💬 NLP

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

This paper introduces STEP, a novel career-path recommendation system that leverages large language models to extract temporal and educational signals from unstructured resumes, utilizing a time-decay GRU, FiLM conditioning, and a specialized contrastive learning framework called ROUTE to outperform state-of-the-art baselines in predicting next job transitions.

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

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your career isn't just a straight line on a map, but a winding river that changes speed, depth, and direction based on how long you've been paddling and how much training you've packed in your boat. For a long time, trying to predict where that river flows next was like trying to guess the weather by looking at a single cloud. But a new system called STEP (Sequential Trajectory of Employment Prediction) is like a super-smart, time-traveling navigator that looks at your entire journey to tell you what the next bend in the river might look like.

Here's how it works, why it's different, and what the authors actually found.

The Magic Map: Turning Chaos into Clarity

First, the team had to deal with a messy problem: resumes. They are like a giant pile of handwritten notes in dozens of different languages, with job titles that mean different things to different people. You can't just feed a pile of messy notes into a computer and expect it to understand.

To fix this, they built a tool called ROUTE. Think of ROUTE as a master translator and librarian. It takes those messy, unstructured resumes and turns every job title and description into a precise, standardized "coordinate" on a giant map of 3,007 different jobs (using a system called ESCO). It does this in two steps: first, it reads millions of career stories to learn the "vibe" of the job market (like a student reading a library of books), and then it practices finding the right connections between jobs, making sure it doesn't get confused by jobs that sound similar but are actually different.

The Navigator: How STEP Predicts the Future

Once the map is ready, STEP takes over. It's not just guessing; it's a sophisticated engine with four special gears:

The Time-Decay Gear: Careers aren't like a video game where every level takes exactly the same time. You might stay in a job for six months or twenty years. STEP uses a special "time-decay" mechanism. Imagine your memory of a specific skill fading like a radio signal the longer you go without using it. If you haven't worked in a specific field for a decade, STEP knows that part of your "career signal" is quieter now. It weighs recent experiences heavier than ancient ones, but it remembers that some deep skills (like being a doctor) fade very slowly.
The Education Filter: A person with a PhD might take a different path than someone with a high school diploma, even if they have the same work history. STEP uses a "FiLM" (Feature-wise Linear Modulation) filter. Think of this like putting on a pair of glasses that changes the color of the world based on your education level. It adjusts the prediction to fit the specific "lens" of the person's degree.
The Attention Spotlight: When you look at a long career history, you don't remember every single detail equally. STEP uses an "attention" mechanism to spotlight the most important parts of your past. It asks, "Which of these past jobs actually matters most for predicting the next one?"
The Temperature Knob: Finally, the system has a "temperature" setting. This controls how confident it is. If the temperature is low, it's very picky and only suggests the most obvious next steps. If it's higher, it's more open to wilder possibilities. The authors found that tuning this knob was the single most important thing for getting good results.

What They Actually Found (The Proof)

The authors tested STEP on four different datasets, ranging from small collections of about 2,000 career histories to massive sets with over 500,000 paths. They compared it against other methods, including simple math models, standard neural networks, and even some large language models (LLMs).

The Winner: STEP came out on top in every single test. On the largest dataset (JobHop v2), it correctly predicted the next job in the top 10 suggestions 41.1% of the time. On another dataset (Karrierewege), it hit 70.3%.
What It Beat: It beat the previous best methods, including complex models that tried to just "read" the text without understanding the sequence of time and education.
What It Ruled Out: The authors explicitly tested using powerful, general-purpose Large Language Models (LLMs) to do the prediction directly. They found that, in this specific task of picking the next job from a fixed list of 3,007 options, these giant LLMs actually underperformed compared to their specialized, smaller model. The paper suggests that trying to generate the next word in a sentence isn't the same as retrieving the right job from a massive database; a specialized "retrieval" engine works better here.

The "Exploration" Gap

There's a twist in the story. While STEP is great at predicting what people actually do next, it tends to be a bit conservative.

The Reality: When people actually change jobs, they often make big leaps, moving to completely different industries (a "novelty" score of about 2.9 on a scale where 4 is a totally different world).
The Prediction: STEP suggests moves that are closer to home (a novelty score of about 1.5 to 1.7).
The Takeaway: The system is excellent at telling you the most likely next step, but it might miss the "wild card" opportunities that people sometimes take. The authors suggest that if you want a career counselor that encourages big, risky jumps, you'd need to tweak the system to be less conservative.

How Far Can It See?

The authors also tested what happens if you ask STEP to predict not just the next job, but the job after that, and the one after that (up to 5 steps into the future).

The Result: The accuracy naturally drops the further out you look, but it doesn't crash. Even 5 steps ahead, STEP still had a 28.7% chance of being in the top 10 suggestions for the JobHop v2 dataset. This suggests it's a reliable tool for long-term planning, not just immediate job hunting.

The Bottom Line

The paper proves that by combining a clean, standardized map of jobs (ROUTE) with a model that understands time, education, and attention (STEP), we can predict career paths better than ever before. It's not a magic crystal ball that solves all labor market problems, and it doesn't fix the historical biases in the data it was trained on. But as a tool for understanding how careers move, it's a significant step forward, showing that specialized, time-aware models are currently the best way to navigate the complex river of a working life.

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

1. Problem Formulation

The paper addresses the challenge of next job prediction within career trajectories. While resumes contain rich information regarding work history, education, and skills, their unstructured, heterogeneous, and multilingual nature has historically hindered large-scale systematic analysis. The authors frame career-path recommendation as a many-to-one sequence prediction task: given an observed sequence of jobs $(j_1, \dots, j_{t-1})$ and auxiliary signals (inter-job time intervals $\delta$ and education level $e$ ), the goal is to predict the next occupation $j_t$ from a finite candidate set (typically the 3,007 leaf-node occupations in the ESCO taxonomy).

The problem is complicated by three factors:

Data Scarcity: Existing public benchmarks are either small, lack temporal/educational metadata, or rely on proprietary data (e.g., LinkedIn) that precludes independent replication.
Representation: Standard sentence encoders often fail to capture the compositional semantics of career text or the specific relationships between job titles and the ESCO taxonomy.
Temporal and Educational Dynamics: Career transitions are not uniform in time (ranging from months to decades) and are heavily influenced by educational attainment, yet few models jointly integrate these signals in a retrieval-based sequential architecture.

2. Methodology

The proposed solution consists of two primary components: ROUTE (for representation learning) and STEP (for sequential prediction).

2.1 ROUTE: Representation of Unique career Trajectories as Embeddings

ROUTE is a two-stage contrastive learning procedure designed to produce high-quality occupational embeddings that serve as the input for the predictor.

Stage 1: Unsupervised Domain Adaptation (TSDAE): A pretrained multilingual encoder (specifically multilingual-e5-base) is adapted to the career domain using Transformer-based Sequential Denoising Autoencoding (TSDAE). The model reconstructs career text from token-deleted inputs, forcing the encoder to capture the full semantic content of heterogeneous resume descriptions.
Stage 2: Guided Contrastive Fine-Tuning (GISTEmbedLoss): The adapted encoder undergoes supervised contrastive fine-tuning on (history, next job) pairs. To address the issue of "false negatives" (where random batch sampling selects semantically similar occupations as negatives), the authors employ GISTEmbedLoss. This uses a frozen guide model to compute pairwise similarities, filtering out false negatives and emphasizing hard negatives to produce a cleaner gradient signal.

2.2 STEP: Sequential Trajectory of Employment Prediction

STEP is a sequential prediction model that operates on the embeddings generated by ROUTE. It integrates four specific architectural components to model career dynamics:

Time-Decay GRU Cell: Unlike standard RNNs that assume uniform time steps, STEP incorporates an exponential decay mechanism into the GRU hidden state update. Before processing a new job, the previous hidden state $h_{t-1}$ is decayed by a learnable factor $\exp(-\lambda \cdot \delta_t)$ , where $\delta_t$ is the time interval between jobs. This allows the model to learn that certain career attributes (e.g., domain expertise) decay slower than others (e.g., specific technical skills).
FiLM Degree Conditioning: Instead of concatenating education level as a simple input feature, STEP uses Feature-wise Linear Modulation (FiLM). The education level is encoded and projected into scaling ( $\gamma$ ) and shifting ( $\beta$ ) parameters that modulate the hidden state at every time step ( $h'_t = \tanh(\gamma) \odot h_t + \beta$ ). This allows the model to condition career progression patterns directly on educational attainment.
Attention-Based Sequence Pooling: Rather than relying solely on the final hidden state, the model uses a learnable attention mechanism to aggregate the sequence of degree-conditioned hidden states into a single trajectory representation. This weights career steps by their predictive relevance.
Learnable Temperature Scaling: The final trajectory embedding is compared against all candidate ESCO occupations via cosine similarity. A learnable temperature parameter $\tau$ scales the logits before the softmax, controlling the sharpness of the probability distribution over the large candidate set ( $K=3,007$ ).

The model is trained end-to-end using cross-entropy loss with label smoothing.

3. Key Contributions

The STEP Model: A novel sequential architecture that jointly models temporal intervals and educational attainment for next job prediction. It achieves state-of-the-art performance across four benchmarks.
The ROUTE Procedure: A two-stage embedding pipeline (TSDAE + GISTEmbedLoss) that significantly improves occupational representation quality, particularly for retrieval tasks, by adapting multilingual encoders to the career domain and mitigating false negatives in contrastive learning.
JobHop v2 Dataset: An improved version of the authors' previous JobHop dataset. It features higher-quality extraction via reasoning-controlled LLMs, a richer schema (including contract types and skills), and a five-level education taxonomy (None, Secondary, Bachelor, Master, PhD), replacing the previous binary flag.
Comprehensive Evaluation: The work evaluates on four distinct datasets (Decorte, Karrierewege, JobHop, and JobHop v2), demonstrating robustness across different scales, languages (English, Dutch, French, German), and annotation richness.

4. Results

The authors report performance using Mean Reciprocal Rank (MRR), Recall@5, and Recall@10.

Overall Performance: STEP outperforms all baselines (including Markov, Linear, MLP, BiLSTM, and direct retrieval) on all four datasets.
- On Decorte, STEP achieves 46.6% R@10, surpassing the previous best (43.01%) and the direct retrieval baseline (46.1%).
- On Karrierewege, STEP reaches 70.3% R@10, a 1.4 percentage point gain over the re-implemented BiLSTM baseline.
- On JobHop v2, STEP achieves 41.1% R@10, outperforming BiLSTM (38.2%) and MLP (38.5%).
Ablation Studies:
- Temperature Scaling: Removing this component causes catastrophic degradation (R@10 drops from 41.24% to 24.69% on JobHop v2), highlighting its critical role in calibrating probabilities over the large ESCO label space.
- FiLM vs. Concatenation: FiLM conditioning provides a marginal but consistent improvement over simple concatenation, confirming that multiplicative modulation better captures degree-career interactions.
- Embedding vs. Predictor: While ROUTE significantly improves direct retrieval performance, the STEP predictor itself is the primary driver of end-to-end gains. The predictor relearns the scoring function, making it robust to variations in the embedding stack, though the full ROUTE pipeline remains the safest default.
Diversity and Novelty: The analysis reveals a trade-off between accuracy and exploration. While STEP achieves high accuracy, it tends to recommend occupations closer to the current role (lower "Novelty" scores) compared to the ground truth transitions, which are often more exploratory. However, STEP offers a better balance of diversity and accuracy than pure retrieval baselines.
Multi-Step Prediction: In autoregressive multi-step evaluation (horizons up to 5), performance degrades gradually rather than catastrophically, retaining meaningful recall (e.g., 28.7% R@10 at horizon 5 on JobHop v2).

5. Significance and Claims

The paper claims that STEP represents a significant step forward in career-path recommendation by successfully leveraging the "richness" of unstructured resume data through Large Language Models (LLMs) for extraction and temporal/educational modeling.

Data Accessibility: By releasing JobHop v2 and the code, the authors address the critical lack of large-scale, publicly available, richly annotated career trajectory datasets, enabling reproducible research in this domain.
Methodological Integration: The work demonstrates that integrating real-valued temporal intervals and educational attainment into a retrieval-based sequential architecture yields superior results compared to standard sequential models (LSTM/MLP) or pure retrieval methods.
Domain Adaptation: The study validates that domain-adaptive pretraining (TSDAE) and guided contrastive learning (GISTEmbedLoss) are essential for creating occupational embeddings that generalize well, particularly when dealing with multilingual and noisy resume data.
Modesty on Limitations: The authors explicitly acknowledge that their models are trained on historical data and may reflect existing labor market biases (e.g., occupational segregation). They caution that high accuracy does not equate to optimal career counseling, as the models tend to be conservative and may under-recommend exploratory career moves. They emphasize that fairness evaluation is a necessary prerequisite for any real-world deployment.

In summary, the paper establishes a new state-of-the-art for next-job prediction by combining advanced representation learning with a temporally and educationally aware sequential model, while providing the community with the necessary data and tools to build upon this work.