基於時間感知神經網絡預測工作轉換

Abstract

人才流動預測在人力資源管理領域中具有重要意義，對企業而言，預測候選 人的下一份工作可作為人才招募決策的依據;對獵頭而言，預測潛在目標人才可 提高招募效率;對員工而言，收集其他合適的公司或職位信息可簡化尋找工作的 過程。然而，現有的模型忽略了人員、公司和職位隨著時間變化的事實，包括職 員技能、公司大小、公司服務或是職位實際內容等的動態變化。這些變化可能導 致對於每個人在同一工作於不同時間下的經驗其實具有不同的意義，使用固定的 公司與職位嵌入反而可能會影響個人在預測下一份工作的預測效能。 因此，我們擴展了目前預測下一份工作任務中最先進的模型，引入時間維度 的概念，通過考慮時間序列嵌入，加權過往的嵌入去更好地捕捉當前時間窗口下 公司和職位的特徵，了解到個人的職位軌跡中每份工作於不同時間可能代表的意 義，從而提高下一份工作預測的準確性。除了評估下一份工作預測的準確性的同 時，我們也探索了不同的時間窗口大小以及調整向過往時間加權的滯後數量，觀 察不同的時間設置對預測準確性的影響。實驗結果表明，我們提出的方法能夠有 效地捕捉在不同時間上公司和職位的表示向量(嵌入)，進而提升預測下一份工作 的模型性能，整體上優於當前表現最佳的方法。

Predicting job mobility is of great significance in the field of human resources management. For businesses, forecasting a candidate’s next job can serve as a basis for talent recruitment decisions. For recruiters, predicting potential candidates enhances their recruitment efficiency. Employees also can benefit from job mobility prediction by pointing the employees to predicted companies and positions to move, streamlining the employees’ job search process. However, existing models overlook the temporal changes in individuals, companies, and positions, such as skills, company size, services, and job content. These changes may lead to different meaning for individuals staying in the same job at different times. The previous models use fixed company and job embeddings may degrade the effectiveness of predicting the individual’s next job. To address this, we extend the state-of-the-art model for predicting the next job by expanding the time dimension. By considering time-series embeddings and weighting previous representations, we better capture the characteristics of companies and positions at each time window. This enables us to understand the varying meanings of each job in an individual’s career trajectory at a specific time window, thereby improving the effectiveness of predicting the next job. Additionally, we explore different time window sizes and adjust the number of lags for weighting to observe their impact on prediction effectiveness. Experimental results demonstrate the ability of our proposed approach in capturing representations of companies and positions at different time windows, leading to enhanced effectiveness of next job prediction, outperforming that of the existing state- of-the-art methods.