Addressing the issue of undeclared work – Part I: Applying associative classification per the CRISP-DM methodology

3.1 The organization

The analysis of the present study is based on real business needs of a labor law inspection and enforcement authority, the Hellenic Labor Inspectorate (HLI), and was carried through close cooperation with the domain experts at its different phases. The HLI lies under the Ministry of Labor and Social Affairs in Greece, aiming to guarantee a decent working environment by ensuring the application of the labor law provisions in both fields of labor relations and occupational safety and health. This is accomplished by providing advice to employers and employees regarding the most effective means of complying with labor legislation and regulations, but also by performing inspections in workplaces to examine the lawful insurance coverage and employment of workers, targeting to protect their rights. Thus, one of the main objectives of the Authority is to efficiently identify and deal with undeclared work and other labor law violations and, at the same time, use its resources most adequately.

The HLI consists of the labor relations inspection services and the occupational safety and health inspection services, having central offices in Athens and 125 local departments countrywide, with around 650 labor inspectors in total offering their services. At present, the labor market consists of around 300K active registered employers and around 2M active registered employees, with both numbers having a generally increasing trend each year. This fact, along with the decreased human force of the Authority per around 30% the last decade and the increasingly complex labor regulatory setting, has unveiled the HLI difficulties in carrying out its role and also the critical need for new and advanced tools to be embodied into its activities to achieve higher inspection yields.

For the last three years, the HLI owns a risk analysis tool able to rank employers based on a calculated risk, using predefined user-specified rules composed of various criteria. This tool can be employed in the planning and selection of onsite inspections, yet it has some major disadvantages resulting in a low level of acceptance by the labor inspectors. To begin with, the risk analysis rules need to be regularly manually updated, following the new scheduled inspections results, since the system does not support this process automatically. Moreover, the creation and update of the rules rely on the experience and perceptions of labor inspection professionals, who can have biased ideas regarding patterns of labor law non-compliance, thus rendering the tool misleading, as it was proved in [8]. Another issue refers to the configuration of the tool, which is done by experts at the central offices of the Authority since it requires specific competency. Yet, this results in the local inspectors not participating in this process and thus getting unwilling to correspond to the tool suggestions and inspect companies scored as ‘risky’, without actually knowing the factors affecting this score. The above issues brought about the gradual abandonment of this risk analysis tool and, as a consequence, the selection of the companies for inspection is still often done arbitrarily, or it is based on reported complaints and whistleblower information ( ∼ 20%), or the inspectors intuition, usually resulting in recurring and superfluous inspections with low efficacy. This situation induced the recent quest for new technology solutions to attain increased benefits and improved allocation of the HLI resources.

3.2 Business objectives and goals

Yearly, the HLI inspectors perform around 35K on-site visits to inspect labor relations and around 30K visits to check the conditions of occupational safety and health. Focusing on the results of the labor relations inspections, only half of them is found with violations, out of which around 4% corresponds to fully undeclared work and nearly 35% to underdeclared work. However, the last survey regarding undeclared work conducted by the EU in September 2019 – Special Eurobarometer 498 [4] – published that in Greece, in the demand side, 27% of the respondents admitted having paid for goods or services believing to have undeclared work involved, and, in the supply side, 59% of the respondents said that they personally know someone not declaring all or part of their income to tax or social security authorities, with the corresponding rates lying at 10% and 33% at EU level. Apparently, the figures between the EU Member States exhibit significant differentiation, with Greece displaying one of the highest rates of fully or partially undeclared work prevalence in the EU.

Thus, considering these statistics and the said operational difficulties the HLI encounters, the primary business objective of the Authority is to detect and assimilate into its business processes an advanced method that will assist it to, first, overcome its functioning constraints and, secondly, use its resources more properly, aiming to reach an increased proficiency in dealing with the problem of undeclared work.

Consequently, this study focuses on the analysis of a progressive tool to accomplish the above two goals of the Authority, which can be converted into the following two success criteria, equally important. First, the tool must offer significant help to the labor inspectors in the selection process of the companies to be inspected, attaining higher effectiveness in detecting labor law violations than the present one, which lies around 50%. In addition, the tool must not need manual update and configuration, but it should work fully automatically. Second, the tool should provide the labor inspectors with understandable knowledge over the factors that affect each of the tool suggestions regarding the selected companies for inspection visits. This knowledge provision will highly enhance the labor inspectors ability to identify by self those factors indicating labor law violations, and, moreover, it will actively keep them involved in the inspections selection process, thus aiming to bend their objections for a wide use of the tool. In addition, at the administrative level, this knowledge will strongly support the decision-makers to design targeted campaigns for specific groups of companies exhibiting increased levels of non-compliance, to advance the transition of undeclared work into the formal economy.

3.3 Situation assessment

3.4 Data mining goals

Following the above thorough analysis of the Organization at different levels, the last task of this phase is the transformation of the business objectives into a clearly specified data mining problem, with the respective success criteria defined in technical terms, to move forward with the analysis.

Hence, this study focuses on the analysis of an innovative classification – prediction method that can use past experiences, i.e., past inspections data, and produce patterns of employers practices indicating compliance or non-compliance to labor law. A tool embodying this machine learning method should be able to predict, with increased accuracy, the employers likely involved in undeclared work and other labor law violations and thus assist the Authority in planning inspection visits. This tool must not need manual configuration or update, but it ought to be able to update itself automatically, learning from the new inspections data inserted into the HLI-IIS.

In addition, the patterns produced relating to the employers practices must be easily understandable and provide the labor inspectors with valuable insights into those attributes that are highly correlated with these practices, thus improving their expertise in detecting violations and their ability to plan effective inspection visits.

Concluding, the success criteria of this data mining problem include the provision of a set of comprehensible rules classifying employers into ‘compliant’ or ‘non-compliant’ with a predictive accuracy higher than the current inspections efficiency of 50%, finally set to 65%.

4.1 HLI-IIS data

Primarily, it was realized that the HLI maintains loads of data regarding the actions taken after an inspection visit, such as imposed fines and other administrative measures, objections, appeals, court decisions, etc. Clearly, though, all these actions are highly related to human decisions and are affected by various unregistered factors, so, it was finally decided to leave them out from this data mining analysis, whose scope is mainly to predict the discoveries of inspection visits and not the terminal outcome of each inspection case. Therefore, only the findings and discovered violations of past inspections were decided to be used for classification purposes in this study, i.e., the class attribute will rely on the onsite findings and infringements and not on any data related to further continuity of the inspection cases.

Table 1 describes the INSPECTIONS table containing all the inspection cases inserted into HLI-IIS since it was put in place. Each inspection case is associated with a specific company and branch, and its type and initiation trigger are indicated. Regarding the type of inspection, considering that this study is focused on addressing undeclared work and other labor relations violations that lie in inspection objects of only one of the two types, this analysis used the past inspections data of only the labor relations inspection services. The occupational safety and health inspections were omitted since they are occupied with labor accidents and labor safety conditions. Consequently, the type of initiation trigger that normally takes four values (see Table 1) concludes to be taking two in our study (i.e., complaint/scheduled), since accident investigations are not included, and labor disputes constitute another type of labor relations inspection that is related to solving issues between employer and employees and is not examined within the scope of our study.

The findings and infringements of all the inspection visits exist in the INSP_FND and INSP_INFR database tables respectively (see Tables 2 and 3). An inspection visit can consist of multiple findings and infringements, with only the second ones to be considered negative discoveries and require further investigation. If an inspection visit is registered to contain only findings, then the employer was found compliant with labor law provisions. Yet, inspection visits to non-compliant employers (i.e., with infringements registered) may include some (positive) findings as well.

In light of this and the main scope of the study to discriminate between compliant and non-compliant employers, the most important data selection criterion for our analysis was to include in the final dataset one entry for each inspection visit when the employer was found compliant, and one entry for each infringement when the employer was found non-compliant to labor law. The reasoning behind this criterion lies in the data mining goal of discovering different patterns of employers non-compliance for different types of violations (e.g., undeclared work, underdeclared work, other violations, etc.), whereas we are not interested in discovering different patterns of compliance for different findings, but only for the compliance existence itself.

The last interesting data source from the HLI-IIS for our project is the COMPLAINTS database table, given in Table 4. This table holds the filed complaints as described in §3.3.2, with each one of them likely connected with an inspection visit. The purpose of using this information in our analysis is to investigate the correlations between the inspections outcome and the complaint data if any.

For each entry in the dataset, some data related to the corresponding company and its employment status are retrieved from the Employment-IS and integrated to analyze possible correlations of those features with the inspection discoveries. Table 5 describes the EMPLOYERS database table with all the registered employers since March 2013 from where interesting attributes for our study can be obtained, such as the legal form, the starting date, and the business sectors of the company.

Employers may have multiple branches registered, which are all saved in the EMPL_BRANCHES database table, illustrated in Table 6. This table can offer information about the specific business sector of the branch and its area.

Concerning the employment details, these are obtained from the various types of declarations the employers are obliged to submit at specific dates or deadlines, according to the labor law regulations. All these employers declarations exist in the EMPL_ DECLARATIONS database table described in Table 7. The volume of data in this table is huge, considering that a company can submit limitless declarations for each of its branches and based on its needs.

More specifically, the Employment-IS offers around 50 different types of electronic services to the employers who are required to use each one of them on different occasions to keep the state informed about their business employment. Yet, this study cannot examine the effects associated with such a large diversity of declaration types, but its interest focuses mostly on the types that modify the working schedule and wage of the employees, which are the major characteristics of a company employment status. Each of these declarations can refer to changes in the employment of one or more employees of the company branch. Thus, an entry in Table 7 can be connected with one or more entries in the DECL_CHNG_EMPLOYEES database table (see Table 8) that contains the declared employment changes for each employee.

While looking into the data and selecting the sources of information based on their relevance to the goal of our project, the researchers and the HLI IT personnel were investigating their quality as well, mainly with respect to the databases tables connections and keys, the attributes missing values, and the values formatting and encoding. In general, the databases were considered well-structured, with no major issues detected, and with the tables connections operating well in the integration stage of the selected data fields, as described in the following section. Some missing values and a few wrong (future) dates were detected, yet they didn’t notably affect the evolvement of the analysis.

5.1 Data selection

Selecting the data to be fed into a data mining model is, one can say, the most significant part of the process since the obtained results depend totally on the data quality and relevance to the project objectives, the distribution of the attribute values, the limits on data volumes, and many other features, with each one being able to affect the quality of the results drastically. Thus, data selection should be made with care, both in the dimension of attributes but also of records – i.e., in columns and rows respectively – and inclusion or exclusion criteria should be documented in detail.

Summarizing here the selection criteria outlined in the previous section, the dataset was defined to be structured over the findings and infringements of the past labor relations inspection visits. Namely, each record in the dataset corresponds to either an inspection visit encountering a labor law compliant employer, thus, its ‘result’ attribute is marked as ‘NO_INFR’ – i.e., no infringements, see also §5.6 and Table 10 – or, to an infringement of an inspection visit detecting a non-compliant employer, with its ‘result’ containing the type of the infringement (see Table 3). Regarding the employment data, these are selected based on the employers declarations of those types that report/modify the employees weekly working hours and wages.

Beyond the above criteria, the time period of data collected and used is another type of criterion to be identified as well. In this study, the decision was grounded on two facts: the first was the obligation starting date of the inspectors inserting their inspection data into the HLI-IIS, which was the beginning of 2018, hence ensuring that the data would be more complete and accurate after that date; the second was the pandemic of Covid-19 appearing to vastly impact the Greek labor market after March of 2020, thus employment data of that year were widely affected by an overly uncommon factor, whose effects, though, lie out of the scope of this study. Consequently, here, only the data of past inspections conducted in 2018–2019 were chosen to be used.

The last data selection criterion applied concerned the area of inspection, opted to be Attica, the broader area of the country capital. The reasoning behind this choice rests on the proven broad diversity of employment trends and patterns between the various regions of the country, with this variety being dominated by multiple seasonality and locality factors, such as tourism, or local festivities at specific periods. Hence, the area of Attica was a safe decision for our initial application of Associative Classification in this domain since it does not exhibit substantial differentiation in employment trends and figures within the year, and data collection with this criterion can supply an adequate volume of information for our analysis.

Fitting the aforesaid criteria, 18.988 records concluded to be collected, with around half of them corresponding to inspection cases of labor law compliant employers and the rest to represent violations found to non-compliant companies.

5.2 Data integration

As previously explained, the dataset was determined to be constructed on the basis of findings and infringements (Tables 2 and 3), and the rest of the attributes considered interesting, to be accordingly integrated. Thus, data fields related to the inspection visit were merged, such as the date and time, the day and type of day, the total workers found, and the initiation trigger (Table 1). The source and the objects of the complaint were also integrated (Table 4) but contained data only in the cases where the inspection is triggered by a complaint. Other information joined pertained to the company (Table 5) – such as the legal form, the business starting date, and the types of business industry – and some others to the inspected employer branch (Table 6), i.e., city, municipality, region, and the branch business sector.

Employment information was challenging to be combined in the dataset since it had to be drawn from the employers declarations (Table 7) and the changes in employment details (Table 8). Yet, both of those tables hold numerous entries connected with an inspected employer branch, which, in this study, constitutes the inspected entity.

After consultation with the domain experts of the Authority, it was identified that obtaining some meaningful insights into the employment status of the inspected entity required a summarization of these data. So, for each dataset entry, the average weekly working hours, and the average hourly wage of all the employees of the inspected location were calculated based on all the declarations submitted in the six months period before the inspection. In addition, the sum of all these declarations was also joined in the dataset as a new attribute. Regarding the total number of employees in the company and of those in the inspected branch-entity, these two figures were taken from the last declaration before the inspection.

Last, the researchers wished to integrate information related to the compliance level of the inspected entity, prior to the last inspection. Actually, they wanted to investigate if the employers are affected by inspection visits and to what degree they subsequently change their level of compliance with labor law. Hence, the sum of inspections conducted before the inspection date was calculated, as well as the sum of detected violations in these inspections, in the period of 2018–2019. These figures, later on, were used to indicate the level of past labor law compliance of the inspected entity.

5.3 Data cleaning and anonymization

Following the integration of all attributes considered interesting in relevance to the project objectives, data cleaning was required before proceeding with the next stages. Few records were observed with no values in some attributes, possibly due to momentary technical complications in the software application saving process, which were finally omitted from the dataset, concluding with 18.548 records.

Anonymization was demanded next by the DPO since any data linking companies details with inspections findings could not be made known to the researchers. Therefore, attributes related to identifying companies and their branches, such as Tax ID, name, and address, were omitted from the dataset, which, at this stage, concluded to contain the attributes as illustrated in Table 9.

At this point, the researchers obtained a dataset collected using some selection criteria identified at the data understanding phase per the project objectives. Yet, this dataset is comprised of 25 attributes (Table 9), which are way too many for a model to process and produce interpretable patterns of employers compliance and non-compliance to labor law. The next step at this stage was, then, a more detailed examination of each one of them and the final selection, based on their interestingness and importance to this study, to come up with a smaller set of attributes.

The primary attributes were selected either to participate themselves in the data mining analysis or to be used in creating other derived attributes for the analysis. Principally, as explained earlier, the attribute ‘result’ plays the role of the ‘class’ in our classification problem and thus was firstly selected since the classifiers to be produced aim to predict the result of the inspection visits. Among the inspection features, the ‘starting time’, the ‘day’, and the ‘type of trigger’ were picked as most intriguing, to check if and how they can be associated with specific violations. If the analysis were to focus specifically on seasonal employment, the ‘inspection starting date’ attribute could have been used as well, after clustering in seasons. Yet, in the selected area of Attica, the ratio of seasonal businesses is not high, denoting that seasonality itself would have a low impact on employment figures and trends in this area along the year. For this reason, the ‘inspection date’ was not preferred for inclusion in the final dataset, compared to other attributes.

Regarding the ‘source’ and ‘objects’ of complaints, these were not of much concern for further investigation to the Authority, and also, values existed in only 19% of the records, i.e., in the cases of initiation due to a complaint (see ratio in Table 10), hence they were omitted from the analysis.

From the company-related attributes, the ‘legal form’ was selected aiming to discriminate the business leadership between single-person and directing-board and explore the way this feature is correlated with specific levels of compliance with labor law. The ‘business starting date’, serving to distinguish between old and new companies, could have been used similarly, yet its values range was identified as very broad and tricky to be partitioned, so it was finally neglected.

Among the four available attributes in the dataset related to the ‘business sector’, that of the inspected entity was chosen, i.e., of the inspected branch of the employer, being the most relevant to the specific inspection case.

The attributes identifying the inspection location area are the ‘city’, the ‘municipality’, and the ‘region’ of the branch, out of which the ‘region’ was selected to further participate in the analysis since its values were less and didn’t need further categorization. Further on, the ‘total number of employees in the branch’ – and not ‘in the company’ – was chosen to identify the size of the inspected entity, and, ending, the rest of the attributes related to the employment status and the level of past labor law compliance of the inspected entity, were all chosen to be used in the construction of derived features, as analyzed beneath.

Derived attributes are considered the new ones that are constructed using one or more of the existing ones. In this study, we built two new attributes that the domain experts were curious about to see how they can be associated with the results of an inspection: the level of past compliance to labor law and the frequency of changes in employment.

The level of ‘past compliance’ was calculated by dividing the sum of detected past infringements by the sum of past inspections – if any – in the inspected entity. Yet, in only around one-third of the cases, previous inspections data were available, whereas, in two-thirds of the cases, the entity was being inspected for the first time, as later noticed.

The ‘frequency of changes’ corresponds to the number of changes in employment details per employee in the inspected entity, in the last semester prior to the inspection. The labor inspection experts believed that this attribute reveals non-compliance when it gets high values, i.e., in the cases of often submissions of changes, though in our first machine learning application in this domain [8] exactly the opposite was proved.

5.6 Discretization and clustering

Most of the primary and derived attributes, as presented earlier, take either numeric values or values from a wide range of category codes. Yet, Associative Classification, the modeling technique we are using in this study, as also many other machine learning methods, requires the selected attributes to be taking discrete values. Thus, before building the classifiers, one last task is necessary, discretization, in which the values range of numeric attributes is partitioned into different non-overlapping intervals, with each one being portrayed by a discrete value. Similarly, the multiple category codes of the categorical attributes must get clustered, i.e., aggregated into groups according to a homogeneity feature, with each group being represented by a single value.

Discretization of quantitative data values and clustering of qualitative data values demanded an in-depth understanding of their meaning and consultation with the labor inspection experts since an arbitrary grouping of data values would produce meaningless data mining results.

Starting from the values of the ‘result’ as the most significant attribute and considering the study focus on addressing undeclared work, the around 200 category codes corresponding to the different infringements were clustered into three main groups: undeclared work (UDW), underdeclared work (UNDER_DW), and other infringements (OTHER_INFR). These three values plus the no infringements (NO_INFR) constitute the new categorical values of this attribute, which, as said, plays the role of the class. Keeping this categorization, we have a four-class classification problem, whereas grouping the violations categories into one, the infringements (INFR), we transform it into a binary, i.e., a two-class problem. In the following section, we are studying both types of classification problems, in combination with the two selected AC algorithms.

The ‘inspection time’ was easy to be partitioned into three 8-hours time zones, the ‘inspection day’ to be split into weekday and weekend, and the around 40 category codes of the ‘legal form’ attribute to be segmented into Sole proprietorship and Corporation, determining the business leadership. Similarly, the ‘business sector’ numerous codes of the inspected entity were divided into four main groups: Hotels-Restaurants-Catering, Production-Construction, Sales, and Services. The ‘region’ categorical values were seven constituting the different parts of Attica and did not need further grouping.

The ‘entity size’ was created by discretization of the quantitative attribute of the ‘total number of employees in the branch’ relying on labor regulations defining the ranges of each of the categories: small, medium, large, and very large size as illustrated in Table 10.

The employment-related attributes, i.e., the ‘average weekly working hours’, the ‘average hourly wage’, and the constructed ‘frequency of changes’, all taking numerical continuous values, had to be discretized as well. Thus, the new feature ‘employment’, illustrating the level of average working time in the inspected entity, was created to take the values: low employment for up to 16 hours per week, medium employment corresponding to 17–32 hours per week, and full employment for more than 32 hours per week. Accordingly, the new attribute ‘payment’ exhibiting the level of average payment in the branch was generated to pick values from: low, medium, high, and very high paid, with their ranges as displayed in Table 10. Last, the ‘frequency of changes’ was partitioned per the labor inspectors perception and expertise ending with the categories: rare changes with up to 2 changes per employee within the last semester before the inspection, medium frequency changes matching to 2–4 changes per employee, often changes and very often changes with ranges 4–10 and more than 10 changes per employee accordingly.

Lastly, the derived numerical attribute ‘past compliance’ had to be turned into categorical. For values above zero, the range was split into four intervals as shown in Table 10, matching to categories: low, medium, high, and very high non-compliance. Along with these, two more categories participated: compliant, to correspond to the cases that no infringements had been detected in the past inspections, and uninspected, to be attributed to those cases where no past inspection had been registered for the inspected entity.

At this point, having completed all data preparation steps, we have constructed a dataset of 12 categorical attributes and 18.548 records. Table 10 presents the final attributes, their categorical values with their corresponding ranges – at the numerical attributes – and their ratio of existence in the obtained dataset.

6.1 Modeling technique

In this study, Associative Classification (AC) is put forward for building our classification models for various reasons. Primarily, studies [19, 20, 21] have proved that this innovative machine learning method is able to create classifiers with higher predictive accuracy than other traditional data mining techniques, such as decision trees [22, 23] and rule induction [24, 25]. Foremost though, the appropriateness of the method relies on its core characteristics that assure fulfillment of our project goals. Among them, AC produces very simple knowledge in the form of rules that are easily interpretable and manually updated by the end-user if need be [19]. Moreover, this data mining approach often discovers further hidden knowledge neglected by other classification methods, hence succeeding in minimized error rate. This feature derives from Association Rule Mining (ARM) [10] being used in the training phase of AC, where all possible correlations among the attribute values and the class values in the dataset are identified and extracted [19].

AC algorithms normally follow three main phases. In the first training phase, it generates the Class Association Rules (CARs), i.e., rules in an if-then form revealing the connections among attribute values and class values. In the second pruning phase, ranking and pruning processes are activated to exclude weak rules from the set of CARs, according to certain user-defined thresholds. At this phase, opposing and replicating rules are also omitted. The last testing phase aims to examine the classifier on a new separate dataset and assess its predictive accuracy and error rate [19].

The AC problem is defined by Liu et al. as follows [11]:

‘Let D be the dataset. Let I be the set of all items in D , and Y be the set of all class labels. We say that a data case d ∈ D contains X ⊆ I , a subset of items, if X ⊆ d . A class association rule (CAR) is an implication of the form X → y , where X ⊆ I , y ∈ Y and I ∩ Y = ∅ . A rule X → y holds in D with confidence c if c % of cases in D that contain X are labelled with class y . The rule X → y has support s in D if s % of the cases in D contain X and are labelled with class y ’.

In this study, we are using two AC algorithms, CBA and CBA2, whose key features are described below.

6.2 Training, testing, and assessment design

As earlier explained, in this project, we are using two types of datasets, one with two classes and one with four classes, plus we are applying two AC algorithms, the CBA and the CBA2. The combination of these two variations generates four different models: (a) CBA with the two-class dataset, referred to as CBA-2class further on, (b) CBA2 with the two-class dataset, called as CBA2-2class, (c) CBA with the four-class dataset, named as CBA-4class, and (d) CBA2 with the four-class dataset, called as CBA2-4class, in the analysis below.

All four above classifiers were built using LAC [31], a Java Library for Associative Classification. The system was running on OS Windows 10Pro 64-bit, with an Intel Core i5-10210u 2.3 GHz processor and 8 GB memory. For the parameters setting of both CBA and CBA2 algorithms, the authors recommendations were followed [11, 12], i.e., the minimum support was set to 1%, and the minimum confidence was set to 50%. Tables 13–16 display the classification results of the models CBA-2class, CBA2-2class, CBA-4class and CBA2-4class respectively.

Table 17 summarizes the measurements at all the evaluation metrics for all four models, where, for the two four-class models, it has been regarded that when a data instance of an infringement class is predicted to fall into a different infringement class, then this is attributed to True Positives (TP), following the structure of Table 12.

In Table 18, though, the evaluation figures are calculated again for the two four-class models, but here all infringement classes are considered as mutually exclusive, i.e., a positive instance is regarded as True Positive only when it is classified into its own infringement class.

Followingly, the classification results and the prediction measurements of each of the four models are analyzed and compared with reference to the data mining goals and success criteria.

7.1 Models assessment per the business goals

In the previous section, the four generated models were assessed per the data mining goals, whereas this last phase deals with the overall evaluation of the project, also with respect to the initial business objectives. Be recalled that the Authority initial objective was to embody into its business processes innovative technology solutions succeeding in two major goals. The first was the increased efficiency of onsite visits through the use of a prediction-recommendation system, which should be automatically configured and updated and contribute to a more effective allocation of the Authority resources; the second was the provision of simple and comprehensible knowledge to the labor inspectors in terms of patterns of labor law compliance and non-compliance that would improve their ability to perceive the system recommendations for targeted onsite inspections and, at the same time, to recognize swiftly cases of employers likely to be involved in undeclared work and other violations.

Out of the four generated models, as thoroughly prior explained, only the first two that were built using the two-class dataset are succeeding in both the business goals, with the CBA2-2class model exhibiting a slightly improved overall prediction performance and offering more and exhaustive knowledge about those characteristics discovered to be linked with compliance and non-compliance to labor regulations. Thus, if one had to choose only one of the four models to be applied, this should be CBA2-2class, which offers a complete and balanced set of readable and understandable rules classifying employers into compliance (NO_INFR) and non-compliance (INFR) classes, with an accuracy of 67.15% for both classes.

Yet, this model cannot offer an enhanced knowledge with regard to the specific patterns associated with each of the infringement sub-classes, i.e., undeclared work (UDW), underdeclared work (UNDER_DW), and other infringements (OTHER_INFR), as the four-class models can. Especially the CBA2-4class model manages to produce CARs for all classes, even for the undeclared work whose ratio in the dataset is very low, at 4%. This accomplishment should not be overlooked since this model succeeds the knowledge provision business goal at a much higher level than the selected CBA2-2class model, even if it cannot reach the set accuracy threshold of 65% of the initial data mining goal. Particularly, Table 22 proves that the patterns of labor law non-compliance can be differentiated for each infringement class, and the provision of such detailed insight into the different types of violations is of significant value for the Authority. Must be reminded, at this point, that one of the HLI major issues is the labor inspectors unwillingness to correspond to the current risk analysis tool inspections suggestions, mainly due to the lack of knowledge of the tool functioning and the factors affecting its results. If a newly established inspections recommendation tool does not effectively address this issue by helping the labor inspectors thoroughly perceive the various patterns of labor law non-compliance and actively involving them in the inspections selection process, it highly risks being gradually abandoned, as well.

Considering the above business risk of the tool disuse, the CBA2-4class model should be unquestionably integrated into the inspections recommendation tool, all the more so if the business goal of improved allocation of the Authority human and financial resources is further contemplated. Every inspection visit is associated with a high operational cost which should be taken into account while deciding which of the tool suggestions necessitate initiating a targeted onsite inspection. With this model Precision at 73.37% – i.e., its correctness of infringements predictions – and notably that of undeclared work lying at 83%, its suggestions should not be questioned, but they should all trigger an inspection visit hence increasing the likelihood of violations detection.

7.2 Determination of best option

Concluding from the aforementioned, a combination of the CBA2-2class model displaying the highest overall accuracy at 67,15%, with the CBA2-4class model exhibiting the peak Precision at 73.37% and supplying the most detailed knowledge regarding patterns associated with all infringement classes, stands as the best option to be followed in setting up the required inspections recommendation tool.

The tool may consist of two classifiers, one built by the CBA2-2class model reaching an adequate and stable predictive accuracy both for the positive but also for the negative class, thus, assisting the Authority in planning on-site inspections at a larger scale, and one made by the CBA2-4class model supplying simple and thorough understanding of those characteristics linked with the different types of infringements.

Knowing the pros and cons of each model, the tool may be set to use both of them to classify employers and benefit accordingly. When both the two models predict a company to be positive – i.e., with infringements – then it should be prioritized to be inspected unhesitatingly, whereas if they classify it into negatives – i.e., compliant – then it can be ranked as less risky for conducting an onsite inspection. Cases of employers classified as non-compliant (INFR) by the CBA2-2class model, but the 4class model foresees them as compliant (NO_INFR), should be candidates for an inspection visit, yet ranked per their classification rule confidence value. Notably, all inspections suggestions of both the two models should come along with their corresponding activated classification rule and its confidence value, supplying the tool users with the necessary knowledge, first, to perceive the reasons – i.e., the attribute values – for which a company is suggested for a targeted inspection and second, to be aware of how confident this suggestion is.

In this way, the labor inspectors will be highly stimulated and actively involved in the whole inspections selection process, improving their domain knowledge and violations identification ability, while, at the same time, they will still be in authority over deciding which of the tool suggestions to follow. Thus, the inspectors will perceive this inspections recommendation tool as a helping means to improve their productivity and not as yet another complicated system producing inexplicable employers ranking lists, which they are inevitably obliged to follow.

The tool will be able to automatically update its classification mechanisms through an automated periodic process of rebuilding the classifiers, by including also, each time, the new inspections data that were inserted into the Inspections subsystem of the HLI-IIS. For this reason, the labor inspectors are suggested to keep conducting a proportion of inspection visits randomly selected so as to give the tool the chance to discover new patterns of labor law non-compliance. Consequently, the tool will need no manual configuration or update, but it will function completely automatically, offering unbiased inspections recommendations, thus contributing to the enhancement of the labor inspectors trust in it.

Closing, the combination of the two models, as explained above, incorporated in a recommendation system, achieves maximum success both in business and in data mining goals and is suggested as the best option in this data mining application.

7.3 Next steps

Following the in-depth analysis of this data mining project through the different stages of the CRISP-DM methodology, several parts were identified that could be further examined and allow an even more meticulous understanding of the models produced knowledge, as well as an increased prediction capability.

More specifically, for every task to be completed, various decisions had to be taken from the researchers of this study in close cooperation with the Authority involved personnel, which were mainly based on their domain expertise and thorough investigation. For instance, starting from the data selection and integration, proceeding with the primary attributes selection and the creation of the derived attributes, till the discretization and clustering tasks, the dataset preparation phase required several choices to be made, where, evidently, different variations would produce differentiated results. Similarly, in the modeling phase, other algorithms could be used which, combined with a diversity of dataset structures, would create a variety of changed models to be analyzed and assessed.

Under this view and per the CRISP-DM methodology, the present analysis constitutes a complete life cycle of our data mining project, which, eventually, may trigger new, more focused business goals and another life cycle, benefitting from the gained experiences and lessons learned.

Particularly, in this cycle, the two models with the four-class datasets were identified to not perform well in detecting an adequate percentage of actual infringement cases, both reaching a Recall measurement of under 40%, hence they could not be considered effective to be used as a core classification model. Yet, their vital knowledge provision constitutes them beneficial, thus they are regarded worthy of a supplementary examination, especially the CBA2-4class model, which is able to handle more effectively the unevenly distributed multi-class datasets.

Grounded on the acquired experience of the first life cycle of this data mining project, the researchers presume that the four-class models prediction capability might have been affected by the dataset construction design leading to the existence of several instances with identical attribute values but with different infringement classes – i.e., when inspection cases pertained to a non-compliant employer with several violations – which must have impeded the models from constructing several strong and accurate rules for each infringement class.

For this reason, a new iteration initiation of the data mining project life cycle was identified as one of the next steps, to examine the predictive potential and the discovered knowledge of the four-class models, but with a variation in the current dataset structure, so as only one data instance per inspection case should exist in the final dataset. This practically means that the cases of non-compliant employers detected with multiple violations should be represented in the dataset by one data record and labeled with the most severe infringement class out of the discovered ones. If the infringement classes are ranked per severity, i.e. UDW → UNDER_DW → OTHER_INFR, then probably the percentage of undeclared work will rise in this new dataset since the total number of data instances will decrease, yet the undeclared work cases will remain the same. Hence, this dataset construction design seems to give undeclared work a higher priority, and the inspection cases detecting it are distinctly labeled only with the UDW class in the dataset, even if they contained other violations as well.

Concluding, it is estimated and remains to be proved that this new construction variation of the dataset, containing only one record per detected non-compliant employer being labeled with his/her most severe violation, might allow the four-class models to generate more accurate and stronger CARs for each infringement class and achieve higher prediction capability.

Additionally, other life cycle iterations could include the analysis and comparison of other AC algorithms, such as the Classification based on Multiple Association Rules (CMAR) [20], the Classification based on Predictive Association Rules (CPAR) [21], the Multiclass Associative Classification (MAC) [32], the Live and Let Live ( L ⁢ 3 ) [33], and others.