Towards big data as official statistics: Case study of the use of mobile positioning data to delineate metropolitan areas in Indonesia

2.2.1Business process layer

The Business layer is the first thing that needs to be designed because this layer determines the data flow, along with the applications and technologies required to improve business value. Considering the huge amount of potential big data to be explored, some might prefer to collect as much data as possible without carefully considering the actual output they need. However, it would be wise to perform a preliminary literature study and assessment to obtain clear objectives for any data analysis based on user needs, and not merely driven by data availability. According to [5], more data does not necessarily mean more insights. Referring to the United Nations Statistical Quality Assurance Framework (UN-SQAF), big data does not need to be treated differently from other potential data sources in the Specify Needs phase [6]. Big data’s potential as a data source in producing statistics needs to be evaluated just like any other data sources through an assessment of whether the collected data would meet user requirements and, if it does, then how to efficiently obtain and process that data.

The business process layer in this proposed framework offers a solution to overcome the different treatments that are needed to handle big data. In the Design phase, it is more recommended to study the structure of the data source and select the required variables compared to the needs of the design frame and sample. In the Build phase, the focus is on preparing the system and infrastructure along with implementing the designed rules and algorithms for data collection in the system. In the Collect phase, data collection activity in principle can start right away without the need of either sample selection or field-force management, although sampling issues and field-force type interactions may need to be addressed later to deal with emerging quality issues. There are also some differences in the Process and Analyze phase, mostly in the statistical method. In addition to traditional statistical methods, the processing and analyzing of big data often uses machine learning and information extraction. It is also important to compare big data results to relevant official statistics and other data during the Analysis phase for benchmarking or calibrating. The result of this benchmarking and calibrating can then be used to determine the appropriate type of publication during the Dissemination phase.

All of the existing processes are open to iteration, meaning it is possible to roll back into previous processes when adjustment is deemed necessary in big data use. It is also important to consider the benchmarking result before deciding whether the results would be published as official statistics, supporting information, experimental statistics, or even a mere research paper or report. In order to publish official statistics from big data sources, a lot more things need to be properly validated and assured. Therefore, it is important to define quality assurance indicators to help monitoring the success in each phase, not least in monitoring whether user needs can be met under the required dimensions of data quality. For example, in UNECE Big Data Quality Framework, there is the selectivity dimension. This dimension explains how it would be helpful to have a reference data set to measure the coverage of the data source obtained from big data. The result of this measurement can then be used to address the issue of data representativeness before publishing the result from big data as official statistics.

Figure 3.

Design of a big data platform.

2.2.2Data layer

The Data layer represents the data components that need to be considered in each process. Figure 2 shows that the data layer encompasses the main output in each process, which if detailed will result in many entities and attributes. For example, the Process phase requires both data structure modeling for the expected results from data cleansing and information extraction, as an input to produce relevant clean data and tabulations as an output. Data needs to be linked to its metadata to help users understand the data and further address data quality issues such as timeliness, completeness, and consistency with other relevant data. Each process should produce metadata that later will be used for the next process. With well-documented metadata, its reusability would be possible. This will furthermore help other divisions in BPS-Statistics Indonesia to learn from and re-create/adapt previous big data projects, be it in the analysis of the purpose, rules and algorithms, collected variables, or the analysis and evaluation of the results.

2.2.3Application layer

The Application layer describes the supporting systems that enable each process to run well. The way each system contributes to the business process is shown in Fig. 2. The previously mentioned data and metadata will serve as inputs and outputs that will flow from/to the data and metadata management systems (DMS and MMS respectively). Among all those systems, DMS and MMS are enablers that provide over-arching support throughout all business processes. Systems like an Application Programming Interface (API) Management Tool also can be used every now and then as a main means of communication for all systems regardless of the business process.

Figure 3 shows the process flow in a DMS as a single source of truth in a big data platform. Data sources can be obtained through administrative products, census, surveys, external websites, and even internal systems. Related to the resources that are needed for big data storing and processing, there is an alternative option to access data from big data providers. Data providers and BPS-Statistics Indonesia can collaborate in sharing resources such as raw data. The data could remain in a data provider’s repository and a data engineer from a BPS-Statistics Indonesia can process the data using the data provider’s resources. The processing results in the form of aggregate data could then be accessed and stored in BPS-Statistics Indonesia repository. All data will be stored in a working database and later than be loaded into an enterprise data lake (EDL). There are several zones in an EDL: landing (for newly loaded data with incomplete metadata), working (for data processing in EDL), and a ‘gold’ zone (for processing results and storing data with complete metadata). Data in the gold zone will then be loaded into a data warehouse for further access by internal systems and the dissemination system via an API. A Business Intelligence (BI) Tool can then access data from the EDL gold zone and the data warehouse.

2.2.4Technology layer

The Technology layer describes the infrastructure used in the big data ecosystem. The biggest difference in the technology layer for big data is the existence of an EDL, which contains several nodes that are used to store and process large data in a distributed approach. Besides the EDL, there are database clusters, data warehouses, and other kind of storages each for certain purposes. There is also a dedicated server for an API Management Tool to support data interoperability and a BI tool to support data visualization. The use of all these technologies are optional and can be adjusted accordingly.

2.2.5Roles

An appropriate division of roles is vital to ensure that business processes run smoothly. In using big data, BPS-Statistics Indonesia takes the role of data consumer and if needed could coordinate with data providers and other NSOs who have conducted big data projects before [7]. Besides that, each role must communicate with all other so that the result would be in concordance with the predefined output. There are several proposed main roles: Data Partnership Officers, Subject Matter Experts, Data Engineers, Database Administrators, Data Scientists, and BI Analysts.

2.2.6Quality management, standard and method management, metadata management, security, and governance

2.2.6.1. Quality management

Quality management is a mechanism to ensure the quality of each performed business process, including inputs and outputs, in order to meet the needs of users. Even if the big data project is a mere exploration or experimental, quality aspects should always be taken into account for optimal results. Quality has an important role in building confidence in the products produced by NSOs. Each NSO can develop their own Statistical Quality Assurance Framework (SQAF) as a guideline for producing better quality statistics. For example, a quality assurance framework (QAF) addressing the assurance of quality at each stage of processing: 1) quality of input data, 2) quality of statistical processes (throughput), and 3) the quality of the product (output). With these three stages of quality assurance, the quality of statistical data will be well maintained because it is always checked at every stage of the process of statistical production. However, it is important to note that assessing the quality of big data is quite different from those of traditional datasets. Referring to Big Data Quality Framework (QAF) designed by UNECE Big Data Quality Task Team, three general principles are proposed when evaluating Big Data quality [8]:

2.2.6.2. Standards and methods management

In conducting statistical activity, BPS-Statistics Indonesia implements statistical standards to ensure user needs are met, not least in the comparability of the produced output, be it nationally or internationally. Beside standards, there would be a statistical method used in the Design phase and its implementation in the Collection, Processing, and Analysis phases. For big data, there are several underlying issues, such as how to create a general and replicable model. With the uncertainty factor in big data, repetitive simulations are required to evaluate the model [9].

2.2.6.3. Metadata management

Similar to the GSBPM, the SBFA uses metadata as a key element that can be obtained and recorded through each phases in statistical business process. Metadata can be collected manually or automatically in using system services and it also can be reused. In statistical activity the functions of metadata could be described as follows [3].

2.2.6.4. Security

As an additional layer, security is related to mechanisms that ensure all components involved in the statistical business process run in a secure manner. For a properly secure architecture, support from governance, including policy and procedure, in business layer is important to ensure compliance during its implementation in data, application, and technology layer.

2.2.6.5. Governance

Governance relates to guidelines consisting of policies and procedures that will be implemented throughout running the processes. These guidelines could be detailed into their technical aspects and later be applied in the application and technology that is being used. For example, a policy is established to ensure that collected data must have and be linked to its metadata.

3.4.1Commuter identification

Based on the definition used by BPS-Statistics Indonesia, the term ‘commuter’ refers to someone who routinely does a round-trip within 24 hours to work/school/course outside his administrative area of residence. There are two aspects to consider: a spatial aspect and a routine aspect. When using MPD, the location of someone’s residence and the location of his/her main activity (work/school/course), or their round-trip routine cannot be determined directly by interviewing respondents as in conventional survey. Therefore, it is necessary to adjust those two aspects to determine commuters using MPD.

The location of residence and the location of main activity in MPD can be described using the Usual Environment scheme [14]. The residence location (home) is the subscribers location in the nighttime, while the main activity location (work) is the subscribers location in the daytime. The highest aggregate time duration of home and work in subdistrict level will be chosen. If the chosen home and work of a subscriber are in a different sub-district, the subscriber fulfills the spatial aspect of a commuter. Subsequently, the subscriber fulfills the routine aspect of commuter when he/she made a round-trip from home to work at least once a week in two different weeks on a one-month observation period. If the subscriber fulfills spatial and routine aspects, he/she is classified as a commuter. The commuting flows were then identified for every subscriber from origin/home to destination/work. BPS-Statistics Indonesia developed home, work, and commuter algorithms using this concept and adapted to conditions in Indonesia.

3.4.2Metropolitan area delineation

According to the Ministry of Public Works of Indonesia, the establishment of a metropolitan area delineation considers various aspects; one of them is the high rate of commuter mobility between the urban core and surrounding hinterlands [15]. The metropolitan area delineation based on commuter patterns is carried out in various countries, including European countries, the United States, and Japan. Many European countries determine surrounding hinterlands that are included in a metropolitan area by using the Functional Urban Area/FUA [16]. FUA consists of the city and its commuting zone. In this approach, based on an example from Eurostat in Fig. 6, a commuting zone can be identified based on commuter patterns using the following steps:

Different country has different thresholds in determining the surrounding hinterlands that fall within the commuting zone. Many European countries use at least 15 percent [16], while the United States uses at least 25 percent [11]. Meanwhile, Japan determines at least 5 percent of worker and student residents [17]. The percentage was obtained by dividing the number of commuting workers with the total workers in each hinterland. However, in MPD, the subscribers can not be classified as workers or students. According to the commuter survey result in Cekungan Bandung, 69.19 percent of commuters in surrounding hinterlands were commuting for work purposes. Hence, in this study, most subscribers were assumed to be workers so that the percentage of commuters used was the number of commuting subscribers divided by total subscribers. In addition, based on 2018 Indonesian Labor Force Survey (Sakernas), 8.91 percent of surrounding hinterlands workers in Cekungan Bandung commuted to its urban core. Therefore, this study used 5 percent threshold.

3.4.3Quality assurance of MPD

To obtain trusted statistics, quality checking is conducted before data processing, the quality of raw data is examined in terms of diurnal distribution records, missing data, removing non-human (machine) data etc. Then, to validate the result of MPD algorithms, this study employed MEILI, an open-source mobile application that records the location of multiple users over a period-of-time and can be annotated to verify each trip. This validation method was used only for identifying the commuting pattern differences between MPD and MEILI results. So that the more sample used, the more commuting patterns can be analyzed. Due to budget constraints, this project used 907 volunteers who were subset from Telkomsel subscribers.

Basically, MEILI is a travel diary survey that uses mobile phone as a data collection mode. While MPD uses passive geo-positioning to determine someone’s location (aggregate BTS location), MEILI uses GPS, an active geo-positioning to determine the location of the mobile phone. Thus, location information resulting from MEILI is more accurate compared to location information from MPD. However, there are some disadvantages to MEILI. This application needs to be installed on a mobile phone with certain specifications, consumes a high amount of battery, and requires a user’s permission to activate the GPS feature.

Through the MEILI application installed on a mobile phone, volunteers annotated their locations, trips, and the purpose of each trip. Thus, the location of volunteers’ home/work can be known by looking at the trip purpose. The location was identified as home when the trip purpose annotation is “my house” or “my family/friend house”, the stay location in the nighttime, or the stay location in days off. On the other hand, the location was identified as work when the trip purpose annotation is “my office” or “my school/course place”, or the stay location on working hours. The identification of home and work locations result from MEILI were then compared to the MPD result can be seen in Table 4. Therefore, the higher the similarity level of home and work identification between MPD and MEILI, indicate the better commuting flows produced from MPD.