The accurate valuation of job skills is crucial for companies to attract, retain, and manage talent effectively. In the knowledge economy, understanding skill value helps bridge the skill gap between employers and employees, fostering a competitive advantage in a rapidly changing technological landscape. At the individual level, skill valuation enables informed career planning and skill development. At the macro level, it reflects the economic equilibrium of the labor market. While traditional survey-based methods have limitations, the vast amount of data from online recruitment platforms offers a unique opportunity for data-driven skill analysis. Existing studies primarily focus on job skill demand; however, a robust quantitative method for assessing skill value based on its impact on salary remains elusive. This is a complex problem because skill value is context-dependent (varying across industries and companies) and skills are often interdependent. Furthermore, ground truth data on skill value is typically unavailable. This research addresses these challenges by proposing a market-oriented definition of skill value and a novel neural network model to estimate it.

Existing research demonstrates a positive correlation between job skill mastery and job salary. However, traditional market survey approaches lack the granularity and timeliness to capture the dynamic nature of skill value. Recent work utilizes online job postings to analyze skill demand, but quantitative assessments of skill value impacting salary are limited. The challenge lies in the context-dependent nature of skill value and the lack of ground truth data for training a reliable model. This research differentiates itself by offering a market-oriented definition of skill value that directly considers its impact on salary, addressing the limitations of previous approaches.

This study proposes a data-driven solution to quantify skill value by mining job advertisement data. The core concept is the Salary-Skill Value Composition Problem, where job salary is considered a composition of context-aware skill values. The proposed Salary-Skill Composition Network (SSCN) is a cooperative neural network with two key components: a Context-aware Skill Valuation Network (CSVN) and an Attentive Skill Domination Network (ASDN). CSVN dynamically models skills, extracts context-skill interactions, and estimates context-aware skill value ranges (lower and upper bounds). ASDN models skill domination, representing the relative importance of each skill in determining salary, using a graph-based approach incorporating skill co-appearance and context-skill interaction information. The salary prediction serves as a cooperative task, providing indirect supervision for skill value assessment. SSCN jointly learns skill value and salary prediction models, using job salary data as an indirect supervisory signal. The model uses a linear composition of skill values weighted by their domination to predict salary. The value is defined as the expected salary for a job only requiring that skill. This framework allows for the measurement of the influence of contexts on individual skills and the influence of skills on job salary. The data consisted of over 800,000 IT-related job postings from a Chinese online recruitment website, pre-processed to yield 215,308 samples. Several baseline models, including linear models (SVM, LR), GBDT, DNN, HSBMF, TextCNN, HAN, Transformer-XL, BERT, Roberta, and XLNet, were used for comparison in salary prediction.

The SSCN model demonstrated superior performance in salary prediction compared to benchmark models, exhibiting a 3.5% to 5.2% RMSE improvement over BERT, which was the best-performing baseline model. The analysis of skill value revealed several key insights: * **Level Influence:** Higher skill mastery levels generally correlated with higher skill value. The model differentiated the impact of various mastery levels, with "Versatile" skills contributing significantly more than "Can Read" skills. * **Temporal Dynamics:** Skill values exhibited fluctuations over time. For instance, Architecture showed a steady increase, whereas GoLang experienced sharp changes. * **Experience:** Longer working experience consistently led to higher skill values. The rate of increase varied across skills, suggesting different career trajectories for different skills. * **Company Influence:** Different companies valued skills differently. The analysis highlighted company preferences and the range of salary variation for specific skills. * **Skill Domination vs. Value:** The study found a trade-off between skill domination (how often a skill is required) and skill value. Generic skills had higher domination, while specific skills had higher values, suggesting a balance between breadth and depth of skill development. * **Skill Influence on Salary:** High-value, high-domination skills like Matrix Calculation had a substantial impact on job salary (18.4% decrease if removed). The study provided a quantitative assessment of this influence. The quantitative findings provided specific and substantial improvements over baselines and were generally aligned with real-world employment trends. The results were validated on an additional dataset of designer-related job postings, showcasing the model's generalizability.

The SSCN model successfully addresses the research question by providing a quantitative, market-oriented framework for skill value assessment. The results underscore the context-dependent nature of skill value and the interplay between skill value and domination in determining salary. The findings have significant implications for talent recruitment, business market analysis, student education, knowledge management, talent development, and job recommendation. The model’s accuracy in salary prediction validates its effectiveness in skill valuation.

This study introduces a novel approach to market-oriented job skill valuation using a cooperative composition neural network (SSCN). SSCN effectively estimates context-aware skill values without relying on labeled data, utilizing salary prediction as a cooperative task for indirect supervision. The superior performance of SSCN in salary prediction validates its ability to accurately assess skill value. Future research could explore incorporating additional data sources to enhance the model's comprehensiveness and generalizability, potentially including longitudinal data for long-term trend analysis. Further investigation into other domains beyond IT and design is also warranted.

The study's limitations include the use of data from a single online recruitment website and the limited time span covered by the data. This might introduce bias and hinder the analysis of long-term skill trends. The lack of ground truth for skill value necessitates indirect evaluation through salary prediction. Future work should address these limitations by incorporating more diverse and comprehensive data sources.