Research Information System
Thesis Type: Postgraduate
Institution Of The Thesis: Gazi University, Fen Bilimleri Enstitüsü, Turkey
Approval Date: 2023
Thesis Language: Turkish
Student: Buğra Kaan TÜRKMENOĞLU
Supervisor: Oktay Yıldız
Open Archive Collection: AVESIS Open Access Collection
Abstract:
The helpdesk system is the main communication tool between users and customer service representatives. Because of the manual selection of the helpdesk ticket field, the ticket field may be assigned incorrectly, the resolution time of the ticket may be extended, and the workforce may be lost. In case of wrong selection ticket priority, non-critical tickets may be given higher priority. Customer service representatives accumulate a large number of tickets due to the necessity of having knowledge about many subjects, the difficulty of transferring the process, and repetitive tickets. To overcome these problems, it is aimed to design models that will classify tickets according to upper and lower fields and their priorities, and that will suggest the closing of tickets that have the same meaning as tickets closed by customer service representatives. A hierarchical approach was used to overcome the difficulty of multi-class ticket field classification. For the training of ticket classification models, four datasets were collected from a company's help desk system, with the upper field, lower field and priority classes, and ticket description. To train the similarity model, tickets belonging to a ticket upper field were mapped as cartesian and manually labeled. Thanks to data preprocessing steps, resampling and dimension reduction, the challenge of multiclass text classification problem was overcome by using traditional machine learning algorithms on an unbalanced dataset. The classification of the tickets to the upper field with 94.4% accuracy and to lower fields with an accuracy between 88.22% and 98.92% was carried out using the SVM method. Priority classification of tickets with an accuracy of 97.35% was performed using the K-NN method. Using the MaLSTM model, 0.235 mean squared error, 0.4848 root mean square error, 0.4603 mean absolute error, 0.2528 Pearson correlation coefficient, and 0.2487 Spearman's rank correlation coefficient were obtained for ticket similarity. This thesis study was supported within the scope of TUSAŞ Scientific Research Program (TUSAŞ BAP).
| Key Words : Help desk, ticket classification, semantic text similarity, machine learning, natural language processing, Word2vec, Doc2vec, LSTM |