Understanding the households’ residential relocation dynamics is of great importance in urban land-use modeling. The duration of residence at a location and its dependence on accessibility is one of the less studied areas of residential relocation. Residential relocation duration data is usually modeled using classical survival analysis approaches. The last few years’ huge progress in the field of data science and machine learning has provided a fruitful opportunity for improving such models and making the most out of the heterogenous and high-dimensional residential relocation data. This study benchmarks the performance of ten machine learning approaches combined with six feature selection methods that are capable of handling survival data. The models are also compared with parametric and non-parametric classical survival approaches. The 30-minute accessibility of home locations to job opportunities and accessibility of household to workplace and education locations were estimated and included. The residential self-selection issue is also addressed by incorporating the direct questioning approach. A comprehensive life-course dataset of 1,024 residential relocations in the metropolitan area of Sydney, Australia is utilized. A linear model on the households’ tendency to increase/decrease their accessibility after relocation, revealed important policy implications including reemphasizing the importance of public transport equity and scattering job opportunities for a more sustainable environment and human well-being. Using a concordance index measurement, the results indicate that classical approaches could reach 52% accuracy without any feature selection and 76% With feature selection approaches, while machine learning approaches have generally a high-performance accuracy regardless of the learner and feature selector.

A retrospective survey of residential relocation in Sydney, Australia, containing household unit members’ socio-demographic information, vehicle ownership, occupation, and education records. The data cannot be made publicly available due to its confidentiality agreement. The below figure presents the frequency of home and job locations of the households in the metropolitan area of Sydney.

The accessibility measures are calculated for each suburb in Sydney. A heatmap of the accessibility measure for both car and public transport is presented in the below figure.

The dataset contains a huge number of variables (features) that cannot all be included here. However, a list of some of the variables with their mean and standard deviation values are provided in the below table. Please note that the survey is conducted in both Sydney, Australia and Chicago, US and both are included in this table for the ease of comparison.

In this study, 10 machine learning algorithms are included and compared to three widely-used classical survival analysis methods. All the learners are as follows:

[1] Cox Proportional Hazard (Cox-PH)-non-parametric [2] Weibull-based Accelerated Failure Time (AFT) parametric [3] Lognormal-based Accelerated Failure Time (AFT) parametric

[1] Cross-validation regularized cox proportional-hazards models (cv.glmnet)

1. Ridge regularization
2. Lasso regularization
3. Elastic net regularization
   [2] Gradient boosting methods
4. Survival gradient boosting model (GBM)
5. Boosted generalized linear survival model (GLMBoost)
6. Extreme gradient boosting survival model (Tree-based) also listed in 4.
7. Extreme gradient boosting survival model (Linear based)
   [3] Survival tree model
8. rpart
9. Extreme gradient boosting survival model (Tree-based)also listed in b.
   [4] Random forest models
10. Survival random forest SRC (RFSRC)
11. Ranger survival model

Feature selection methods uses in this study are as follows:
[1] The univariate model score
[2] Minimum redundancy, maximum relevance (MRMR)

[3] feature importance
[4] minimal depth
[5] feature hunting
[6] ranger impurity

Targets package is used in this study for arranging a pipeline for all the study steps from the data preparation to data analysis. Two pipelines are included in this study, (1) the accessibility pipeline (accessibility_targets.R) and (2) the main pipeline (_targets.R). The accessibility command includes the process of retrieving raw data (i.e., land-use, transportation network, and OSM data), finding each suburbs’ land-use status, households’ travel time to work and school using car and public transport, and calculating the 30-minute accessibility of each suburb members to jobs. The main pipeline includes the data preparation, model preparations, regression modelling, machine learning benchmark, and post-processing steps. tar_make() command can be used for running each pipeline.