客服中心早已成為企業與客戶之間溝通的橋樑，客服員在兩者間更扮演重要的角色，企業在人才的培養與網羅都需要一套制定的評分標準，該如何避免人為因素影響客服員績效分數，並傳承這寶貴的知識，各家企業無不絞盡腦汁思考。 本研究以國內某客服中心各系統的歷史資料，做為研究數據的基礎，並參考了學者研究提出的23個客服中心營運相關的量化指標；實驗的過程歷經：資料各屬性與績效分數的相關係數分析、利用特徵值選取技術篩選特徵集、應用倒傳遞類神經網路預測客服員的績效分數。 經過特徵值篩選後的屬性集合，隱含著公司主管的評分要素，也是客服員績效評分的核心標準，由此可知主管在進行績效評比時，集合內的各屬性有相當大的機率被列為績效考核的重點；而預測模型的輸出值與實際客服員績效分數，其平均誤差值為2.78分，兩資料間的相關係數為0.9821，實驗結果說明所建置的網路預測模型輸出的預測分數相當接近實際的客服員績效分數，誤差值也比研究所設定的標準3分來的低，研究最後，再重新針對錯誤的資料進行修正，誤差值下降至1.97分，相關係數為0.9907，數據結果相當良好。 Call center has become the communication bridge between enterprises and customers, agents play an important role to between enterprises and customers. Enterprises in agents training and snares set of assessment criteria need to be developed. How to avoid human factors affecting service agent performance score, and pass this valuable knowledge, all enterprises are brainstorming. In this study, the sample data are based on historical data of internal systems, and with reference to 23 quantitative indicators about call center operations. The experimental process includes analyzing data correlation coefficient between performance score and each attribute, selecting feature sets by feature selection technology, and predicting the performance score of call center agents by back-propagation neural networks. By selecting a specific attribute set, the experiment result not only contains enterprises managers to score factors but also the main standard for performance appraisal of call center agents. We can know that each attribute in the feature set has a high probability to be considered when managers are making performance appraisal. The average error and the correlation coefficient between the output value from the prediction model and the real call center agents performances score are 2.78 and 0.9821, respectively. The experiment results show that the output from the network prediction model is very close to the real call center agents performance score. Its deviation is also lower than the expected value 3 of this study. In the end of the study, we correct some erroneous and further lower data the deviation to 1.97 and the correlation coefficient to 0.9907.This is very well.
