

ERP systems are usually standard software. If a new

version of the software is developed by the software

manufacturer, this is often associated with changes to

the data model. On the one hand attributes may be

added, on the other hand existing attributes may be

no longer used. In addition, the data model is adapted

in a similar way by the customer. The data proﬁling

process has to take all these changes into account.

A Reference Model for Product Data Proﬁling in Retail ERP Systems

319

4 REFERENCE MODEL

In order to enable an efﬁcient proﬁling of product data

despite the challenges and problems formulated in the

last section, we have developed a reference model

consisting of a reference process and an architectural

model of a data proﬁling tool.

4.1 Reference Process

4.1.1 Data Partitioning

The main problem with proﬁling product data is the

amount and variety of data. In order to achieve mean-

ingful results efﬁciently, the data must be partitioned

appropriately for data proﬁling. Table 4 shows the

most important partitioning criteria by data segments.

Table 4: Partitioning criteria by product data segments.

1. Product type

2. Product category

Partition criteria

Logistic data

…

Basic data

Sales data

Purchasing

data

…

Product data segments

Partitioning criteria

4. Business

organisation

3. Location, e.g.

store, warehouse

In general, product data can be partitioned by type

and category. Especially, basic data can be partitioned

by product type and category. In the case of logistic

data, locations can be used for partitioning in addi-

tion. Purchasing and sales data can also be partitioned

by business department. As is well-known, partition-

ing leads to smaller amounts of data, reduces data

variability and simpliﬁes the data proﬁling.

4.1.2 Data Proﬁling and Exploration

For each element of the partition, we suggest a hierar-

chical approach to data proﬁling as illustrated in Fig-

ure 1. Using this concept, the product data are anal-

ysed and evaluated at the merchandise category level

in the ﬁrst step. This allows to detect anomalies by

comparing results for different categories with avail-

able meta data and expert knowledge. In addition, the

results of different categories are compared with each

other. Differences between categories often indicate

anomalies caused by erroneous data.

Due to the large number of attributes and cate-

gories, heat maps are often used in this step to visual-

ize the results. Examples are given in Section 5. The

information obtained in this way is used to check, cor-

rect and extend existing meta data.

The data proﬁling process can thus be executed

using a drill down approach. This means that the

results are ﬁrst analysed for main categories, then

for categories, subcategories and ﬁnally for product

items as illustrated in Figure 1. This allows the data

proﬁling to be performed efﬁciently even with large

amounts of data containing many attributes.

In addition to the hierarchical approach the refer-

ence process also proposes the order in which data

segments should be considered. Figure 2 shows the

order of data segments.

The basic data should be considered in the ﬁrst

step. Usually, they are independent of the locations,

e.g. stores and warehouses. In addition, the assign-

ment of the products to the stores and warehouses has

to be considered. This is necessary because the assort-

ments of the stores differ - not all products are sold in

all stores. Afterwards, the knowledge gained is used

to look at the logistic data of the products which de-

pend on locations. In many cases, the products are

transported to the stores via multi-level logistic net-

works. Therefore, the logistic data for the warehouses

should be considered before the logistic data for the

stores. Finally, the purchasing and sales data of prod-

ucts that are dependent on business partners are anal-

ysed.

For each step we propose to develop speciﬁc

scripts to run through the reference process in a time

and cost efﬁcient way.

4.2 High Level Architecture of the Tool

To evaluate the reference process, we have developed

a prototype of the tool as a package in the statisti-

cal programming language R. It offers functions and

scripts for product data proﬁling of the different data

segments and for the visualization of the results, espe-

cially for basic and logistics data. Figure 3 describes

the high level architecture and the application of the

package.

4.2.1 Data Import

Our tool contains scripts for importing the data and

meta data, building the base for data exploration. In

addition to the product data, we have also provided

the option of loading change documents that describe

changes of the product data depending on time and

transaction. When loading, the data types of the

ERP system are converted to the corresponding R data

types. In order to automate this process, the data

dictionary of the ERP system is accessed. The type

mapping and the encoding of missing values are pro-

vided in an XML document. Basic data cleaning steps

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

320

Figure 1: Description of the hierarchical top down data proﬁling process with example categories for a fashion retail data

base.

Figure 2: Order of data segments as considered in the data

proﬁling reference process.

Figure 3: High-level architecture model of our data proﬁl-

ing package based on R. Systems are shaded in grey.

are also performed. For instance, non-informative at-

tributes (completely missing attributes and constant

attributes) are identiﬁed and removed. The different

encodings of missing values (e.g. 00.00.0000, 0.000,

” ” etc.) are replaced by the value NA.

4.2.2 Reference Data

Reference data are required for certain proﬁling steps.

In particular the following reference data are required:

1. Merchandise Category Hierarchy.

To support the hierarchical data proﬁling ap-

proach the merchandise category hierarchy of

the company must be part of the reference

data. Additionally, the standard classiﬁcation

systems UNSPSC (www.unspsc.org), eCl@ss

(www.eclass.eu) and GPC (www.gs1.org) used in

many companies should be provided.

2. Conversion Factors for Units of Measurement.

Usually, the dimensions and volumes of prod-

ucts are given in different units of measurement.

A conversion function allows the user to easily

convert the values into a deﬁned target unit us-

ing conversion factors provided by the reference

data. Custom-speciﬁc types of packaging units

like pallets containing different amounts of pieces

are converted to the general basic unit pieces for

example.

3. Dimensions and Weights of Standardized Packag-

ing Units.

Packaging units such as boxes and pallets that are

referenced in product data are often standardised.

If their dimensions and weights are known, de-

tailed analyses can be carried out when proﬁling

the product data. For instance, a comparison of

the gross and net weight of a product in a partic-

ular packaging unit can be done in detail. Often

this reveals erroneous data records.

4.2.3 Data Proﬁling Scripts

Our package provides functions for analysing single

attributes, multiple attributes, and data rules. They

form the basis for the scripts developed for each data

A Reference Model for Product Data Proﬁling in Retail ERP Systems

321

segment. Bar charts and box plots are used to visu-

alize the results of univariate analyses. For bivariate

analyses scatter diagrams, heat maps, calendar-heat

maps and bag plots are used. For each data segment a

script deﬁnes the key ﬁgures and diagrams to support

the proﬁling of product data. The package is written

in R and uses custom-made code as well as standard

libraries. It is tailored to be used with data in SAP for

Retail format but can be adapted to work with other

systems, too.

5 BEST PRACTICES

The scripts are created with the process knowledge

of product data management experts and the techni-

cal knowledge of IT experts and correspond to best

practices for the proﬁling of product data. In the fol-

lowing we describe some of the scripts and give some

examples regarding the basic data segment.

5.1 Maintenance Status

An important step in data proﬁling is the analysis of

the data segments that are created for each product

item. Different segments are required depending on

the product type and category. A heat map showing

the frequency of data segments by type (or category)

can help to understand the dependencies and to iden-

tify erroneous items. Typically there are some items

with missing data segments.

5.2 Creation and Change Dates

As mentioned before, retail ERP systems often con-

tain data about products that are obsolete and do not

belong to the current product assortment. The cre-

ation date and the date of the last change of a data

record provide important information about this. Our

script determines the number of new product data

records by day, week and month. Based on a multi-

variate analysis the merchandise categories are identi-

ﬁed in which most products are created per day, week

and month. In addition the frequencies of attribute

values depending on the creation dates are analysed.

This provides information about product items and

categories that were created a long time ago and have

not undergone any changes. Together with technical

experts, it should be checked whether these products

can be excluded from further data proﬁling. In many

cases this leads to a considerable reduction in effort.

In our case study it was determined that only about

20% of the products belong to the current assortment.

5.3 Merchandise Categories

The distribution of product categories and their sub-

categories can easily be visualized using a tree map as

shown in Figure 4. It displays the grouping of prod-

ucts along their hierarchy levels. It is useful to get a

quick graphical overview about the distribution of the

product data by category. Additionally, it is helpful

in determining whether products are categorized cor-

rectly.

Figure 4: Tree map illustrating the frequencies of items by

category and subcategory using textile product data. All

articles are ﬁrst grouped by category and then subdivided

by subcategory. The category headwear for example has

the subcategories caps, bonnets and hats.

The purpose of this analysis is to determine the

main characteristics of the merchandise category hier-

archy, the number of main categories, categories, and

subcategories for example. As a rule, the individual

categories differ in the number of subcategories they

contain. The frequency of the subcategories per cate-

gory can be visualized by a box plot. Outliers can also

be detected and further investigated. Categories with

a high number of subcategories and products must be

further subdivided for data proﬁling.

5.4 Frequencies Product Items

The number of different product types and product

categories as well as the frequencies of the product

items per type and category are of importance. They

allow us to derive the number and size of the parti-

tions, which are needed to plan and execute the sub-

sequent proﬁling steps.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

322

5.5 Dimensions, Volumes and Weights

The dimensions, volume and weight of products are

important for space planning in the shop and for the

transportation planning of the products. These data

are analysed taking into account the different mer-

chandise categories and product types. Outliers are

identiﬁed. By comparing the different merchandise

categories, further anomalies are found. It should be

noted that several packing units are often assigned to

one product as shown in Figure 5. Product items hav-

ing an infrequent combination of units are suspicious

and should be analysed in detail.

Figure 5: Frequencies of items by combinations of pack-

ing units. Red ﬁelds indicate that a packaging unit is used,

while blue ﬁelds stand for unused packaging units. The

right-hand bars depict the total amount of a combination

for all items. The bars above show the total amount of each

packaging unit for all items. The combinations in the upper

rows with very small frequencies are possibly errors.

In addition, we have implemented various data

rules (e.g. product dimensions are very small, very

large, etc.) to identify erroneous product data records.

Figure 6: Heatmap of the percentage distribution of items

by defect classes for each category. For category 12, about

20% of all items are correct, 60% display defect D09, while

the remaining 20% belong to the defect classes D03, D04,

D05 and D10, for example.

Figures 6 and 7 show examples of the results of

such analyses. Figure 7 is the result of a drill-down

by disaggregating the merchandise category.

Figure 7: Heatmap of the percentage distribution of items

by defect classes for each subcategory. The items of cate-

gory 12 as seen in the previous ﬁgure are split into subcate-

gories and analysed regarding the same defect classes. The

detailed analysis shows that the subcategories 1210, 1212,

1222 and 1223 contain the majority of all errors in category

12. All other subcategories are almost error-free. This illus-

trates the validity of our top down approach from suspect

category to ﬁner scaled subcategories.

5.6 Missing Data

Another objective of data proﬁling is to gain infor-

mation about missing attribute values. An univari-

ate analysis ﬁrst determines the suspicious attributes.

Multivariate analyses are then carried out in different

levels of detail for each product type, product cate-

gory or creation period.

Figure 8: Visualization of attributes with missing values by

categories. The colour determines how often values of an

attribute for a certain category are missing. Red indicates

always missing values, blue always ﬁlled values and violet a

combination of both missing and ﬁlled values. For category

13 all values of attribute NETGW (net weight) are missing

for example.

Figure 8, for example, shows missing values in a

three-level gradation for each merchandise category.

Suspicious categories are then analysed in more de-

tail.

5.7 Reference Products and Locations

In many companies, reference product and location

data records are used to simplify the creation of prod-

uct data. In SAP Retail, for example, a reference

A Reference Model for Product Data Proﬁling in Retail ERP Systems

323

product item can be deﬁned for each merchandise cat-

egory. The reference data contain default values that

can be adopted by the user. Using various multivariate

analyses, we propose to determine the frequency of

deviations from the default values. The reference lo-

cations must be taken into account for the logistic data

segment. Frequent deviations are suspicious, they can

indicate incorrect default values for example.

5.8 Change Documents

In some cases certain attribute values of product items

cannot be explained by a multivariate analysis of the

current product data record. To support a root cause

analysis, we have integrated some functions into our

tool that enable an analysis of change documents.

A change document describes when, how and with

which transaction of an ERP system an attribute value

was changed. This should enable a complete analysis

of the product data over their life cycle. Currently,

among other key ﬁgures, the frequency of changes by

attribute and product category and the frequency of

value changes with respect to a speciﬁc attribute are

determined. Product categories with many changes

must then be subjected to further proﬁling steps.

6 CONCLUSION

In the digitalized environment of retail data quality

is absolutely essential. To ensure a high data quality

data proﬁling has to be performed. In this paper we

have presented a reference model and best practices to

proﬁle product data in a retail ERP context efﬁciently

and comprehensively. The objective of our approach

is to reduce the effort needed to obtain information

about the content, structure and quality of product

data by taking into account their speciﬁc properties

and domain knowledge. A ﬁrst evaluation of the ap-

proach was promising, but additional research is nec-

essary.

Moreover, in our opinion the importance of data

proﬁling and exploration is increasing. For instance,

many companies are analysing the potential of ma-

chine learning to improve their processes. As is well

known, machine learning requires training and test

data of high quality. If product data are used, our

approach can deliver useful information to support a

subsequent clean up and labelling of the data for ma-

chine learning purposes. Due to the enormous amount

of data the cleaning of the product data cannot be

done manually, though. Semi-automated cleaning us-

ing simple static data rules can only address a part

of the problem. Therefore our future research also fo-

cuses on the development of intelligent algorithms for

product data cleaning itself.

ACKNOWLEDGEMENTS

The authors would like to thank the German Federal

Ministry of Education and Research for supporting

their work through the KMU-innovativ programme

under grant number 01—S18018.

REFERENCES

Abedjan, Z., Golab, L., and Naumann, F. (2015). Pro-

ﬁling relational data: a survey. The VLDB Journal,

24(4):557–581.

Di Blas, N., Mazuran, M., Paolini, P., Quintarelli, E., and

Tanca, L. (2014). Exploratory computing: a draft

manifesto. Proceedings of the 14th International

Conference on Data Science and Advanced Analytics

(DSAA), pages 577–580.

Domingos, P. (2012). A few useful things to know about

machine learning. Communications of the ACM,

55(10):78.

Guido, A. L., Paiano, R., Pandurino, A., and Pasanisi, S.

(2015). Searching issues: a survey on data exploration

techniques. IJETTCS, 4(6):183–188.

Kandel, S., Parikh, R., Paepcke, A., Hellerstein, J. M., and

Heer, J. (2012). Proﬁler: integrated statistical analy-

sis and visualization for data quality assessment. Pro-

ceedings of the International Working Conference on

Advanced Visual Interfaces.

Olson, J. E. (2008). Data quality: The accuracy dimension.

Morgan Kaufmann.

Papenbrock, T., Bergmann, T., Finke, M., Zwiener, J., and

Naumann, F. (2015). Data proﬁling with metanome.

Proceedings of the VLDB Endowment, 8(12):1860–

1863.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

324