iModuleBuddy - A Hybrid AI-Based Academic Planning System

Abstract

iModuleBuddy is a study planner that helps postgraduate students create personalized study plans. It combines course recommendation with long-term planning and considers students' professional background, career goals, and individual study preferences. The system integrates structured data from the ESCO ontology and course descriptions with vector-based similarity methods and retrieval-augmented generation (RAG). A key component is the JobRanking algorithm, which prioritizes courses based on the relevance of a student's career history. The system uses a multi-agent architecture: one agent aligns professional experience with suitable courses, while another organizes these into a multi-semester plan. Based on user input, iModuleBuddy generates diferent study plans-career-focused, balanced, and preference-based-along with explanations of how the recommended courses contribute to career development. The system is currently under development, with the career-focused plan already implemented and the other variants in progress.

1. Introduction

The availability of flexible and part-time study programs reflects an institutional efort to attract students with professional backgrounds and support those who wish to continue working while earning a degree [ 1 ]. While universities provide flexible degree options, students are responsible for planning their studies and balancing academic requirements with personal and career interests [ 2, 3 ]. This planning involves both short-term course selection and long-term strategies to meet degree requirements while considering academic prerequisites and individual preferences. Challenges include limited personalized guidance, high student-to-advisor ratios [ 4 ], and frequent course updates [ 5 ]. Consequently, students often struggle to create study plans that align with their professional experience and aspirations, as current planning systems primarily focus on compliance with regulations rather than individual career goals [ 2, 6, 7, 8 ].

In prior works, we explored several approaches to support students in academic planning. [ 9 ] created a performance prediction model that uses course description embeddings and measures of student similarity to forecast academic outcomes. The model combines indicators based on students’ academic interests and past grades. However, small datasets and varied course combinations limit prediction accuracy. [ 10 ] proposed a knowledge-based recommender system that links course learning objectives to job-related competencies using the European Skills, Competences, Qualifications and Occupations 2025. ∗Corresponding author. †These authors contributed equally.

CEUR Workshop Proceedings

ceur-ws.org ISSN1613-0073 (ESCO)1 ontology and a Large Language Model (LLM). This system ranks courses based on how well they support skills relevant to selected careers and explains each recommendation. ChatGPT was used to extract learning objectives and assign them to competencies, resulting in some inconsistencies. Diferent competencies were sometimes assigned to the same objectives, raising concerns about reliability and reproducibility. [ 11 ] developed an ontology-based recommendation system that uses semantic representations of course content and student preferences to generate personalized course suggestions. Although the ontology was designed to generate tailored recommendations, further refinement is necessary to improve the accuracy and relevance of these recommendations.

Building on these foundations, this position paper introduces iModuleBuddy, an academic planning system under active development. It is based on the assumption that academic planning should consider various individual and contextual factors. These include students’ professional experience, personal interests, career goals, course content, learning objectives, and scheduling constraints. Grades and similarities to other students were not considered, as they are context-dependent and may not accurately reflect motivations or future potential. iModuleBuddy aims to improve how study plans are generated and builds on earlier research, addressing practical limitations such as reliance on historical data, limited adaptability, and lack of explanation in existing systems. A knowledge graph connects occupations, competencies, and courses through structured representations to address issues such as cold-start problems and shifting student goals, which are common in systems based on past behavior. A RetrievalAugmented Generation (RAG) framework with a LLM accesses relevant course content and provides explanations. To align study recommendations with students’ professional backgrounds, iModuleBuddy employs a JobRanking algorithm that sorts courses based on career relevance, extending earlier eforts to link academic content with job-related competencies. iModuleBuddy produces three study plans: career-focused, balanced, and preference-based. By synthesizing insights from our earlier data-driven, competency-oriented, and ontology-based approaches, iModuleBuddy aims to deliver a flexible and transparent planning experience that integrates students’ career goals, professional backgrounds, and personal preferences. iModuleBuddy is still under active development; the career-focused study plan functionality is implemented, while the balanced and preference-based plans are currently in progress.

This paper is structured as follows: Section 2 overviews current research on study planning systems and identifies the research gap. Section 3 introduces the research methodology. Section 4 highlights the practical relevance of the iModuleBuddy study planner, and Section 5 describes the system architecture and the role of the diferent components. Finally, Section 7 concludes the paper and outlines the next steps.

2. Related Work

Study planning and course recommendation systems have been explored from various angles, including collaborative filtering, content-based methods, and hybrid approaches that combine machine learning and domain knowledge [ 12, 5, 6 ].

Content-based recommendation systems typically align course selection with career goals by analyzing job market data. [ 13, 14 ] developed a tool that maps course content to career-relevant skills extracted from job descriptions. In contrast,[ 15 ] introduced a system that utilizes LinkedIn profiles for course recommendations. Nonetheless, incomplete or biased data may limit the accuracy of such methods. Decision-support systems (DSS) such as IDiSC+ [ 8 ] enable students to create long-term academic plans based on constraints like graduation timelines and budgets but do not consider career relevance. Other DSS solutions [ 16, 17 ] focus on optimizing study duration rather than aligning courses with professional backgrounds. Beyond these approaches, social network-based systems attempt to enhance study planning through peer recommendations. [ 18 ] developed a model leveraging Facebook connections for course suggestions. These systems often fall short when students lack meaningful social connections or require personalized career-oriented recommendations. 1https://esco.ec.europa.eu/en/classification/occupation_main

Regarding applied techniques, AI-driven methods leverage knowledge graphs and machine learning to provide tailored course recommendations. [ 19 ] uses reinforcement learning and graph-based modeling to improve the selection of online courses and dynamically adapt recommendations to students’ evolving learning behaviors. [ 20 ] proposes a hybrid model that constructs course knowledge graphs using rule-based and deep learning techniques, improving recommendation accuracy through structured dependency mapping. [ 21 ] explores using OpenAI’s embedding models with LLMs to refine course suggestions through RAG, demonstrating improved contextual relevance.

Despite recent advancements, most existing systems generate a single study plan, making it dificult for students to assess how diferent course selections might align with their goals or constraints. In addition, there is insuficient integration between course recommendations and overall study planning, treating them as isolated decisions rather than part of a longer-term academic path. While factors such as past academic performance, career aspirations, and personal preferences are often included, students’ professional experience is not considered. Although a considerable number of students with professional experience enrol in higher education. For example, [ 1 ] found that 38% of higher education graduates in a Swiss sample had completed vocational education and training (VET) before entering university, suggesting that these students possessed work experience. Similarly, [ 22 ] found that among an Australian university student sample, 53.6% were employed during their studies, and another 46.4% had worked previously, while only 3.2% had never worked. These findings demonstrate that many students, particularly those entering postgraduate programs, possess professional experience. These ifndings highlight that professional experience is part of many students’ profiles.

Data-driven approaches that predict student performance using historical enrollment data may face challenges, particularly when dealing with sparse datasets or rapidly changing curricula [ 2, 7, 8 ]. These insights suggest that academic planning systems could benefit from expanding their focus beyond academic data by incorporating diverse student backgrounds, including professional experience.

In our earlier work, we explored how textual course descriptions could be coded to improve recommendations, although small data sets and diferent pathways remain a challenge [ 9 ]. Further, our research introduced ontology-based methods to represent academic knowledge, such as learning objectives, prerequisites, and competencies, enabling more powerful queries and structured reasoning [ 11 ]. Additionally, our earlier work demonstrated how linking learning objectives to job-related competencies could enable more targeted course choices [ 10 ]. However, purely knowledge-based solutions can be dificult to scale or maintain, whereas purely data-driven solutions fail to exploit domain knowledge and career-oriented constraints. iModuleBuddy addresses these issues by including professional background information through a JobRanking algorithm and ofers multiple plan variations (career-focused, balanced, preference-based).

3. Methodology

The development of iModuleBuddy follows the Design Science Research (DSR) methodology, which ofers a structured approach for designing and evaluating innovative information systems artifacts [ 23 ]. DSR involves iterative phases: problem awareness, suggestion, development, evaluation, and conclusion. During the problem awareness phase, challenges in course selection were explored through a literature review and World-Café workshops [ 24 ] conducted with postgraduate students from the MSc in Business Information Systems program at the University of Applied Sciences and Arts Northwestern Switzerland (FHNW). Students reported that they find it hard to select courses that will both qualify them for their desired future job and fulfill academic requirements. They also noted limited personalized guidance and a lack of consideration for professional experience. These findings reflect gaps identified in prior data-driven and ontology-based course recommendation research [ 9, 11, 10 ]. In the suggestion phase, key requirements for iModuleBuddy were derived from workshop insights, interviews with alumni, and a review of existing AI-driven study planning systems. Emphasis was placed on integrating professional experience, generating multiple personalized study plans, and clarifying the relevance of each recommended course for diferent career goals. During the development phase, iModuleBuddy was designed to link structured career information with flexible course recommendations. To achieve this, a knowledge graph based on the ESCO ontology is combined with a language model that retrieves and explains relevant courses using a Retrieval-Augmented Generation (RAG) approach. Course relevance is further refined through a custom JobRanking algorithm that takes the student’s professional background into account. The system currently produces a career-focused study plan. Additional plan types—balanced and preference-based—are in development. Evaluation of iModuleBuddy will occur in future stages, assessing both individual components and overall performance once all plan variations are fully implemented.

4. Application Scenario

The MSc BIS study program at the FHNW provides a practical context for developing iModuleBuddy. As a Swiss postgraduate program, it reflects the characteristics of Switzerland’s education system. Most students begin the study program in their late twenties, with a median starting age of 28 and an average of 29.5 years. A few students enter earlier, around age 21, mainly as exchange students from countries where postgraduate studies follow directly after a bachelor’s degree. Admission typically requires at least one year of professional experience, with exceptions for exchange students under institutional agreements. To understand the professional backgrounds of students, 463 LinkedIn profiles from those enrolled between 2013 and the spring semester of 2023 were analyzed. This period includes a total of 1045 matriculated students. The analysis shows that students in their mid-twenties generally have 1–2 years of work experience, while those in their thirties or older often have significantly more. Older students (aged 39–42) report 3 to nearly 15 years of experience, while even younger students (around 24) may already have up to 5 years. This pattern reflects Switzerland’s dual education system, where vocational training often starts at age 15, followed by early entry into the workforce around age 19 [ 25 ]. Figures 1 and 2 visualize the distribution of student age and professional experience.

The MSc BIS curriculum consists of 90 ECTS credits. Of these, 36 ECTS are allocated to elective courses, which students choose from a pool of over 20 options. The study program is ofered as full-time (1.5 years) and part-time (2.5 years) formats. Insights from World-Café workshops, together with an analysis of student enrollment event logs from FHNW’s internal course registration system (2013–2023), show that students create individualized study plans, with few following the same combinations of courses per semester. While core courses are typically taken early in the program, elective course selection varies significantly. The workshops revealed that multiple factors influence course choices, including alignment with career goals, prior knowledge, relevance to professional experience, assessment format, scheduling flexibility, and lecturer reputation. Peer recommendations play a limited role and are typically only valued if they come from trusted individuals. Social influence from friends or former alumni in students’ private networks has little impact on their final decisions. Many students work up to 80% alongside their studies and therefore, balance their workload across the semester. Figure 3 illustrates the course selection process as described by the students. Despite this structured approach, students struggle with the selection process. Many feel uncertain about choosing the right courses or course combinations. In the first two weeks of each semester, when course changes are still permitted, it is common for students to switch courses. Further, students often seek input from the head of the program or lecturers they trust, not only to confirm whether their course choices are appropriate, but also to explore alternative options and to discuss whether they should enrol in certain courses during the same semester or distribute them across diferent ones.

5. System Development and Architecture of iModuleBuddy

This section provides an overview of the iModuleBuddy architecture, describing the key components and processes involved in retrieving, ranking, and generating customized study plans. 5.1. System Architecture of iModuleBuddy iModuleBuddy is designed as a multi-agent system with three layers — (1) data, (2) processing, and (3) application — which work together to generate personalized study plans. This architecture reflects the complexity of academic planning, where students must weigh multiple, and sometimes conflicting, factors such as career goals, professional experience, prior knowledge, personal interests, assessment formats, and scheduling constraints. To support this process, the system uses two specialized agents careerAgent and scheduleAgent, which operate collaboratively within a shared workflow. Each agent focuses on a distinct aspect of the planning task. The careerAgent receives student inputs, such as career goals and prior professional experience, and determines whether occupation-based recommendations are necessary. When required, the agent queries the Neo4j knowledge graph to retrieve relevant courses, considering metadata such as learning outcomes, required skills, and ECTS credits to ensure alignment with the student’s target occupations. Once the relevant courses are identified, the scheduleAgent organizes them into a multi-semester plan, considering prerequisites, degree requirements, and previously completed credits. Finally, the agent refines the semester-level plan into a detailed weekly schedule, incorporating specific teaching sessions, locations, and times to provide students with a concrete academic roadmap.

The data layer holds both structured and unstructured information. This includes course details, professional roles, user profiles, and the ESCO ontology. At its core is a Neo4j knowledge graph that links occupations, skills, and courses. Each course entry includes metadata like learning objectives, prerequisites, and scheduling. The ESCO ontology helps map occupations to skills, supporting course recommendations based on career paths. User profiles store information on completed courses, earned credits, and personal preferences such as the planned study duration.

The processing layer is responsible for retrieving and ranking relevant data. To assess how well courses support specific skills, both course descriptions, including learning objectives, and ESCO skill descriptors are transformed into vector representations using the mxbai-embed-large model from Ollama. Similarity scores are then calculated to evaluate the match between each course and the skills associated with both the student’s target occupation and their past work experience. These scores help identify courses that are most relevant to the student’s profile. A central component of this layer is the JobRanking algorithm [ 1 ], which evaluates job experiences based on duration, recency, and job type. When the careerAgent receives a student’s background, it uses the algorithm to prioritize experiences that are most influential for course selection.

The application layer uses RAG to generate personalized study plans in natural language. This approach allows the system to retrieve relevant course descriptions and align them with the student’s professional background or career goals. It explains how each course supports students’ goals and contributes to their development. The application layer coordinates two specialized agents that produce multi-semester plans based on user input and processed data. A LLM, Claude 3 Sonnet from Anthropic, is used to generate these explanations and assist in matching course content to student profiles. When activated, the careerAgent analyzes the student’s professional background and aligns it with one or more ESCO occupations (e.g., “Enterprise Architect,” “ICT Network Architect”). Using the similarity scores from the processing layer, it identifies courses whose learning outcomes best match the skills required by the selected occupations. It then retrieves relevant courses from the knowledge graph, considering completed credits and prior experience, and ranks them so that the most relevant options appear first. The scheduleAgent refines this ranked list by applying user-defined constraints, such as semester workload limits, scheduling preferences, and course availability (Spring or Autumn). It verifies prerequisite requirements and constructs a multi-semester plan. For instance, a foundational course like “Business Intelligence” might be scheduled before an advanced elective requiring analytical skills. The agent also generates a detailed weekly schedule for the upcoming semester, including specific days, times, and campus locations. Figure 4 illustrates the architecture of the iModuleBuddy study planners.

6. Data Flow and Execution Process

The process begins when the student provides their career goals, prior professional experience, and study preferences, such as the number of courses per semester. The system converts ESCO skills and course learning outcomes into vector embeddings, computing similarity scores to find the best matches for the student’s target occupation. On the Application Layer, the careerAgent gathers all potentially relevant modules from the Data Layer, leveraging the embedding scores and factoring in partial completions or previously earned credits. The scheduleAgent takes the ranked list of modules, applies constraints (e.g., “no more than 12 ECTS in a single semester”), and distributes courses across the planned study duration. It also generates a week-by-week timetable for the following semester. A final multi-semester study plan and a detailed weekly schedule for the immediate term are produced. This plan highlights both mandatory and elective modules, alignment with the student’s career goals, and a feasible timetable. To accommodate diferent student requirements, iModuleBuddy generates multiple study plans. The career-focused plan prioritizes courses essential for the student’s professional aspirations, ensuring that their studies directly support career advancement. The balanced plan combines core and elective courses to foster academic breadth and career relevance. The preference-based plan tailors recommendations to personal interests while ensuring compliance with degree requirements. Each plan explains how the selected courses contribute to specific professional roles or competencies. Figure 5 illustrates how the system organizes courses over semesters and ofers a weekly schedule. planning solution that supports postgraduate students in designing a personal study plan.

7. Conclusion

iModuleBuddy is a hybrid AI-based academic planning system that generates personalized study plans aligned with students’ professional backgrounds and career goals. It combines sub-symbolic AI, such as Large Language Models, with symbolic AI methods like knowledge graphs and the ESCO ontology. By pairing the careerAgent, which aligns course recommendations with professional experience and career goals, with the scheduleAgent, which structures a multi-semester plan and weekly schedule, iModuleBuddy fills a gap in traditional course planning systems: recognizing the importance of students’ professional experience and ofering multiple study plans. The system’s main features, such as vector embeddings for learning outcomes, the ESCO ontology for skill matching, and the JobRanking algorithm, ensure that recommended courses are aligned with the student’s career path. Neo4j was chosen for its lfexible data modeling and eficient retrieval capabilities. However, future versions may utilize ontologydriven graph databases for improved semantic reasoning and scalability. An upcoming evaluation phase will examine how efectively the system’s JobRanking algorithm and skill-based course mapping perform in real-world postgraduate programs, using student interviews and surveys to gauge user acceptance and outcome quality. Overall, iModuleBuddy demonstrates a new approach to academic planning, merging AI-driven personalization with structured knowledge to better serve postgraduate students’ career aspirations. The current development focuses on the MSc BIS program at FHNW. The next step will be to assess how well the approach generalizes to other academic contexts, including possible extensions to programs such as the FHNW School of Music or partner universities.

Declaration on Generative AI

While preparing this work, the author(s) used Grammarly to check grammar and spelling. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.

References

[1] M. E. Oswald-Egg , U. Renold, No experience, no employment: The efect of vocational education and training work experience on labour market outcomes after higher education , Economics of Education Review 80 ( 2021 ) 102065 . doi: 10 .1016/j.econedurev. 2020 . 102065 .

[2] M. A. Z. Khan , A. Polyzou , Session-based Methods for Course Recommendation , Journal of Educational Data Mining 16 ( 2024 ) 164 - 196 . doi: 10 .5281/zenodo.11384740.

[3] S. Van Rossen , E. Kluijtmans , S. van Brussel , M. van Harsel , J. Smarius , E. van der Stappen , Recommender Systems for Students in Flexible Education: An Exploration of Benefits and Risks , in: INTERACT, Springer, 2024 . doi: 10 .1007/978-3- 031 -61698-3_ 18 .

[4] C. Karrenbauer , C. M. König , M. H. Breitner , Individual Digital Study Assistant for Higher Education Institutions: Status Quo Analysis and Further Research Agenda , in: Innovation Through Information Systems , volume 48 , Springer, 2021 . doi: 10 .1007/978-3- 030 -86800- 0 _ 8 .

[5] D. B. Guruge , R. Kadel , S. J. Halder , The State of the Art in Methodologies of Course Recommender Systems-A Review of Recent Research, Data 6 ( 2021 ) 18 . doi: 10 .3390/data6020018.

[6] G. Lampropoulos , Recommender systems in education: A literature review and bibliometric analysis , Advances in Mobile Learning Educational Research 3 ( 2023 ) 829 - 850 . doi: 10 .25082/ AMLER. 2023 . 02 .011.

[7] V. Maphosa , M. Maphosa , Fifteen Years of Recommender Systems Research in Higher Education: Current Trends and Future Direction, Applied Artificial Intelligence 37 ( 2023 ). doi: 10 .1080/ 08839514. 2023 . 2175106 .

[8] A. Mohamed , A decision support model for long-term course planning, Decision Support Systems 74 ( 2015 ) 33 - 45 . doi: 10 .1016/j.dss. 2015 . 03 .002.

[9] D. Mäder , M. Spahic-Bogdanovic , H. F. Witschel , Student Performance Prediction Model Based on Course Description and Student Similarity, in: Advanced Information Systems Engineering Workshops. CAiSE , volume 521 , Springer, 2024 . doi: 10 .1007/978-3- 031 -61003- 5 _ 9 .

[10] N. Beutling , M. Spahic-Bogdanovic , Personalised Course Recommender: Linking Learning Objectives and Career Goals through Competencies , in: Proceedings of the AAAI Symposium Series , volume 3 , 2024 . doi: 10 .1609/aaaiss.v3i1. 31185 .

[11] G. Beuchat , K. Hinkelmann , M. Spahic-Bogdanovic , Ontology-Based Course Recommendation , in: Conference Society 5 . 0 - Innovation for Sustainable and Inclusive Social Good, University of Technology, Mauritius, 2024 . doi: 10 .5281/zenodo.11619278.

[12] R. Burke , Hybrid Web Recommender Systems , in: P. Brusilovsky , A. Kobsa , W. Nejdl (Eds.), The Adaptive Web , volume 4321 LNCS, Springer, Berlin, Heidelberg, 2007 . doi: 10 .1007/ 978-3- 540 -72079-9_ 12 .

[13] S. Tomy , E. Pardede, Course Map: A Career-Driven Course Planning Tool , in: ICCSA, volume 10961 , Springer, 2018 . doi: 10 .1007/978-3- 319 -95165-2_ 13 .

[14] S. Tomy , E. Pardede, Map My Career: Career Planning Tool to Improve Student Satisfaction , IEEE Access 7 ( 2019 ) 132950 - 132965 . doi: 10 .1109/ACCESS. 2019 . 2940986 .

[15] L. F. Lessa , W. C. Brandão , Filtering Graduate Courses based on LinkedIn Profiles, in: Proceedings of the 24th Brazilian Symposium on Multimedia and the Web, Association for Computing Machinery , 2018 . doi: 10 .1145/3243082.3243094.

[16] M. Shakhsi‐Niaei , H. Abuei‐Mehrizi , An optimization‐based decision support system for students' personalized long‐term course planning , Computer Applications in Engineering Education 28 ( 2020 ) 1247 - 1264 . doi: 10 .1002/cae.22299.

[17] M. Khamechian , M. E. Petering , A mathematical modeling approach to university course planning, Computers Industrial Engineering 168 ( 2022 ) 107855 . doi: 10 .1016/j.cie. 2021 . 107855 .

[18] S. Channarukul , N. Saejiem , K. Bhumichitr , R. Jiamthapthaksin , V. Nicklamai , K. Terdvikran , Socialaware automated course planner: An integrated recommender system for university registration system, in: ECTI-CON , IEEE, 2017 . doi: 10 .1109/ECTICon. 2017 . 8096295 .

[19] H. Guan , An Online Education Course Recommendation Method Based on Knowledge Graphs and Reinforcement Learning , Journal of Circuits, Systems and Computers 32 ( 2023 ). doi: 10 .1142/ S0218126623500998.

[20] X. Chen , C. Yin , H. Chen , W. Rong , Y. Ouyang , Y. Chai , Course Recommendation System Based on Course Knowledge Graph Generated by Large Language Models , in: TALE - Proceedings, IEEE, 2024 . doi: 10 .1109/TALE62452. 2024 . 10834324 .

[21] Q. EL Maazouzi , A. Retbi , S. Bennani , Optimizing Recommendation Systems in E-Learning: Synergistic Integration of Lang Chain, GPT Models, and Retrieval Augmented Generation (RAG) , in: SADASC , volume 2167 , Springer, 2024 . doi: 10 .1007/978-3- 031 -77040- 1 _ 8 .

[22] D. Bennett , E. Knight , I. Li , The impact of pre-entry work experience on university students' perceived employability , Journal of Further and Higher Education 47 ( 2023 ) 1140 - 1154 . doi: 10 . 1080/0309877X. 2023 . 2220286 .

[23] V. K. Vaishnavi , W. Kuechler , Design Science Research Methods and Patterns , 2 ed., CRC Press, 2015 . doi: 10 .1201/b18448.

[24] J. Brown , D. Isaacs, The World Café Community, The World Café: Shaping Our Futures Through Conversations That Matter , 1 ed., Berrett-Koehler Publishers , Inc., 2008 . doi: 10 .5749/minnesota/ 9780816676224.003.0004.

[25] Federal Department of Foreign Afairs FDFA, Basic vocational education and training, 2021 . URL: https://www.eda.admin.ch/aboutswitzerland/en/home/bildung-wissenschaft/bildung/ berufsbildung-lehre.html.