Job2Skills: Skill Ranking for Job Titles

• Job2Skills is a system that maps job titles to a prioritized list of relevant skills using machine learning and NLP methods.
• It leverages weak supervision with neighbor-voting, transformer-based encoders, and IDF weighting to optimize skill ranking.
• The approach supports applications in recruitment, labor market analytics, and cross-lingual job recommendations by generalizing to novel job titles.

A Job2Skills system addresses the computational problem: given a job title, produce a ranked list of skills such that the most important or relevant skills are prioritized for that role. This mapping problem underpins applications in recruitment, labor market analytics, upskilling pathways, and job recommendation engines. The current state of the field employs a spectrum of machine learning, natural language processing, and network science methods, ranging from transformer-based neural models to co-occurrence-driven graph clustering and market-aware statistical ranking.

1. Formal Problem Definition and Objectives

The core Job2Skills task is: for a given job title $T$ (e.g., “Python Developer”), output a real-valued importance score $r(T, s) \in [0, 1]$ for each skill $s$ in a fixed taxonomy $S = \{S_1, S_2, \dots, S_{|S|}\}$, thereby inducing a ranked list by descending $r$ values. The scale of underlying resources varies—with taxonomies reaching $|S|\approx 37,000$ and input job title corpora at $N\approx 170,000$ unique entries (Anand et al., 2022). The Job2Skills output must generalize to novel, possibly multilingual titles and surface both ubiquitous and niche competency requirements in appropriate order.

2. Labeling Strategies: Weak Supervision via Job Title Neighborhoods

The principal challenge in Job2Skills construction is label scarcity. Annotating the relative importance of thousands of skills per job title is infeasible. Anand et al. (Anand et al., 2022) design a weak supervision protocol as follows:

• Embed each title $T$ using Sentence-BERT into $v_T \in \mathbb{R}^{768}$.
• Compute cosine similarity to all other titles, retaining a “neighbor” set $\mathcal{N}(T)$ with cosine $\geq 0.75$.
• Accumulate a multiset of all skills assigned to $\mathcal{N}(T)$.
• For each skill $s$, compute the raw frequency $c(T,s)$, then normalize:

$y(T, s) = \frac{c(T,s)}{\max_{s'} c(T,s')}$

yielding a weakly supervised, continuous “importance” target $y(T,s) \in [0,1]$.

This neighbor-voting method leverages the redundancy in similar job titles and their skill tags to estimate skill relevance, reducing noise compared to direct bag-of-words matching.

3. Model Architectures and Learning Paradigms

Encoder-based Regression

The central architecture in (Anand et al., 2022) consists of a frozen or fine-tuned Language-agnostic BERT Sentence Encoder (LaBSE) which outputs $v_T$; a massive fully-connected layer with learnable weights $\{w_s, b_s\}$ for each skill $s \in S$ scores every skill:

$$u_s = w_s^{\top} v_T + b_s \ , \qquad \hat{y}_s = \sigma(u_s)$$

Skills are sorted by descending $\hat{y}_s$ at inference. Training minimizes mean squared error between predictions and pseudo-labels $y(T,s)$:

$$L(\theta) = \frac{1}{|B|} \sum_{(T, s) \in B} (\sigma(w_s^{\top}v_T + b_s) - y(T,s))^2$$

where $\theta$ may comprise only linear layer weights (frozen LaBSE) or the full encoder stack (fine-tuned end-to-end).

Inverse Document Frequency (IDF) Specialization

To mitigate the dominance of common skills (e.g., “Excel”, “Communication Skills”), an IDF multiplier is introduced at inference:

$\mathrm{IDF}_s = \log\left(\frac{N}{f_s}\right)$

where $f_s$ is the number of titles in which $s$ appears. The final score:

$$\mathrm{score}_{\mathrm{final}}(T,s) = \sigma(w_s^{\top} v_T + b_s) \cdot \mathrm{IDF}_s$$

This enhances the prominence of specialized or rare skills in the output ranking.

4. Evaluation Protocols and Empirical Results

Systems are evaluated by comparing predicted skill rankings against the pseudo-ground-truth $y(T,s)$ for held-out test job titles. Metrics include:

• Average Precision at 20 (AP@20):

$$\mathrm{AP@20}(T) = \frac{1}{\min(K, 20)} \sum_{i=1}^{20} P(i) \cdot \mathrm{rel}(i)$$

where $P(i)$ is precision at cut-off $i$, $\mathrm{rel}(i) = 1$ if the $i$th ranked skill's $y(T,s)>0$, $K$ is the number of positive skills (Anand et al., 2022).

• Mean AP@20 across all test titles.

Quantitative results:

Model Configuration	Mean AP@20
LaBSE (frozen encoder)	0.621
LaBSE + Linear (end-to-end FT)	0.722

IDF application qualitatively shifts rankings: for “Stock Broker”, generic skills like Sales and Marketing are demoted in favor of domain-specific skills such as Financial Markets and Equities.

5. Practical Considerations, Multilinguality, and Limitations

The Job2Skills approach is designed to scale to large taxonomies via vectorized operations: one LaBSE embedding per title, followed by a $|S| \times d$ matrix multiplication, pointwise sigmoid, and IDF scaling. This is highly parallelizable on GPU hardware.

Cross-lingual generalization is feasible when using a frozen, multilingual encoder: the system can embed titles in 109 languages, e.g., ranking skills for “Arzt” (German), “Docteur” (French), etc. However, fine-tuning for higher accuracy in English degrades alignment for other languages.

The authors recognize key limitations:

• Only job title text is used; full job descriptions—which offer richer context—are not exploited.
• The weak-supervision protocol may miss rare-but-critical skills excluded from the neighborhood pseudo-labels.

6. Representative Outputs and Interpretability

Selective examples illustrate the interpretability and face-validity of Job2Skills outputs:

• Front-End Developer: JavaScript, Software Development, Java, CSS, AngularJS.
• Stock Broker (with IDF): Financial Markets, Securities, Equities, Proprietary Trading, Commodities.
• Therapist: Therapy, Healthcare, Patient Care, Nursing, Group Therapy.

Applying IDF-based boosting generates outputs that more closely align with expert intuitions about the distinguishing skills for specialized job roles.

---

All methodology, evaluation, and ranking outputs outlined above are components of the Job2Skills system as formulated in Anand et al.'s "Is it Required? Ranking the Skills Required for a Job-Title" (Anand et al., 2022).