以多任務基於注意力的方法預測工作轉換

Abstract

在就業市場中，有各式各樣的工作機會。一個人的下一份工作有許多選擇，同時可能涉及他曾經任職過的多間公司和不同職位，這使得個人工作轉換的預測變得更加複雜。同一家公司在不同時間點的定位可能會發生變化，同一職位在不同時間點的性質和所需技能也可能有所不同。在個人的工作經歷中，各個工作發生的時間點不同，工作之間的相關程度也有所不同，這使得用模型來捕捉這些特性變得更加困難。 為了解決這些問題，我們延伸了目前考量公司、職位在時間動態變化的最先進模型，使用注意力機制並引入多任務學習來增進模型的表現。透過注意力機制，可以權衡當前工作與過去工作的重要性；透過多任務學習，可以利用輔助任務，參考額外的相關資訊，進而更新模型共用的參數提升表現。 實驗結果顯示，我們的模型優於目前最先進的方法，且注意力機制、多任務的引進這兩項改動都分別對這項任務有所改善。

In the job market, there are various job opportunities available. An individual's next job can have many choices, potentially involving multiple companies and different positions they have previously held, making the prediction of job transitions more complex. The positioning of the same company may change at different points in time, and the nature and required skills of the same position may also vary over time. In an individual's work history, the timing of each job differs, and the degree of relevance between jobs also varies, making it more challenging to use models to capture these characteristics. To address these issues, we extend the current state-of-the-art model that considers the dynamic changes of companies and positions over time by incorporating attention mechanisms and introducing multitask learning to enhance the model's performance. Through the attention mechanism, the importance of the current job and past jobs can be captured; through multitask learning, additional related information can be referenced using auxiliary tasks, thereby updating the shared parameters of the model to improve performance. The experimental results demonstrate that our model outperforms the most advanced methods now in use, and the incorporation of the attention mechanism and multitask learning both separately contribute to the improvement of this task.