Evaluierung der Benutzerinteraktion in der Telekommunikationsindustrie : Evaluation of user interaction in the telecommunication industry

Abstract

The telecommunication domain is marked by a vast number of products and services on the one side and on the other side, telecommunication providers have access to substantial data. The main goal of this thesis is to address the challenge of understanding users of the telecommunication domain with their information needs and preferences. Based on this knowledge, personalized offers of services and products can be given to users. Recommender systems aim to support their users during their decision making processes by suggesting products or services, which match the users needs. To achieve this goal, we propose four different approaches. First, text mining is applied on chat conversations between users and a telecommunication chatbot to determine users topics of interests. Secondly, further text mining techniques are applied on users feedback scores and comments to decide if users are satisfied with the received answers or not. Thirdly, event sequence analysis is applied on sequences of events, which are extracted out of chat conversations. Finally a prediction model is generated and trained based on users clicks on the homepage of a telecommunication company. Our approaches are tested using data of an Austrian telecommunication company. Summarizing the results of this work, it is possible to address the challenge of understanding users information needs, preferences and satisfaction. Analyzing the chatbot data, billing, homenet and un-/locked simcards are determined as users most popular topics of interests in chat conversations. Moreover, feedback scores given to chatbots answers and feedback comments in textual form are analyzed. The majority of feedback is negative and considerable differences in feedback for different topics is observed. Finally, the chatbots answer quality is calculated by analyzing users chat behavior after receiving an answer. The majority of users continue chatting after reaching an answer node. Based on the clickstream data, a prediction model for users future clicks is generated. Using the test dataset as input to the model, approximately half of the clicks are predicted correctly. In general, the introduced techniques and models can be applied to any domain, chatbot or clickstream dataset.