Social demographics imputation based on similarity in multi-dimensional activity-travel pattern: A two-step approach

In response to the absence of demographics in increasingly emerging big data sets, we propose a novel method for inferring the missing demographic information based on similarity in people’s daily multi-dimensional activity-travel patterns as well as the characteristics of the area they move about. Instead of using isolated activity-travel attributes to infer social demographic features, our proposed method first calculates the similarity of people’s multidimensional daily activities and travels as well as characteristics of their visiting locations, between those for whom the social demographics are to be imputed (target) and those with known demographics (base) using a polynomial function. The weights of the function are determined using the permutation feature importance method, and then dynamic time warping is used to align the multidimensional activity sequences of the base and target sample and measure their similarities. For each person in the target database, a matched list is created consisting of those with the most similar activity-travel sequences in the base sample. A support vector machine is then trained using the base sample as input to impute the demographics of the target sample. The proposed model is trained using a national travel survey and validated by applying it to a GPS dataset. The results show that the proposed method outperforms existing methods in predicting four selected demographics: gender, age, education level, and work status, with an accuracy range between 91% and 94% for the national dataset and 88% to 91% for the GPS data. This study highlights the importance of considering the multidimensional and sequential nature of peoples’ daily activity-travel patterns in the imputation of demographic features.