基於公司感知的遞迴式神經網路方法的人才流動預測

Abstract

人才流動預測是協助企業人力資源團隊制定有效的人才管理策略的關鍵任務。此問題可以透過線上專業網絡的歷史工作轉換資料來分析與預測企業間未來的人才流動。過去的研究方法使用成對的公司資料結構進行特徵工程和預測。然而，成對的公司資料結構只能提供整體人才流動網絡的聚合特徵，並不能利用從焦點公司到其他公司的分佈。此外，先前的研究也使用股票數據作為額外的數據來源，但股票數據容易被眾多市場訊號影響，並不一定能反應一間公司對人才的吸引力。為了解決這些限制，我們提出了一個深度學習模型，採用了公司列式的資料結構和公司評價網站的評分資料作為特徵，並使用遞迴式神經網路以擷取人才流動時間序列的特徵。此外，我們模型加入了公司感知結構和嵌入層去有效地捕捉了各焦點公司的人才流動模式。與現有模型相比，不同職位的預測表現提高了 3-4%。

This thesis addresses the problem of talent flow prediction by leveraging historical job transition data from Online Professional Networks to forecast future talent movements, a crucial task for human resource teams in developing effective talent management strategies. Previous approaches used a pair-wise structure for feature engineering and prediction. However, the pair-wise structure can only provide aggregated features of the network and does not leverage the distribution from a focal company to other companies. Additionally, previous studies have used stock data as an additional data source, yet stock data can be sensitive to many market signals.To address these limitations, we proposed a deep learning model employs a company list-wise structure and company rating data as features and feed into a RNN-based model to learn the time series features. After that, our model's company-aware structure and embedding layer effectively capture each focal company’s unique talent flow patterns. It demonstrates a 3-4% improvement in predictive performance over existing models across various positions.