Design of information tree for support related queries: Axis Communications AB : An exploratory research study in debug suggestions with machine learning at Axis Communications, Lund

Sammanfattning: Context: In today's world, we have access to so much data than at any time in the past with more and more data coming from smartphones, sensors networks, and business processes. But, most of this data is meaningless, if it's not properly formatted and utilized. Traditionally, in service support teams, issues raised by customers are processed locally, made reports and sent over in the support line for resolution. The resolution of the issue then depends on the expertise of the technicians or developers and their experience in handling similar issues which limits the size, speed, and scale of the problems that can be resolved. One solution to this problem is to make relevant information tailored to the issue under investigation to be easily available. Objectives: The focus of the thesis is to improve turn around time of customer queries using recommendations and evaluate by defining metrics in comparison to existing workflow. As Artificial Intelligence applications can have a broad spectrum, we confine the scope with a relevance in software service and Issue Tracking Systems. Software support is a complicated process as it involves various stakeholders with conflicting interests. During the course of this literary work, we are primarily interested in evaluating different AI solutions specifically in the customer support space customize and compare them. Methods: The following thesis work has been carried out by making controlled experiments using different datasets and Machine learning models. Results: We classified Axis data and Bugzilla (eclipse) using Decision Trees, K Nearest Neighbors, Neural Networks, Naive Bayes and evaluated them using precision, recall rate, and F-score. K Nearest Neighbors was having precision 0.11, recall rate 0.11, Decision Trees had precision 0.11, recall rate 0.11, Neural Networks had precision 0.13, recall rate 0.11 and Naive Bayes had precision 0.05, recall rate 0.11. The result shows too many false positives and true negatives for being able to recommend. Conclusions: In this Thesis work, we have gone through 33 research articles and synthesized them. Existing systems in place and the current state of the art is described. A debug suggestion tool was developed in python with SKlearn. Experiments with different Machine Learning models are run on the tool and highest 0.13 (precision), 0.10 (f-score), 0.11 (recall) are observed with MLP Classification Neural Network.