GAT-MDN predicts salary distributions using graph attention networks

Traditional salary prediction models output a single point estimate and treat job attributes like location, occupation, and industry as independent categorical features—ignoring both the inherent multi-modality of compensation data and the rich hierarchical relationships that govern pay norms. To address this, Zhipei Qin and colleagues introduce GAT-MDN, a unified framework that simultaneously tackles uncertainty and relational structure. For each attribute domain, GAT-MDN constructs a graph with edges encoding hierarchical parent-child containment and weighted similarity links derived from a Sentence-Transformer. Parallel Graph Attention Networks then learn context-sensitive node representations from these multi-relational graphs. A priority-based hierarchical selection module assembles a composite feature vector that gracefully handles missing or coarse attributes, and a Mixture Density Network head maps this vector to the parameters of a Gaussian Mixture Model, yielding a full conditional salary distribution.

Extensive experiments on a real-world Dutch job-posting dataset of over 1 million records demonstrate that GAT-MDN significantly outperforms a non-graph MLP-MDN baseline in both Negative Log-Likelihood (NLL) and Mean Squared Error (MSE). The paper shows that incorporating graph-structured attribute relationships and probabilistic modeling leads to more accurate and interpretable salary predictions. This approach can help job seekers and employers better understand compensation ranges, support fairer negotiations, and reduce information asymmetry in labor markets.