[Paper Review] Multi-task learning of time series and its application to the travel demand

Boris Chidlovskii|arXiv (Cornell University)|Dec 21, 2017

Time Series Analysis and Forecasting12 references5 citations

TL;DR

This paper proposes a multi-task learning framework for joint prediction of interdependent time series in public transport demand forecasting, using regularization-based support vector regression with two alignment strategies: dynamic time warping (DTW) for clustering similar series and a continuous profile model (CPM) to infer a shared latent representation. The CPM-based approach outperforms single-task learning by 3%-12% in mean squared error across multiple scenarios, demonstrating improved robustness and accuracy in travel demand prediction.

ABSTRACT

We address the problem of modeling and prediction of a set of temporal events in the context of intelligent transportation systems. To leverage the information shared by different events, we propose a multi-task learning framework. We develop a support vector regression model for joint learning of mutually dependent time series. It is the regularization-based multi-task learning previously developed for the classification case and extended to time series. We discuss the relatedness of observed time series and first deploy the dynamic time warping distance measure to identify groups of similar series. Then we take into account both time and scale warping and propose to align multiple time series by inferring their common latent representation. We test the proposed models on the problem of travel demand prediction in Nancy (France) public transport system and analyze the benefits of multi-task learning.

Motivation & Objective

• Address the challenge of predicting multiple interdependent time series in intelligent transportation systems, particularly for public transport demand.
• Overcome limitations of single-task learning by leveraging shared information across related time series to improve prediction accuracy.
• Develop a regularization-based multi-task support vector regression model adapted for regression tasks on time series.
• Address time and scale warping in time series by introducing a common latent representation for alignment.
• Evaluate the proposed methods on real-world travel demand data from Nancy, France, with external features like weather, calendar, and traffic.

Proposed method

• Use dynamic time warping (DTW) to measure similarity between time series and cluster related series for multi-task learning.
• Propose a continuous profile model (CPM) to infer a shared latent representation across all time series, enabling alignment via Viterbi-like inference.
• Apply regularization-based multi-task learning to support vector regression, sharing model parameters across related tasks to improve generalization.
• Integrate external features (weather, calendar, traffic) into the prediction model to enhance predictive performance.
• Use Viterbi-style decoding to align observed time series to the inferred common latent profile for improved temporal alignment.
• Train and evaluate models using historical boarding data from Nancy’s public transport network, with varying aggregation periods (1h and 10min).

Experimental results

Research questions

• RQ1 Can multi-task learning improve the accuracy of time series prediction in public transport demand forecasting compared to single-task learning?
• RQ2 How effective is dynamic time warping (DTW) in identifying clusters of related time series for multi-task learning in this context?
• RQ3 Does inferring a common latent representation (CPM) lead to better alignment and prediction performance than DTW-based clustering?
• RQ4 How do external factors such as weather, holidays, and traffic conditions affect the predictive performance of the models?
• RQ5 How does data delay (e.g., missing last observation) impact the robustness of multi-task learning models compared to single-task models?

Key findings

• Multi-task learning with the continuous profile model (CPM) reduces mean squared error (MSE) by 3% to 12% compared to single-task learning across all experiments.
• The DTW-based clustering method improves prediction accuracy by 2% to 7% over single-task learning, but is less robust than CPM, especially under data delay.
• In the 1-hour period zone experiment, CPM-based multi-task learning outperforms single-task learning by 3% to 12% in the 0-delay scenario.
• Under data delay (1-delay scenario), CPM-based multi-task learning maintains superior performance, improving accuracy by 3% to 16% over single-task learning.
• The inclusion of external observations (weather, calendar, traffic) reduces MSE by 40% to 60% compared to models without such features.
• The CPM-based method shows greater stability and consistent performance gains across varying numbers of stops (2 to 20) and different delay conditions (P=40% and 90%).