Modelling Assignment: Implementation Report For An Ontology-Based AI Job-Matching Service

Introduction

The rapid evolution of job markets has increased the demand for efficient and accurate job-matching systems. Traditional methods often struggle to personalise matches, resulting in inefficiencies and missed opportunities (Chen, 2023). Ontologies provide a formal framework for organising and reasoning about domain knowledge (Stumme, 2002), making them well-suited for an Artificial Intelligence (AI) driven job-matching service. This report outlines the development of a prototype ontology using Protégé, demonstrating its effectiveness in connecting candidates with appropriate job opportunities. The report also explores the ontology’s impact, scalability, limitations, and potential for real-world applications.

Background and Justification

Job-matching is the dynamic process of aligning candidates with jobs within an organization (Weller et al., 2019). The job market is characterised by different skills, and preferences thus creating a complex matching problem and the risk of mismatch. Traditional methods commonly rely on keyword searches or manual screening processes, which can introduce errors, exhibit bias, or fail to account for complex relationships like compatibility with job preferences (ExactBuyer Blog, n.d.). Ontologies provide a structured way to represent knowledge, enabling reasoning about entities and their relationships. For a job-matching service, this means capturing data about jobs, candidates, and their properties in a formalised schema.

Ontology-based systems can represent complex relationships, such as skills required by jobs. They can also enable reasoning, such as inferring matches based on overlapping skills. Job preferences such as work arrangements can also be used to personalise recommendations for candidates.

This prototype ontology was designed with the following components:

• Classes: Entities such as Job, Candidate, Skill, and WorkPreference.
• Subclasses include FullTimeJob, and PartTimeJob, for added granularity. Figure 1 shows the list of created classes in Protégé.
• Object Properties: Relationships such as ‘requiresSkill’ (Job -> Skill), hasSkill (Candidate -> Skill), and posts (Company -> Job) define connections between entities. Figure 2 shows the list of object properties in Protégé.
• Data Properties: Attributes such as experienceInYears (Candidate) and experienceRequired (Job) capture relevant details. Figure 3 shows the list of data properties in Protégé.
• Reasoning: Property chaining (e.g. between ‘hasSkill’ and ‘requiresSkill’) enables automated reasoning in matching candidates to jobs based on skill overlap.

These design decisions align to enhance job-matching efficiency and personalisation.

Protégé Implementation

The following was the process involved in building and testing the ontology using Protégé (version 5.6.4): Defining core classes and subclasses.

• Creating object and data properties to represent relationships and attributes.
• Updating appropriate domains and ranges for the object properties.
• Adding individuals (instances of classes/subclasses) to test the ontology.
• Applying property chaining on object properties to enable automated reasoning using the HermIT reasoner in Protégé.
• Testing using the DL Query.

Key features of the ontology are as follows:

• Classes: Job, Skill, Company, Candidate, WorkArrangement
• Sub-classes: FullTime, PartTime, Design, Management, Programming, Hybrid, Onsite, Remote
• Object Properties: hasSkill, requiresSkill, posts, hasWorkArrangement, prefersWorkArrangement, matches
• Data Properties: name, email, jobTitle, experienceInYears, experienceRequired, postedDate, workHours
• Individuals: Candidate_1, Candidate_2, Candidate_3, ABC_Company, XYZ_Company, Data_Analyst, Machine_Learning_Engineer, Product_Manager, Python, Agile, Figma, HybridWork, RemoteWork

See Figure 4 for the representation generated by OWLViz in Protégé.

The initial challenges of working on the ontology involved balancing granularity and simplicity and maintaining consistency in property definitions. These issues were resolved through iterative testing and validation.

Results

• Testing Procedure

The ontology was tested using queries in Protégé’s DL Query and SPARQL tools. DL Queries were used to test class relationships and inferred properties while the SPARQL queries were used to verify specific instances and retrieve data across the ontology. The test cases included: Find jobs with remote work arrangements. Listing all jobs available. Finding jobs that required specific skills. Finding candidates qualified for specific roles. Inferring a match based on skills and work preferences.

• Outputs

• Query: Find jobs with remote work arrangements. Result: Machine_Learning_Engineer was correctly retrieved
• Query: List all available jobs. Result: Machine_Learning_Engineer, Product_Manager, and Data_Analyst were correctly retrieved
• Query: Find jobs requiring Python as a skill . Result: Machine_Learning_Engineer and Data_Analyst were correctly retrieved
• Query: Find candidates with Python in their skill set . Result: Candidate_2 was correctly retrieved
• Query: Infer a match based on skills and work preferences . Result: Reasoning inferred Candidate_2 is matched with Machine_Learning_Engineer based on skill and work preference overlap.

See Figure 5 for DL and SPARQL queries and results in Appendix.

The ontology effectively supports job matching through logical reasoning. Its structure facilitated accurate matches.

Critical Assessment

Strengths

• Transparency: The ontology-based reasoning provides clear explanations for matches.
• Scalability: The ontology can be scaled to include new skills, job types, salary preferences, location, etc.
• Personalisation: Captures candidates’ work preferences, enhancing user satisfaction.

Limitations

• Data Quality: The effectiveness depends on the accuracy and completeness of input data.
• Maintenance: Updating the ontology to reflect market changes can be resource-intensive.

Recommendations

Future iterations could include:

• Integrating machine learning models to refine recommendations.
• Automating ontology updates based on real-time data.
• Expanding the ontology to support multilingual job descriptions and global markets.

Conclusion

This modelling project demonstrates the utility of ontologies in building AI-driven job-matching systems. The ontology supported personalised and efficient job recommendations by organising domain knowledge and enabling reasoning. While challenges such as maintenance remain, the approach’s scalability makes it a valuable component of AI applications in recruitment.

References

• Chen, Z. (2023) ‘Collaboration among Recruiters and Artificial intelligence: Removing Human Prejudices in Employment’, Cognition, Technology & Work, 25(1), pp. 135–149. Available at: https://doi.org/10.1007/s10111-022-00716-0.
• ExactBuyer Blog (no date) blog.exactbuyer.com. Available at: https://blog.exactbuyer.com/post/ai-traditional-job-matching-accuracy-comparison.
• Stumme, G. (2002) ‘Using ontologies and formal concept analysis for organizing business knowledge’, in J. Becker and R. Knackstedt (eds). Physica-Verlag HD, pp. 163–174. Available at: https://doi.org/10.1007/978-3-642-52449-3%E2%82%84.
• Weller, I. et al. (2019) ‘How matching creates value: Cogs and wheels for human capital resources research’, Academy of Management Annals, 13, pp. 188–214.

Appendix

Figure 1 Figure 1 shows the list of created classes in Protégé.

Figure 2 Figure 2 shows the list of object properties in Protégé.

Figure 3 Figure 3 shows the list of data properties in Protégé.

Figure 4

Figure 5: Queries and results