Next activity prediction from individual daily mobility patterns

Next activity prediction from individual daily mobility patterns Xianchen Wang, Nilufer Sari Aslam, Chen Zhong The Bartlett Centre for Advanced Spatial Analysis is part of The Bartlett, UCL Contact: [email protected] Aim and Objectives ➢Model the dynamics of individual's daily movements ➢Predict human activities in the next time interval Figure 1: Illustration of the next activity prediction ➢Individual-level semantic trajectory data extracted from mobile phone applications ➢The dataset consists of records in four dimensions: user ID, time, location and activity type ➢There are 1475 users with 104,069 trajectory records within 30 days (November 2021) ➢Human’s daily activities: ➢home (H) ➢work (W) ➢eating and drinking (EaD) ➢education (Ed) ➢entertainment (En) ➢frequent shopping (S1) ➢non-frequent shopping (S2) ➢others (O) Figure 2: Visualisation of all individual trajectories Figure 3: Quantitative distribution of different types of activities Methodology ➢Use embedding layer to learn semantic relationship ➢Construct training and test datasets using sliding windows Figure 4: Next activity prediction framework Figure 5: Structure of proposed MTL-LSTMs model ➢Multi-category classification task ➢The primary challenge lies in the imbalance among activity categories ➢Transform the task into multiple binary classification tasks, leveraging concepts from ensemble learning and multi-task learning Figure 6: Evaluating the model performance using various performance metrics Figure 7: Precision-recall curve for MTL-LSTM ➢The inverse relationship between activity entropy and prediction accuracy, i.e., lower entropy with higher prediction accuracy Figure 8: Relationship between activity entropy and prediction accuracy Conclusion ➢Among these metrics, the highest score in MTL-LSTM is captured in recall Results and Findings Data Description ➢The proposed approach effectively improves recall, particularly for routine activities, while highlighting challenges in predicting less structured behaviours ➢Further refinement of multi-task learning strategies and ensemble techniques can enhance prediction accuracy for irregular activities ➢The prediction results illustrate higher accuracy for routine (regular) activities, such as home, work, and education ➢Less structured activities, such as eating, entertainment, and shopping, obtain lower accuracy from individuals' daily activity chains Calculation formulas for different metrics