MASTER User behaviour analysis and prediction based on device logs

Nowadays, with the increasing usage of IoT (Internet of Things) devices in daily lives, the data collected from those connected devices can help understanding user behaviours and delivering better services to individual users. In this thesis, we explored the device log data collected from connected air purifier devices to analyze usage behavior patterns. Further, we integrated the device log data with external data set to invesitage the impact of external factors on the user behavior. We studied different user behaviour patterns from two aspects. The first one only focuses on the ON/OFF state of the device and four daily usage patterns based on the power state are discovered. Then those patterns are utilized in the prediction model as extra features to improve the performance of prediction. The second one splits power state ON as two operating modes AUTO and OVERRIDE, and the focus is put on the three operating modes: AUTO/OVERRIDE/OFF. Then three metrics are proposed to evaluate user behaviours from different aspects: aggregated operating mode ratio, histogram and distribution of operating mode ratio. Then we study potential factors that may have impact on user behaviours. These analyses produce more insights about how people use their devices and know how these factors affect user behaviours. Then two prediction models predicting power state and operating modes separately are established using those factors in analysis part as features with two purposes. The first one is to validate the observations in analysis part by comparing prediction performance with and without those factors. Results show that improvement of prediction performance caused by those factors are significant. The second one is to facilitate the further goals like operating mode automatization by evaluating the prediction model. Results show that the final two predictors taking analysed factors as features perform pretty well with average weighted F1 score 0.990 and 0.977 separately. Finally main findings of this thesis are generalized as well as limitations and possible future jobs. This thesis shows the feasibility of using IoT device logs to analyse user behaviours and the possibility to supply better services to users based on the understanding of user behaviours. User behaviour analysis and prediction based on device logs iii