A Survey of Visualizing Business Data
By classifying business intelligence appropriately, we allow ourselves to spot opportunities for investment and exploitation, increasing our ability to turn the data and insight we collect into profit. Business intelligence and its research can be
divided into a taxonomy. This paper breaks that down. Even without data, are there areas that may contain similar opportunities?
3. Business Visualisation Articles
3.1. Business Intelligence (BI)
The category of business intelligence (BI), which consists of two sub-classifications: Internal intelligence and external intelligence. Business intelligence is seen as the generation of meaningful insight into business data.
3.1.1. Internal Intelligence (II)
This subsection contains all inwardly focused BI visualisation research. If the emphasis is placed on internal structures or processes within the business, then we classify the paper to be internal.
Primary Data as Intentional, Active Digital Collection (II): This section contains research that use digital hardware to collect data for the purpose of internal intelligence visualisation research. Only a small amount of research is found in this category, presumably due to the availability of existing data and the high cost of collecting new data. Otsuka et al. presented methods of visualising internal staffing relationships within a company. Each staff member wears a digital name tag that records interactions with other staff members. All internal interactions are recorded and then visually mapped. A topographic map is used to display this interaction. To visualise the interaction, a topological map called a dendrogram is used. See Figure 3.
Figure 3. The dendrographic representation of the hierarchical relationship. The higher up the member is, the closer to the centre their node is drawn. Clusters of users are highlighted in green. The dots represent individuals and are also given unique identification numbers. The dots are coloured by the role classification of the employee and the definition of the contours around the dots within each group highlight how well the group integrates. Image courtesy of Otsuka et al.
The employees are clustered into groups showing the social and working dynamic of the organisation. Each employee can fall into multiple groups and therefore can be represented by multiple nodes. The process of adding new members automatically generates
the dendrograms. This is done by either creating a new group, adding the member to an existing group, or merging two groups to make the member fit. The novel algorithm classifies the interactions into a hierarchy in such a way that members can belong
to multiple groupings, and the staff hierarchy can also be seen. The research informs managers on the employee networks in their own company.
The strongest influences in the field of visualising hierarchical structures have been MoireGraphs, Multi-Tree Hierarchies, and ConeTrees. This research data have strong links with social networking data. Visualisation projects of this nature are becoming
increasingly more popular as social networks grow.
Primary Data as Intentional, Active Research Study (II): In the following, internal intelligence (II) is derived from data sourced using direct research studies. Sedlmair et al. examined the analytical role of visualisation tools used in large
automotive companies. Nine different challenges are discussed along with a set of recommendations for planning and evaluating large company visualisation software. The paper also explores two case studies within the automotive industry that highlight
the challenges of information visualisation in the environment of large businesses.
The article outlines many field characteristics that present challenges when attempting to perform analytical data visualisation in a large business environment. These challenges range from the aligning the software capabilities with the corporate aims,
through data acquisition processes, to the end result of either publication, research, or appealing to the stakeholders. The recommendations for the field are established as a counter to the challenges, e.g., the challenge of publication and stakeholder
appeasement can be overcome by making all publication conditions clear at the start of the project, agreeing what components of the research can be published through a documented communication.
- Case Study 1: AutobahnVis. The AutobahnVis software provides an overview and navigation of error detection in network communication logs. The challenges that arose while developing the software were largely the complexity of the data and the specialised skill required to interpret it. It had to be acquired from busy staff members within the company, resulting in a large time cost and expense to the project. The complexity of the project is reflected in the design, and therefore presented several challenges along the way.
- Case Study 2: MostVis. The MostVis software is designed as an alternative visual access to auxiliary information. It presents large hierarchical data related to the bus systems of car models. The visual hierarchy tree runs from left to right and shows complex information about a car's auxiliary data. Company stakeholders accepted the resulting software research and provided funding to expand further, highlighting the importance of stakeholder support in visualisation research.
The next paper uses primary interview data to increase understanding of a company's employees. Kandel et al. conducted an interview study with 35 enterprise analysts with the aim to better understand their day-to-day operations and how visualisation tools
are used from the analysts perspective. The study is conducted with 35 participants from 25 organisations within 15 different industries.
It was found that analysts generally fall into three main categories; hacker, scripter, and application user. These three groups of people have very different tool requirements. Six of the "Hackers" claimed that visualisation tools such as Tableau or
D3 are useful only as reporting tools as they did not offer any data flexibility and can only be used to present information. They already know what information they would like to portray. Scripters did use statistical visualisation packages to produce
visualisations for exploration purposes, and they found that using the same package for visualisation and analysis helped them transition smoothly between visualisation and analysis. The extent to which application users created visual designs was
through simple packages such as Excel or used standard reporting tools such as Crystal Reports.
Analysts report that a primary benefit of visualisation is error detection. When working with large amounts of data, errors in the collection often go unnoticed, and visualisation highlights these errors. In general, it is reported that visualisations
are best used alongside statistical analysis. The findings indicate that imagery of large amounts of high dimensional data is too complex, and simple visualisations do not scale to this level.
Prior to Kandel et al., Sedlmair et al. discussed the difficulties of evaluating visualisation tools in the corporate environments. Kwon and Fisher discussed the difficulties of using visualisation tools from the perspective of someone who is not trained
in the field. Other research has looked at the processes of analysts but often do not focus specifically on business or on visualisation.
Burkhard presented a framework for the creation of business strategy visual designs. Building on the knowledge visualisation framework, this research identifies the aspects of strategic data suited for visualisation by isolating the different perspectives
of the visual design. The resulting guidelines produce valuable strategy based imagery suitable in a business context – focussing on internal operations of a business.
- The Function perspective distinguishes functions of visualisations based on the desired outcome; i.e., if the goal is to create new insight, recall the data, produce motivation, elaboration etc.
- The Knowledge perspective identifies the type of knowledge that is required to be transferred, i.e., what, who, where, why, and how?
- The Recipient perspective highlights the target group recipient, i.e., individual worker, team leader, senior management, workgroup etc.
- The Visualisation type perspective examines the type of visual design suitable for the above context. i.e., sketches, diagrams, maps, images, interactive visualisations, stories.
Secondary Data as A Priori Database (II): The literature in this section studies pre-existing data for internal intelligence visualisation. The first paper describes some visual designs to inform a company's internal decision making after
the original data is collected.
Wright presented six case study examples of info-vis software used in a business context. These real-world examples are among the first recorded utilisations of information visualisation software used in the day-to-day management of the business. In this
article, case studies focus on 3D visualisations of pre-existing business data ranging from financial information to management support.
- Fixed Income Management: In this case study, a dataset of financial portfolios is depicted using 3D line graphs. Emphasis is placed on the 3D nature of the visual design as it enables thousands of data points to be plotted compared to a smaller number in 2D. The more holistic view enables investors to quickly see the state of their portfolio or compare multiple 2D visual designs.
- Derivatives Risk Management: The software conveys the risk involved in options trading. A virtual environment contains multiple visual representations including a virtual screen showing the yield curve, a surface plot mapping the current profit and loss, and a grid map that shows the relative profit and loss. Users can interact by adjusting the extraneous variables such as interest rates to change the forecast visual designs.
- Management Decision Support: A geospatial map is used to display the locations of a chain of businesses and then 3D bar charts are overlaid on top to show the metric values used to analyse the businesses. This enables managers to evaluate and balance multiple business locations.
- Credit Scoring: This design uses a geospatial map to display credit scores in the United States. The software enables the market risk of permitting loans to be analysed.
- Retail Sales Analysis: Again using geospatial maps, this enables the user to compare the retail value of stores across the U.S. both individually or aggregately in each state. Three-dimensional bar charts or raised map tiles are used to show the sales from each sector or store.
- Management Reporting: This managerial software uses a virtual environment and 3D bar charts to show the portfolios of a business. The portfolios are grouped into asset classes and represent the main axis of data. A virtual screen shows 60 scenarios that would affect the portfolios and users can select each to see the effect. Another virtual screen shows the currency conversion rates which change with the scenario.
Figure 4. The transformation of the treemaps into other shapes: (a) the original treemap; (b) the original transformed into a pyramid but without density uniformity; (c) the modified version of (b) now adjusted to show uniform density; and (d) the transformation of the original into a pie chart. Image courtesy of Vliegen et al.
The first proposed mixed treemap takes the structured layout of the slice and dice algorithm and combines it with the square readability of the squarified treemaps algorithm. This way, the higher level nodes remain ordered, but the lower level nodes become
squarified. The matrix modification enables data comparison of different sizes by subdividing the rectangle into a grid. When the number of cells in the grid exceeds the number of items, then dummy nodes are added and blank cells are used. Another
modification involves the transformation of the treemap visualisations in pixel space to reflect the aesthetics of other popular visual designs, such as pie charts.
Prior to Vliegen et al., Harris provided an extensive overview of traditional depictions of business data. The treemap was first introduced by Johnson and Shneiderman in the early 1990s. Further work enhancing the original design is also presented.
Continuing on this theme, we look at research utilising pre-existing databases for the purposes of exploring internal business operations and functions. Nicholas et al. presented a novel way of conveying the failure rates of automotive components and
the effect that this has on customer satisfaction. The base design of the visual layout extends the standard chord diagram and depicts three-way relationships using both curved angles and glyphs. Previous attempts at chord diagrams focus on two-way
relationships. This method enables a third relationship to be added and compared. The dataset is collected from by automotive company recording the failure rates of automotive components. The automotive data are divided into 11 autonomous fault categories
(engine, transmission, etc.). The aim of the visual design is to ascertain which combination of component failures yield the most dissatisfied customer. See Figure 5 on page 16.
Figure 5. The extended chord diagram where glyphs are added to show the relationship between automotive component failure rates and customer satisfaction. Image courtesy of Nicholas et al.
The multi-chord diagram represents the frequency of failures between components with the curve thickness and then may use colour to represent how dissatisfied the customer is. See Figure 5. In the glyph extension, the lines of the chord diagram are
shortened and only display the intersection of three component fails. This reduces overlap and improves visibility by reducing clutter while retaining the vital information. The benefit of this design is that the company can now clearly see what products
typically fail together as well as how badly the hardware failure affects customer satisfaction. Recommendations can be made to improve the worst offending components that have the largest impact on customer satisfaction. Less focus is placed on the
failing components that do not have a negative influence on customer satisfaction.
The inspiration for the chord diagram extension came from Bostock, Ogievetsy and Heer in their "Data-Driven Documents" paper. A radial technique by Kerren and Jusufi enables the visualisation of an undirected hyper-graph. Nicholas et al. used elements
of this visual design.
This next example of derived internal intelligence using a pre-existing industry acquired data involves the analysis of call centre data. Roberts et al. presented an analytics system that visualises call centre data provided by their industry partner.
They modified the traditional treemap to accommodate time series, event-based data such that 24 h of call centre activity can be presented in one view. Novel interactions and filtering methods are used to modify the view from a full day of data to
individual call records.
The top hierarchical level of the treemap shows leaf nodes representing time frames (24 h > 1 h > 10 min > 1 min). When an hour node is selected, the graphic is zoomed in smoothly to reveal the individual call records as the new leaf nodes of
the treemap. A selection of sliders enables the user to filter the call records by a range of metrics, narrowing the scope of the calls in a focus + context environment. Roberts et al. noted that queue times sharply increase around 13:00 each day.
This is attributed to shifting staff levels within the call centre. See Figure 6 on page 17.
Figure 6. A focus + context treemap visualisation of call centre data: (a) all callers who have waited longer than 15 min in the queue but spoke to an agent for less than that period of time; (b) a combination of temporal and event based filters; (c) the increased queue time at the start of 13:00; and (d) all abandoned calls during the hour. Image courtesy of Roberts et al.
A common area of research in the call centre is the customer service quality being provided by the staff. This research extends Blanch and Lecolinet's work on navigating treemaps using a zoom interaction
Roberts et al. continued their work exploring call centre event data, using parallel coordinates plots (PCP). See Figure 7. The focus of this research is to develop new brushing techniques to overcome the challenges associated with overplotting.
The main contribution lies in the "sketch-based" brushing that can be easily applied, modified, and moved around the plot to enhance the analytics of the software..
Figure 7. The sketch brush placement with the dynamic brush interval range glyphs which were automatically placed at 1 standard deviation from the mean. Image courtesy of Roberts et al.
Additionally, a range of glyph-based user options guide the user in their brush placement and provide additional information about the surrounding metadata. A priority rendering feature lets the user select a point on the n-dimensional plot that they
want to focus on, and the draw order of the polylines change to show that data drawn on top, helping with overplotted graphs. The software also includes an automatic brushing feature that can be used to either scale the parallel coordinates axis or
apply a brush sketch according to the distribution of the data. This feature aims to make complex datasets accessible to new users.
Secondary Data as a Business Process (II): This sub-section focuses on internal intelligence generated from existing business processes. The derived internal intelligence of this literature aims to inform decision making.
Broeksema et al. presented a visual analytics system for operational decision making in a business management environment. The system displays the decision-making process in the software. A car insurance case study is used. The VA toolset is used
to analyse a dataset from the car insurance industry. The data used is from an auto-quote request that has been stripped of confidential information.
A set of 144 rules process the input information and generate a quote. Discounts are applied based on variables such as car safety features, no claims bonuses, etc. These data are then fed into the decision map visualisation. The decision map is based
on work by Zizi et al. but instead maps concepts as opposed to instances. Two diverging factors in the data are plotted against each other in a scatterplot and the space is segmented according to the most prolific value in that data space. See Figure
8.
Figure 8. The decision map as an interactive analysis of variables of interest: (a) the analysts has selected the complete list of available attributes associated with drivers and then selected the age group bar chart to show where the clusters of age groups sit in the plot; (b) the simplified version where all non-contributing variables are removed; and (c) a breakdown of the decision process and classification of the "students" demographic. Image courtesy of Broeksema et al.
The decision map is primarily based on Zizi et al.'s dynamic map. Decision support systems often are used with financial data.
Secondary Data as a Business By-Product (II): This segment describes visualisation research related to internal intelligence using business by-product data. In this example, the by-product is data recording delivery times and product sales. Gresh and Kelton present a visualisation application focused on the presentation of business intelligence data using 2D and 3D visual designs in the delivery industry.
Customer delivery times are the focus of this research. The business has delivery targets to achieve, however, due to the vast stock range of hardware and the small delivery targets, the business needs to optimise their warehouse locations and stocking. This optimisation analysis is done through the visualisation of the delivery times. When using the software, the user is presented with a control window that enables them to choose a subset of the data. The user is then shown a combination of two- and three-dimensional visual designs. These images show the target service level in comparison to the actual service level.
Comparable software has been created, but does not exploit the full potential of the data. A more customised approach is required for this subset.
Keim et al. proposed a new type of bar chart that can be used in the visual analysis of large transaction datasets. The design enables the user to see transaction value correlations and outliers. The bar chart is created from sales data by separating the range of transaction values into tiers and then assigning each transaction a tier. The bar is drawn as a sorted accumulation of all transactions within that period. The tiers are coloured such that the bar appears to be subject to a continuous colour gradation while still visualising each transaction as discrete. See Figure 9.
Figure 9. The Value-Cell bar chart. Each bar represents a month worth of sales. Within the bar, tiered sales values for each month are shown. The red tips of some bars represent outlying large sales during that time period. The colour legend shows the value of the transactions. The tiers blend together to create a gradation effect. Image courtesy of Keim et al.
The design of the value-cell bar chart shares elements with Keim's VisDB system that sorts coloured pixels in relation to a user query. Further related work introduced a pixel-based bar chart for large transaction datasets.
3.1.2. External Intelligence (EI)
This subsection contains outward focused BI visualisation research. If the emphasis is placed on the external environment such as direct competition within the market sector, then we classify the paper as external.
The types of stakeholders examined in this paper range from the marketing team to the end users. Each has a different set of expectations for the visual design. The range of requirements spans across two axes. The first ranges from effective to attractive,
and the second from simple to complex. Although this is a generalisation, often effective visualisations are not considered aesthetically pleasing, whereas attractive graphics are not always the most effective. The y-axis is the scale of visual complexity.
Some stakeholders require a complex visualisation to show off the work while others require a simple graphic to aid in information retention. See Figure 10 on page 20.
Figure 10. The stakeholder visualisation requirements map. The x-axis ranges from "effective" to "attractive" and the y axis ranges from "simple" to "complex". The marketing departments expectations of a visualisation rank highly in both complexity and attractiveness, but indicate less interest about its effectiveness. Image courtesy of Keahey.
Secondary data as A Priori Database (EI): This subsection contains one visualisation research paper that uses a pre-existing database for analysis in the field of external intelligence. Ferreira et al. utilised the many urban taxi behaviour
data that have been collected in New York City. Primarily, geospatial data are used in the visualisations to map out the routes the taxis traditionally make. This research intends to find answers to some previously unanswered questions such "What
is the average trip time?", "How does taxi activity vary throughout the week?", and "What effect do major events have on the taxi behaviour?".
The TaxiVis software enables the user to select the time frame. The geospatial paths appear on a map widget, and the raw data appear on a data summary widget. See Figure 11. A heat map is overlaid onto a map of the city to display the majority of the data. This heat map highlights the trip density within the city as plotting individual points would occlude most of the data due to its volume. Side-by-side comparisons are used to show each day's activity throughout the week. It is easy to observe that Monday is the quietest day with activity progressively getting busier throughout the week. The same side-by-side comparison is used to compare the taxi activity during major city events such as presidential visits, or natural disasters. Using day-to-day comparisons of the same map visualisations shows the progression of such events.
Figure 11. Taxi activity in Manhattan during the week of Hurricane Irene. Image courtesy of Ferreira et al.
Previously, Ge et al. proposed an analytics method for taxi drivers to calculate the most financially efficient way of finding passengers. Peng et al. modelled the day-to-day habits of taxi drivers; however, neither of these research topics is visualisation
focused. There has been research into the visualisation of movement data, but not explicitly related to taxi data.
Secondary Data as a Business Process (EI): Here, we look at two papers by Hao et al. that explore the external environment of a business that impacts the day-to-day operations of the business.
Hao et al. presented BizViz, a visualisation software that interactively visualises business operations processes. The BizViz software analyses the relationships between important external operational parameters. The primary focus is on data distribution
of up to three operational parameters but the user can drill down to a one-on-one comparison of parameters where the transaction sets can be seen.
The design uses a circular chord-like plot to present its data. Separated into three sections, the circle presents one attribute on the left half, one attribute on the centre dividing line, and the final attribute on the right side. See Figure 12.
Edges denote relationships. When drilled down to two attributes, the circle is split into two, and the half circles are plotted with horizontal lines at varying density to represent the two selected operational parameters. The data specifically deal
with fraud prevention and credit card usage in retail.
Figure 12. The BizViz visualisation. The left side of the circle shows each geographical region the data was collected from. Lines are drawn from each region to the centre line which holds a range denoting fraud value. From there, the line crosses over to the right axis where the fraud count is tallied. The user can select a region to drill down into and reveal comparisons of the fraud value vs. fraud count. Image courtesy of Hao et al.
Hao et al. later published further research in this field, continuing the parallel/chord visualisation in the financial security sector. They focused on business process impact by adapting the previous visualisation methods use case to show the new data.
The examples show the source attribute to be the customer type (measured by importance or size). The intermediate attribute is mapped to a time frame for ordering (delays, or order time), and the destination attribute is mapped to the outcome (order
accepted/rejected, penalty costing, etc.).
Secondary Data as a Business By-Product (EI): This subsection contains external intelligence visualisation research using data sourced from the by-product of business operations.
Liu et al. explored the potential of billboard placement through the visualisation of GPS taxi data. Using a combination of geospatial visualisation and glyph designs, Liu et al. explored the optimum placements of billboards across a city. A geospatial
heat map is used to display the flow of taxi traffic, highlighting high volume areas where billboard potential is maximised. The solution view uses layered circular glyphs to evaluate potential points of interest where billboards can be placed. Using
a combination of the location view and solution view, the user can select which location is best suited to their billboard as a function of cost and footfall.
The billboard location selection process is a sub-problem of Multicriteria Decision Making, using a geospatial context. Taxi GPS data have been extensively researched with applications from route optimisation to urban planning. Visualisation is often
used to present data of this nature. Chen et al. provided a full survey of traffic data visualisation.
Data Overview: We provide an overview of the data descriptions in each paper and their availability in Table 4. It shows a description of each data source in the survey categorised to match Table 1. It shows that most data in the
business visualisation literature are still proprietary.
Table 4. This table summarises the data sources in each research paper, identifying the accessibility of the data as well as a brief description.
| Classification | Paper Ref | Access | Description | |
|---|---|---|---|---|
| Business Intelligence | Internal Intelligence | Wright | Proprietary | Case Study from portfolio management, derivatives management, customer credit scores |
| Gresh and Kelton | Proprietary | Private IBM business by-product data | ||
| Eick | Proprietary | Log data from web servers used to analyse the efficiency of their website | ||
| Burkhard | Proprietary | Case study from Swiss Federal Institute of Technology using business strategy data | ||
| Vliegen et al. | Proprietary | Unspecified business data | ||
| Keim et al. | Proprietary | Transaction datasets | ||
| Otsuka et al. | Proprietary | Digital nametags collect employee interaction data | ||
| Sedlmair et al. | Survey | Existing software evaluation | ||
| Kandel et al. | Proprietary | Interview Study with industry experts | ||
| Du et al. | Survey | A survey of business process visualisation literature | ||
| Aigner | Proprietary | Text from interview study | ||
| Broeksema et al. | Proprietary | Decision model data | ||
| Bai et al. | Proprietary | Geospatial data for utility network coverage | ||
| Lafon et al. | Proprietary | User Study of unspecified business data visualisation | ||
| Nicholas et al. | Proprietary | Private customer survey database from automotive company | ||
| Roberts et al. | Proprietary | Private call centre interaction database | ||
| Ghooshchi et al. | Proprietary | Business Processes from undefined source | ||
| Kumar and Belwal | Public | Multiple public data sources looking at different aspects of a business | ||
| Bachhofner et al. | Proprietary | Business processes from industry contacts | ||
| Lea et al. | Proprietary | Business process data was used alongside simulated data to test prototypes | ||
| Roberts et al. | Proprietary | Call centre event data from industry partner | ||
| External Intelligence | Hao et al. | Proprietary | Case study data from financial transactions, service contracts data | |
| Hao et al. | Proprietary | Case Study Data: Financial transactions, service contracts data | ||
| Bresciani and Eppler | Public/Proprietary | Case study from Gartner, Argument Map, Five Forces Process | ||
| Bertschi | N/a | Critical Discussion of knowledge visualisation in business. No data used | ||
| Ferreira et al. | Proprietary | Data provided by Taxi and Limousine Commission of New York City | ||
| Keahey | Proprietary | Expert opinion data | ||
| Liu et al | Proprietary | GPS trajectory data | ||
| Ramesh et al. | Public | Data mined from "various sources". Presented for insight into the external operations of a business | ||
| Business Intelligence | Business Ecosystem | Wattenberg | Public | Public stock market data |
| Merino et al. | Public | Stock market data | ||
| Otjacques et al. | Proprietary | Human resources data | ||
| Wu and Phillips | Public | Public Dow Jones 30 data | ||
| Basole et al. | Proprietary | Business ecosystem data | ||
| Ko et al. | Proprietary | Generic Point of Sale data | ||
| Basole et al. | Commercially and Publicly Available | The Thomson Reuters SDC Platinum database and Capital IQ Compustat database | ||
| Basole | Proprietary | Case study: global supply chain data, competitive dynamics data, venture capital network data | ||
| Deligiannidis and Noyes | Proprietary | Data obtained from US Department of Commerce Census Bureau | ||
| Basole and Bellamy | Proprietary | Supply network Structure data | ||
| Lu et al. | Public | Twitter data + IMDb | ||
| Basole et al. | Proprietary | Three commercial datasets are used that cover finance, relationships, and public opinion | ||
| Iyer and Basole | Proprietary | The visualisations use IoT data to show the "big players" in the technology industry | ||
| Basole et al. | Proprietary | User study generated data looking at the effectiveness of different visual designs for decision support | ||
| Schotter et al. | Proprietary | Investment data is used alongside geospatial data | ||
| Basole et al. | Proprietary | Combination of multiple proprietary datasets including geospatial and commercial data | ||
| Customer Centric | Customer Behaviour | Woo et al. | Proprietary | Audio data from customers in call centre |
| Hanafizadeh and Mirzazadeh | Proprietary | Six-dimensional vector customer dataset | ||
| Shi et al. | Proprietary | Generic search engine data | ||
| Rodden | Proprietary | Private Youtube site navigation data | ||
| Yaeli et al. | Proprietary | Digitally collected customer path tracking data | ||
| Dou et al. | Proprietary | Survey conducted on Reddit.com | ||
| Kameoka et al. | Proprietary | Dataset provided by industry parnter – supermarket PoS data | ||
| Nair et al. | Proprietary | Large customer behaviour dataset – unspecified origin | ||
| Wu et al. | Proprietary | Telco data obtained from China's largest telecommunications operator | ||
| Nagaoka et al. | Proprietary | Customer behaviour collected from digital devices | ||
| Sijtsma et al. | Public | Twitter data mined to collect the customer experience and expectation of various retail stores | ||
| Sathiyanarayanan et al. | Public | Email exchange at company level | ||
| Customer Centric | Customer Feedback | Brodbeck and Girardin | Proprietary | Questionnaires distributed to customer of the public transport network |
| Chen et al. | Public | Amazon.com reviews | ||
| Ziegler et al. | Proprietary | Unspecified textual customer feedback data | ||
| Oelke et al. | Public | Amazon.com reviews | ||
| Wu et al. | Public | TripAdvisor data used | ||
| Hao et al. | Public | Twitter data | ||
| Saitoh | Proprietary | Web scraped customer review data | ||
| Kang et al. | Proprietary | Combination of production and customer service data direct from manufacturer | ||
| Fayoumi et al. | Proprietary | Web scraped social media data from Twitter | ||
| Haleem et al. | Proprietary | Web scraped customer reviews | ||
| Saga and Yagi | Public | Customer feedback collected from web crawler using specified keywords about the examined product | ||