CareerScape: Modeling Structured Career Paths

• CareerScape is a comprehensive framework that models career trajectories as structured sequences linking education, skills, jobs, and labor-market institutions.
• It employs heterogeneous graphs, sequence models, and large language model pipelines to drive career guidance, recommendation, and fraud detection.
• Empirical studies using large-scale resume data demonstrate its effectiveness in analyzing mobility, progression, inclusion, and detecting machine-generated profiles.

Searching arXiv for recent and relevant papers on CareerScape-like systems, career trajectories, recommender systems, and graph/LLM career modeling. CareerScape is an Editor's term for an integrated class of socio-technical systems and analytic frameworks that represent careers as structured trajectories linking education, skills, jobs, organizations, and labor-market institutions, and that use those representations for guidance, recommendation, mobility analysis, and, in one explicit graph formulation, detection of machine-generated career histories (Chen, 16 Jan 2026, Yamashita et al., 24 Sep 2025). Across this literature, careers are modeled as sequences of steps, concept trajectories, ESCO- or SOC-coded occupations, or role hierarchies, and are operationalized through heterogeneous graphs, LLM extraction pipelines, recommender systems, and neural predictors (Nadjem et al., 2020, Johary et al., 12 May 2025).

1. Conceptual scope and research strands

CareerScape addresses a recurring problem in career research: careers are neither purely individual preferences nor purely labor-market outcomes. One normative formulation describes career decision-making as a socio-technical problem in which individuals exercise bounded agency while navigating labor market institutions, organizational incentive structures, and information asymmetries, and decomposes trajectories into Wealth, Autonomy, and Meaning, including archetypes as points in $(W, A, M)$-space (Chen, 16 Jan 2026). In parallel, trajectory modeling work treats a profile as a succession of steps and predicts next-step “concepts” rather than only specific jobs, allowing continuation and reorientation to be analyzed within the same formalism (Nadjem et al., 2020).

A second strand treats CareerScape as an integration problem. “Community-Based Data Integration of Course and Job Data in Support of Personalized Career-Education Recommendations” models courses, jobs, and skills as a heterogeneous graph and uses skill communities to bridge the weak lexical overlap between educational and labor-market vocabularies (Zhu et al., 2020). A third strand treats CareerScape as an experiential environment: “The Future Time Traveller Project” uses a 3D virtual world with a 2020-to-2050 narrative, scenario-based missions, and structured reflection tasks to connect future-of-work exploration to personal career thinking (Xenos et al., 2019).

Strand	Representative formulation	Core representation
Normative decision framework	Wealth, Autonomy, Meaning	points in $(W, A, M)$-space
Concept recommender	diploma/job “concepts”	sequence of steps
Education–career integration	courses, jobs, skills	heterogeneous graph
Future-oriented guidance	2020/2050 missions	3D virtual world

This suggests that CareerScape is best understood not as a single tool but as a family of architectures for making career structure explicit: multi-objective decision spaces, sequence models, graph-based education–job linkage, and interactive exploratory environments.

2. Data infrastructures and trajectory representation

The empirical substrate of CareerScape research is the large-scale normalization of resumes and professional profiles. JobHop defines an unstructured resume set $D^u$, a structured resume $d^s=(E,Q)$ with work experiences $E$ and qualifications $Q$, and a normalized resume $d^n=(E^n,Q)$ in which job titles are mapped to ESCO occupation codes. Its released dataset contains over 2.3 million work experiences extracted from and grouped into more than 391,000 resumes, after filtering 442,555 resumes from 385,052 individuals and retaining 391,194 usable resumes (Johary et al., 12 May 2025). In the same line of work, trajectory analysis often uses an explicit sequence model such as $\text{Profile}_u=(s_1,s_2,\dots,s_T)$, where each step is either an academic or job step annotated with concepts, fields, dates, and text (Nadjem et al., 2020).

Two additional infrastructures extend this logic. “Leveraging LLMs for Career Mobility Analysis” constructs Career229K, a set of 228,710 linear five-year career trajectories for college-educated U.S. workers, with each trajectory carrying company, location, refined 8-digit SOC occupation code, NAICS industry code, and state–year–SOC wage information (Achananuparp et al., 15 Nov 2025). “Career Mobility of Planning Alumni in the United States” uses LLMs to extract structured career histories from more than 130,000 LinkedIn profiles of planning alumni and then computes multisector experience, industry span, location moves, market diversity, organizational engagement, and skill aggregates at profile scale (Wang et al., 12 May 2026).

This infrastructure makes several forms of analysis possible. First, trajectories can be treated as transition systems: JobHop computes transition matrices over ESCO groups and observes a dark-red diagonal, indicating strong within-group persistence, with especially visible transitions among Managers, Professionals, and Technicians and Associate Professionals (Johary et al., 12 May 2025). Second, trajectories can be aligned to institutional taxonomies, such as ESCO or O*NET-SOC, allowing external wage tables, occupational hierarchies, and regional labor-market statistics to be merged into the same analytic object (Achananuparp et al., 15 Nov 2025). Third, the same representation can support both descriptive labor-market research and downstream recommendation.

3. Guidance, recommendation, and progression systems

At the simplest algorithmic level, CareerScape recommendation can be expressed as concept ranking. “Predicting Personalized Academic and Career Roads” scores candidate concepts with

$\tilde{S}(H \mid C)=S_HF(H,C),$

where $F(H,C)$ is the joint frequency of a target concept and a contextual concept, and shows that contextual concepts improve over the baseline when predicting current diploma or current job concepts (Nadjem et al., 2020). Using 7,500 users, 17,500 diploma steps, and 24,000 job steps, the paper reports, for current-job concept prediction, MRR values of 0.730 for the baseline, 0.763 using last diploma, 0.798 using previous job, and 0.800 using next job (Nadjem et al., 2020).

A richer progression architecture appears in “Steve: LLM Powered ChatBot for Career Progression.” Steve parses PDF resumes with PyPDF2 and tiktoken, uses GPT Function Calling to extract structured profiles, maps the candidate to a JSON career tree with next-positions and second-jump-positions, compares extracted skills against a skills.json file, assigns skill levels in $(W, A, M)$0, and retrieves courses through Qdrant using all-MiniLM-L6-v2 embeddings (Renji et al., 3 Apr 2025). The result is a modular pipeline from resume to career-stage inference, skill-gap identification, and course recommendation.

An explicitly integrated academic variant is the WBSA+CGE architecture for CS and SWE students. The Web-Based Student Assessment platform manages profiles, tasks, faculty mentoring, secure chat, and institutional workflows, while the Career Guidance Expert uses a Multilayer Perceptron trained on academic and extracurricular data to predict one of six master fields—AI, DS, DEV, SEC, SDE, and UI/UX—with a reported validation accuracy of 94.71% (Faruque et al., 14 Jun 2026). The interface then maps the top probabilities to primary and secondary goals and to curated roles, research domains, and master’s programs (Faruque et al., 14 Jun 2026).

CareerScape guidance is not limited to dashboards. The Future Time Traveller environment organizes orientation through an Information Centre, a World of 2020, a World of 2050, and a return-to-2020 reflection area, combining web quests, escape rooms, treasure hunts, and reflective outputs such as “Message to humanity” and “Message to myself in the future” (Xenos et al., 2019). This suggests that scenario-based immersion and recommender-style inference are complementary rather than competing guidance paradigms.

4. Graph, neural, and LLM architectures

One paper explicitly names “CareerScape” as a detection framework rather than a guidance system. “Unmasking Fake Careers” constructs a global heterogeneous career graph from genuine resumes, with node types for job titles, companies, and standardized job descriptions, and relation types for title–title transitions, company–company transitions, title–company worked_at, and title–description has_description. Description edges are created when $(W, A, M)$1 with $(W, A, M)$2, user subgraphs are augmented with trusted two-hop neighbors from the global graph, and a heterogeneous GNN with relation-specific message passing, duration-aware edges, and a self-attention subgraph encoder is trained with binary cross-entropy to classify resumes as genuine or machine-generated (Yamashita et al., 24 Sep 2025). On the combined dataset, CareerScape reaches an F1-score of 0.86, versus 0.80 for GraphSAGE and R-GNN, 0.62 for LightGBM, 0.54 for GPTZero, and 0.49 for DetectGPT, with reported relative improvements of 5.8–85.0% across baselines and generator regimes (Yamashita et al., 24 Sep 2025).

LLM pipelines are also central upstream of these graphs. JobHop uses one-shot prompting with Gemma2-9b to extract work experiences and qualifications into JSON, achieving JSON accuracy 99.5%, title accuracy 81.6%, and overall accuracy 82.1% on a 200-resume hand-annotated extraction set; exact ESCO mapping then uses a commercial Nobl.ai classifier with 72.6% ESCO code accuracy, 77.8% ESCO level 4 accuracy, and 84.7% ESCO group accuracy on a 600-experience mapping set (Johary et al., 12 May 2025). In U.S. mobility analysis, FewSOC uses few-shot GPT-3.5 Turbo plus taxonomy-constrained mapping to assign 8-digit O*NET-SOC codes, achieving precision 0.72 versus 0.65 for the original LC-SOC labels, with a statistically significant 10.42 percentage-point advantage in crowdsourced evaluation (Achananuparp et al., 15 Nov 2025).

These architectures show two distinct but convergent uses of structure. In guidance systems, structured profiles are needed to compare a person to role templates, skill taxonomies, or historical trajectories. In fraud detection, the same structure exposes unlikely role jumps, anomalous company transitions, or repetitive duration patterns that disappear when careers are reduced to text strings. A plausible implication is that a mature CareerScape stack requires both reliable extraction and graph-aware inference.

5. Mobility, progression, and empirical determinants

CareerScape research increasingly analyzes not only where people can go, but which transition patterns are associated with moving up. In the early-career U.S. study of college graduates, upward mobility is defined by whether the occupational wage in year 5 exceeds the occupational wage in year 1. Logistic models show that intra-firm occupation change has the strongest association with upward mobility, followed by inter-firm occupation change and inter-firm lateral move, with coefficients of 0.9506, 0.9338, and 0.7965, respectively; women and Black graduates have lower overall odds of upward mobility, and women’s returns to job changes are lower across all move types (Achananuparp et al., 15 Nov 2025).

Planning alumni exhibit a different but related pattern. Negative Binomial and Cox models show that multisector experience, industry span, larger professional networks, and greater organizational engagement are associated with more frequent and faster upward transitions. In the main model, Multisector has IRR 1.10 and HR 1.21, Industry span has IRR 1.06 and HR 1.03, Connections has IRR 1.0009 and HR 1.0007, and Organizations has IRR 1.04 and HR 1.08 (Wang et al., 12 May 2026). Technical skills act as early-career velocity signals, soft skills become increasingly decisive in later stages, and AI-related skills present limited additional advantage, with AI skill IRR reported at 0.92 (Wang et al., 12 May 2026).

Resume-scale seniority analysis sharpens the micro-level determinants. Over more than half a million resumes, logistic models with $(W, A, M)$3 regularization show that total job time is the most important positive factor in predicting whether the most recent job is senior, while largest employment gap is modestly negative and number of jobs is often near zero. In Manufacturing, for example, total job time has coefficient 0.361, largest gap size has coefficient -0.131, and buzzwords have coefficient 0.291; personal pronouns are consistently negative across many sectors, whereas counts of skills, awards, certifications, publications, and patents are generally small (Wright et al., 2021). This strongly supports the claim that previous experience outweighs other aspects of human capital in seniority attainment (Wright et al., 2021).

The transition from education to work adds another layer. Among 30 recent CS graduates, 28 had completed a capstone project, only 2 had done internships, less than fifty percent had a mentor in their first role, and over two-thirds had no substantial professional development opportunities (Whalley et al., 2024). Yet those with mentors overwhelmingly reported them as extremely helpful, and those with ongoing professional development reported significant gains including growth of leadership skills and accelerated career progression (Whalley et al., 2024). In contingent labor, AMT workers similarly demonstrate that career progression depends not only on skills but on whether those skills can be translated into recognized professional capital: 60 of 98 respondents said AMT contributes to their career goals, but only 19 listed it on a résumé or CV (Kasunic et al., 2019).

Together, these results position CareerScape as an empirical framework for analyzing internal promotion, external mobility, sectoral breadth, network accumulation, and the legibility of experience itself.

6. Accessibility, inclusion, and governance

A comprehensive CareerScape must also account for uneven access to information, work, and representation. The one stop career centre for people with disabilities in Malaysia frames the problem as one of fragmentation, prejudice, low employer confidence, and built-environment barriers, and proposes a web-based “no-wrong door” system in which disabled school graduates, teachers, school counsellors, parents, and caregivers can enter qualifications and skills, search jobs, and access employment information. The paper explicitly grounds the design in universally accessible services and reasonable accommodations and modifications (Nasution et al., 2020). This work treats CareerScape not only as analytics but as service integration.

SkillTrade extends that inclusion logic to peer learning and startup hiring. It is free for individual users, supported by donations, and charges startups a small fee only when they successfully hire. It supports Skill Exchange Users, Learn-Only Users, Startup Users, and Developers/Admin Users; uses React.js, Tailwind CSS, Node.js, Express.js, MongoDB, Socket.IO, and Google Meet/Zoom integration; and reports 50+ registered users, 70% active engagement in skill exchange or job applications, 87% successful skill exchanges, and 85% startup satisfaction with the hiring process (Purushotham et al., 21 Jan 2025). The platform’s explicit target groups include housewives, people with disabilities, and financially constrained learners (Purushotham et al., 21 Jan 2025).

Governance issues follow from the same architectures that make CareerScape powerful. Steve notes dependence on manually curated trajectories and role-specific skill files, as well as subjective proficiency estimation from resumes and Q&A rather than objective testing (Renji et al., 3 Apr 2025). JobHop is limited by resume anonymization, ESCO mapping error, and the regional bias of the VDAB population (Johary et al., 12 May 2025). The planning alumni study notes that LinkedIn underrepresents alumni without online profiles and omits organizational-level factors (Wang et al., 12 May 2026). The fake-career CareerScape warns against employee surveillance and discriminatory candidate filtering, emphasizing responsible deployment for integrity rather than punitive automation (Yamashita et al., 24 Sep 2025).

A plausible implication is that CareerScape systems should be judged on three linked criteria: representational fidelity, institutional accessibility, and governance discipline. The research record already contains the building blocks for all three, but it also shows that no single architecture resolves the tensions among prediction, explanation, fairness, and intervention.