Dual-Branch Mutual Learning Framework for Human Resources Recommendation

With the rapid expansion of the Internet, online recruitment services are transforming traditional hiring practices. Human resource recommendation systems aim to reduce information overload by connecting job seekers with relevant opportunities. However, the evolving job market poses challenges such as cold-start and scalability. To address these challenges, we propose a Dual-Branch Mutual Learning (DBML) framework specifically designed for job recommendation in human resource management. At its core, DBML enhances a Bidirectional Long Short-Term Memory (BiLSTM) network by introducing a residual structure, resulting in an R-BiLSTM equipped with forward and backward residual gates. These R-BiLSTM units are deployed in parallel to encode unstructured textual features from job descriptions and resumes. The encoded representations are then processed through depthwise separable convolutions for efficient feature extraction, followed by a soft attention mechanism that captures fine-grained interactions between the text inputs. For cold-start scenarios, DBML integrates pre-trained GloVe embeddings into both job and resume encoders, allowing the model to generalize semantic similarities even for unseen entities. Experiments on large-scale datasets demonstrate that the proposed DBML achieves state-of-the-art results in accuracy, AUC, and F1 score for job-resume text matching. Moreover, DBML significantly improves scalability, reducing training time by 28.6% and parameter usage by 21.4% compared to a standard baseline model. In cold-start tests, DBML outperforms existing methods, achieving a 7.35% improvement in F1 score when applied to unseen job postings.