When Occupation-derived Skills Don't Help: On the Limits of Skill-Augmented Career Path Prediction

Ege Süalp; Elena Senger; Toine Bogers; Barbara Plank

cs.CLOct 14, 2026

When Occupation-derived Skills Don't Help: On the Limits of Skill-Augmented Career Path Prediction

Authors: Ege Süalp, Elena Senger, Toine Bogers, Barbara Plank

Organizations: MaiNLP, Center for Information and Language Processing, LMU Munich, Germany · Fraunhofer Institute for Systems and Innovation Research ISI, Germany · NERDS, IT University Copenhagen, Denmark

Abstract

Skills are widely assumed to play a central role in career transitions, motivating their use in career path prediction (CPP) systems. In this work, we critically examine whether explicit skill representations provide complementary signal beyond text-based job representations. We evaluate multiple strategies for incorporating skills into CPP models, including score fusion and neural architectures, under two conditions: predicted skills and ground-truth ESCO annotations. Surprisingly, across two benchmark datasets, none of these settings yield substantial improvements over strong text-only baselines. Even when using ground-truth skills, gains remain marginal and inconsistent. Through detailed analysis, we identify key factors underlying this result: high redundancy between textual and skill representations, limited discriminative power of ESCO's taxonomy, and representational collapse in skill embeddings. Furthermore, we find that skill prediction difficulty correlates with overall sample difficulty, limiting the potential for complementary signals. Our findings challenge the common assumption that explicit skill modeling inherently benefits career prediction and suggest that dense text representations already capture most skill-relevant information.

Figures & tables

		Augmented			Clean
Dataset	Configuration	MRR	R@5	R@10	MRR	R@5	R@10
DECORTE	Text-only baseline	0.266	0.356	0.420	0.276	0.374	0.414
	Skill Pooling	0.270	0.356	0.426	0.283	0.396	0.427
	Concatenated Encoding	0.262	0.342	0.425	0.278	0.392	0.449
KARRIEREWEGE+	Text-only baseline	0.439	0.527	0.604	0.334	0.425	0.502
	Skill Pooling	0.438	0.525	0.602	0.334	0.421	0.502

Table 1: Skill-enhanced MLP results, by skill integration method. For each method, the best-performing configuration is reported.

Figure 1: Discriminative power analysis via score margins (

M=s_{top-1}-s_{top-2}

). Distributions show the density of correct predictions (green) versus incorrect ones (red) for Text, Skill, and Hybrid models. A higher margin for correct predictions indicates superior calibration and decision certainty. The sharper separation in the Text-only model signifies higher discriminative power, whereas the overlap in the Skill-only model suggests lower certainty in distinguishing between occupational classes.

Figure 2: Rank dynamics between text-only and hybrid models. (A) Frequency of improved vs. degraded samples by initial rank bucket. (B) Distribution of rank displacement magnitude. Degradations concentrate in high-rank buckets (1–10) but are individually small; improvements occur mainly in low-rank buckets and involve larger shifts that are metric-insensitive.

Embedding	Top 10	Top 20	Top 50
Text	38.6%	53.7%	73.7%
Pooled Skills (pjmath-BGE)
Mean pooling	80.0%	91.6%	98.1%
Weighted IDF	77.3%	90.0%	97.7%
Log pooling	76.2%	89.1%	97.3%
Concatenated Skills (MPNet-Decorte)

Table 2: PCA cumulative explained variance comparison between text and skill embeddings.

Figure 3: Pairwise cosine similarity and PCA projections for text vs. pooled skill embeddings. Pooled skill embeddings ( $\mu\approx 0.90$ ) shift dramatically toward near-identical representations compared to text ( $\mu=0.07$ ); PCA projections show text embeddings spanning the latent space broadly while skill embeddings collapse into constrained, structured clusters.

Figure 4: Pairwise cosine similarity and PCA projections for text vs. concatenated skill embeddings. Even after mitigating collapse via IDF-weighted concatenation ( $\mu=0.66$ ), the skill modality remains more concentrated than text ( $\mu=0.07$ ), lacking the same latent diversity.

Quartile	TEXT Acc@1	SKILL Acc@1	HYBRID Acc@1
Q1 (Specific)	11.75%	10.20%	12.20%
Q2	16.89%	11.11%	16.44%
Q3	20.89%	14.89%	19.33%
Q4 (Generic)	24.83%	23.50%	23.95%

Table 3: Performance by skill genericness quartile (DECORTE-based IDF). Higher genericness correlates with better performance across all models, reflecting dataset composition rather than skill utility.

Trajectory Type	Samples	MRR T	$\Delta$ MRR
Lateral/Same-Level	1,291	0.298	$-0.003$
Progression (to Manager)	263	0.216	$-0.004$
Regression (from Manager)	248	0.150	$+0.020$

Table 4: Skill enhancement effect by career trajectory type. Progression = target is Manager but last job is not; Regression = last job is Manager but target is not.

Category	Count	MRR T	$\Delta$ MRR
Professionals	617	0.309	$+0.012$
Managers	468	0.323	$-0.014$
Technicians	351	0.206	$+0.004$
Clerical Support	159	0.195	$+0.005$
Service & Sales	137	0.194	$-0.028$
Elementary	13	0.119	$+0.059$

Table 5: Skill enhancement effect by target ISCO-1 group (DECORTE test set). The four smallest ISCO-1 categories (Armed Forces, Agriculture, Craft & Trades, Machine Operators; 57 samples combined) are omitted due to insufficient sample size for reliable estimates.

STEP: Career-Path Recommendation via Temporal and Educational Trajectory Modeling
Jul 13, 2026Iman Johary, Guillaume Bied, Alexandru C. Mara +1Sequence Modeling
AIDA-IDLab, Department of Electronics and Information Systems, Ghent University, Ghent, Belgium
UniSkill: A Dataset for Matching University Curricula to Professional Competencies
Mar 3, 2026Nurlan Musazade, Joszef Mezei, Mike ZhangEducational Data Mining
Åbo Akademi University, Finland · University of Copenhagen, Denmark · Pioneer Centre for AI, Denmark
Disentangled Skill Representations for Predictive Human Modeling
Aug 24, 2026Mariah Schrum, Deepak Gopinath, Srijan Srivatsa +2Human Behavior Prediction Disentangled Representation Learning
Toyota Research Institute Los Altos, California, USA

Jul 13, 2026J/K move · Enter open · S save

STEP: Career-Path Recommendation via Temporal and Educational Trajectory Modeling

Iman Johary, Guillaume Bied, Alexandru C. Mara, Tijl De Bie

AIDA-IDLab, Department of Electronics and Information Systems, Ghent University, Ghent, Belgium

Sequence Modeling

Open paper AskPDF arXiv

Career paths encode decades of skill acquisition, role transitions, and educational investment, and understanding them at scale underpins workforce planning, labor market policy, and job recommendation. Resumes are a rich source of information about career paths: they contain detailed descriptions of work experience, education, and skills. Yet their unstructured, heterogeneous, and multilingual nature has long prevented large-scale systematic analysis. With the advent of large language models (LLMs), it is now possible to source rich career trajectory data containing temporal and educational signals from unstructured resumes, enabling new opportunities for career-path recommendation. Exploiting this opportunity, we present STEP (Sequential Trajectory of Employment Prediction), a novel career-path recommendation system that leverages temporal and educational signals to predict the next job in a career trajectory. STEP integrates a time-decay Gated Recurrent Unit (GRU) cell to model temporal dynamics, Feature-wise Linear Modulation (FiLM) conditioned on educational attainment, and attention-based sequence pooling to select relevant features for next job prediction. To improve internal occupation representation for STEP, we introduce ROUTE, a two-stage contrastive procedure that first adapts a multilingual encoder to the career domain via unsupervised denoising autoencoding, then performs supervised contrastive fine-tuning with guided negative selection. We evaluate STEP on four datasets of career trajectories, including an improved version of our publicly available JobHop dataset, and show that it outperforms state-of-the-art baselines in next job prediction. The dataset and code are publicly released to support reproducible career-trajectory research.

Abstract

Explore similar work

STEP: Career-Path Recommendation via Temporal and Educational Trajectory Modeling

UniSkill: A Dataset for Matching University Curricula to Professional Competencies

Disentangled Skill Representations for Predictive Human Modeling