New Perspectives on the Use of Online Learning for Congestion Level Prediction over Traffic Data

Eric L. Manibardo, Ibai Laña, Jesus L. Lobo, Javier Del Ser

TL;DR

This paper investigates the effectiveness of online learning for predicting traffic congestion levels from speed data, a task complicated by evolving traffic patterns (concept drift). They demonstrate that classifying congestion levels directly, rather than predicting speed and then classifying, yields better results, and highlight the critical importance of analyzing class distribution before model design.

What Problem Does This Paper Address?

Predicting traffic congestion is crucial for intelligent transportation systems. Traffic data is inherently a "data stream" – it arrives continuously and can be massive. Traditional machine learning models trained on historical data can become obsolete as traffic patterns change over time (concept drift). This necessitates "online learning" methods that can adapt incrementally to new data. Furthermore, the authors point out that many existing studies overlook the dynamic nature of traffic data, particularly the distribution of different congestion levels, which can significantly impact model performance. This paper aims to provide a comprehensive comparison of various online and offline learning methods for congestion prediction, specifically addressing the challenges posed by concept drift and varying class distributions.

Methods

The study compares a wide range of learning algorithms, categorized into:

  • Shallow Learning Methods: These include traditional algorithms like Naive Bayes, K-Nearest Neighbors (with drift detection), Perceptron, and ensemble methods like Adaptive Random Forest (ARF), Additive Expert Ensemble (AEE), and Dynamic Weighted Majority (DWM). Many of these are implemented using the Scikit-Multiflow library, designed for stream learning.
  • Deep Learning Methods: A novel hybrid convolutional-recurrent neural network (OCLSTM) is proposed, combining 1D convolutional layers for short-term dependencies and Long Short-Term Memory (LSTM) units for long-term temporal relationships. Variations with GRUs and without convolutional layers are also explored (OLSTM, OCGRU, OGRU).
  • Extreme Learning Machine (ELM): An online sequential implementation (OS-ELM) is also included.
  • Naïve Method (NM): A simple baseline that predicts the last observed value.

The study evaluates both offline (batch training on a fixed dataset) and online (incremental learning) versions of these methods. A key aspect is the comparison between two strategies: (A) predicting speed and then discretizing into congestion levels, and (B) labeling the data first and then performing classification.

Datasets

The experiments utilize real-world traffic speed data collected from inductive loops on four freeways in Seattle, USA (I-5, I-405, I-90, and SR-520) throughout the entire year 2015. This dataset provides speed measurements every 5 minutes for each of the 322 inductive loops. For the study, four specific locations (ATRs) were selected to represent different traffic profiles.

Key Insights and Ideas

  • Classification vs. Regression for Congestion: Directly classifying congestion levels (Strategy B) significantly outperforms predicting speed and then discretizing (Strategy A), especially in online settings. This suggests that modeling the discrete states of traffic is more robust than predicting a continuous value that then needs interpretation.
  • The Criticality of Class Distribution Analysis: The paper strongly advocates for analyzing the distribution of congestion classes over time before designing a model. This pre-analysis can reveal the inherent predictability of different levels and explain why models might struggle with certain classes, particularly during transitions.
  • Concept Drift is a Major Hurdle: As expected, online learning methods generally outperform their offline counterparts, especially when dealing with evolving traffic patterns. However, the degree of improvement varies, and some offline models can still perform well if the initial training data is representative.
  • Forecasting Horizon Impact: Performance degrades severely as the forecasting horizon increases. This degradation is most pronounced for the "congestion" class, as traffic states tend to transition quickly between free-flow and bottleneck, making intermediate congestion states harder to predict over longer periods.
  • No Single "Best" Model: While adaptive learning methods like ARF, HT, HATT, and VFDR show high performance in online settings, and the proposed deep learning architectures also perform well, there isn't one universally superior model. The best choice can depend on the specific traffic profile and the chosen strategy.

Key Results

  • Strategy B (Classification) Superiority: Table I shows that the OCLSTM model, when used for classification (Strategy B), achieves a higher overall F1-score (UMF1) than when used for regression (Strategy A), both in offline and online settings. This holds true across different locations.
  • Online Learning Benefits: Table II demonstrates that online learning generally leads to improved UMF1 scores compared to offline learning for most methods, particularly for adaptive algorithms like ARF, HT, HATT, and VFDR. For example, ARF achieves UMF1 scores of 0.961 and 0.948 in online settings for two locations, significantly higher than its offline counterparts.
  • Deep Learning Performance: The proposed OCLSTM and its variants (OLSTM, OCGRU, OGRU) generally perform better than the Naïve Method (NM) in online settings, with OCLSTM often showing competitive results.
  • Horizon Degradation: Figure 2 visually confirms that performance, especially for the "Congestion" class, drops significantly as the forecasting horizon increases from 1 to 20 (5-minute intervals). The "Bottleneck" class also suffers, but the "Congestion" class is identified as the most challenging due to its transitional nature.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that the performance of some models, particularly OS-ELM, can be sensitive to hidden layer initialization.
  • The dataset, while extensive, is from a specific geographical region (Seattle), and its generalizability to other urban environments might need further validation.
  • The paper focuses on speed data; incorporating other traffic variables could potentially improve predictions.
  • Future Work:
    • The authors suggest exploring transfer learning to leverage models trained on one highway segment for predicting traffic on another, potentially reducing the need for extensive training data at new locations.
  • Further investigation into the similarities between traffic profiles at different road network points could inform more sophisticated transfer learning approaches.
  • Developing models that excel at identifying the precise moments of transition between traffic states remains a key area for future research.

Summary updated

Abstract

This work focuses on classification over time series data. When a time series is generated by non-stationary phenomena, the pattern relating the series with the class to be predicted may evolve over time (concept drift). Consequently, predictive models aimed to learn this pattern may become eventually obsolete, hence failing to sustain performance levels of practical use. To overcome this model degradation, online learning methods incrementally learn from new data samples arriving over time, and accommodate eventual changes along the data stream by implementing assorted concept drift strategies. In this manuscript we elaborate on the suitability of online learning methods to predict the road congestion level based on traffic speed time series data. We draw interesting insights on the performance degradation when the forecasting horizon is increased. As opposed to what is done in most literature, we provide evidence of the importance of assessing the distribution of classes over time before designing and tuning the learning model. This previous exercise may give a hint of the predictability of the different congestion levels under target. Experimental results are discussed over real traffic speed data captured by inductive loops deployed over Seattle (USA). Several online learning methods are analyzed, from traditional incremental learning algorithms to more elaborated deep learning models. As shown by the reported results, when increasing the prediction horizon, the performance of all models degrade severely due to the distribution of classes along time, which supports our claim about the importance of analyzing this distribution prior to the design of the model.
BibTeX
@inproceedings{manibardo2020new,
  author = {Eric L. Manibardo and Ibai Laña and Jesus L. Lobo and Javier Del Ser},
  title = {{New Perspectives on the Use of Online Learning for Congestion Level Prediction over Traffic Data}},
  booktitle = {IJCNN},
  year = {2020},
  url = {https://openreview.net/forum?id=35y5egqCEk},
}
Figure