Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Bilal Farooq, Ryerson University
IATBR 2018, Santa Barbara
§ Topics covered§ Route choice prediction§ Mode choice prediction§ Discrete-continuous mix prediction§ Spatial structure of travel/activities§ Travel/activity pattern inference
§ Artificial Neural Networks [2 paper], DNN [3], CNN [1], RBM [1]
§ Decision Tree/Random Forest [3]
§ Clustering approaches [3]
§ Ensemble machines [1]
1. What is the scope of data-driven learning in the context of travel behaviour modelling
2. What are the key gaps in the research?
3. Develop three research projects that address these gaps
§ July 16, 2018§ 4:00-4:15PM Introduction to the workshop§ 4:15-4:30PM Participants introduction§ 4:30-5:15PM Developing the problem statement§ 5:15-6:30PM Identifying key research gaps
§ July 18, 2018§ 4:00-4:15PM Recap of the workshop§ 4:15-4:20PM Formation of three groups§ 4:20-4:50PM Research projects sketch§ 4:50-5:00PM Presentation/feedback§ 5:00-6:00PM Interaction with time use and travel workshop
§ Dr. Shadi Djavadian (Ryerson)
§ Melvin Wong (Ryerson)
§ Georges Sfeir (AUB)
§ Vishnu Baburajan (IST)
Discriminative models
§ Good for:§ Extraction and analysis of travel patterns
§ Purpose of trip§ Mode of transportation§ Travel activity/diary
§ Classification of travellers
§ Key advantages in the case of activity/mobility surveys using GPS data from smartphone§ Cheaper§ Managing big data sources§ …
Discriminative models
§ Classifying major modes/purpose only§ Ignoring the purpose since it’s not an easy task to detect?§ Abstract representation of purpose
§ Such models good for capturing unique patterns§ Our responsibility to put semantic meaning to them
Generative models
§ Good for:§ Predictive modelling§ Exploring the distribution and correlations of variables§ Dealing with missing data§ Population synthesis § Merging multiple data sources
§ Spatio-temporal transferability of models§ Assumption that behaviour remains the same
Generative models
§ Imbalance data: applications can be risky§ Case of elderly population without smartphones§ Use of probabilistic models based on historical data to predict
missing part of the data (e.g. When phone is off)
§ Such models can be useful for diagnostics§ Case of identification of latent classes
§ When and how to use data-driven learning techniques?§ Alchemy!
§ Interpretation of the model; what can be done with them and what cannot; and what is their use
§ Incorporating dynamics in data-driven models§ Beyond LSTM/time series models
§ Use in forecasting (especially the generative models)
§ Data-driven estimation techniques for hypothesis-driven modelling§ Advancements in stochastic gradient decent
§ Exploring the abstract representation of travel purpose (and mode)
§ Benchmark datasets§ Openly available datasets from North America, Europe, Asia
§ Using such techniques to capture unexplainable dimensions of hypothesis-driven modelling
§ Individual specific modelling§ Rich longitudinal data on individuals
§ Privacy preserved model estimation
§ Incorporating context-aware variables in data-driven approaches
§ Improving predictive accuracy of discrete choice models with machine learning while maintaining interpretability§ Exploration of hybrid model formulations
§ Selection processes for variables/features for interpretable and uninterpretable parts of utility function
§ Exploration of models for the uninterpretable information
§ Trade-off analysis
§ Benchmark dataset for comparative analysis
§ Definition of dataset§ Which decision variable? Or families of decisions?
§ Balanced data§ What location?§ 1 day vs multiple days§ Size of data§ Related transportation data
§ Role of Kaggle sort of data repositories
§ Use of synthetic data?
§ Predictive power: what usage and and what cost
§ To what extent is privacy important in travel behaviour?
§ What could be the implications of masking/filtering private data in travel behaviour?
§ Training of privacy aware and counterpart models
§ Quantification of: § Improvement in privacy§ Semantic data needs
§ Joint discussion on: § Theory/Hypothesis-driven and Data-driven approaches
§ Large dataset can inspire new theories
§ Predictability vs Transferability
§ Interpretability§ What’s inside!?
§ Bayesian origin of machine learning
§ What problems are good to use this tool and what are not?