Source-linked AI summary
Deep Sequence Learning with Auxiliary Information for Traffic Prediction
Binbing Liao, Jingqing Zhang, Chao Wu, Douglas McIlwraith, Tong Chen, Shengwen Yang, Yike Guo, Fei Wu
TL;DR
Traffic prediction is difficult because large-city road conditions involve complex interactions and available large-scale multimodal datasets are limited. The paper releases Q-Traffic and integrates geographical and social attributes, road intersections, and crowd-query impacts in a hybrid encoder-decoder Seq2Seq model. The model achieves the best reported performance across whole-period and event-focused evaluations, with query impact showing the largest improvement among the auxiliary domains.
Problem
Traffic prediction must handle complex interactions among roads and crowds, while large-scale datasets combining traffic speed with relevant auxiliary information are limited.
Method
A hybrid encoder-decoder Seq2Seq framework integrates offline geographical and social attributes, road-intersection spatial dependencies, and quantified online crowd-query impact.
Results
The hybrid model achieves the best reported performance, with overall MAPE 8.63% and 2-hour forecasting MAPE 9.78% on the whole testing set, and overall MAPE 9.22% during events.
Takeaways & Limitations
Appropriately integrating the three auxiliary domains boosts traffic-speed prediction during both event periods and the whole time period, with query impact showing larger improvement than the other domains.
Takeaways & Limitations
The query-based event assumption may be imperfect because users may not travel to the destinations they search; the dataset filters queries using proximity and recency assumptions.
Abstract
from arXiv · showhide
Predicting traffic conditions from online route queries is a challenging task as there are many complicated interactions over the roads and crowds involved. In this paper, we intend to improve traffic prediction by appropriate integration of three kinds of implicit but essential factors encoded in auxiliary information. We do this within an encoder-decoder sequence learning framework that integrates the following data: 1) offline geographical and social attributes. For example, the geographical structure of roads or public social events such as national celebrations; 2) road intersection information. In general, traffic congestion occurs at major junctions; 3) online crowd queries. For example, when many online queries issued for the same destination due to a public performance, the traffic around the destination will potentially become heavier at this location after a while. Qualitative and quantitative experiments on a real-world dataset from Baidu have demonstrated the effectiveness of our framework.
1 INTRODUCTION
Traffic prediction in large cities is important but difficult because dynamic road conditions, sparse multimodal datasets, and interacting geographical, social, spatial, and crowd-query factors complicate forecasting. The paper introduces Q-Traffic and a hybrid Seq2Seq framework that integrates these auxiliary information sources.
- Motivation: Traffic prediction supports network planning, route guidance, and congestion avoidance, but large-city environments make accurate forecasting challenging.The limited potential for new roads increases the importance of network management.
- Auxiliary information: Geographical road structure, junctions, holidays, and weekends influence traffic dynamics and contribute to prediction difficulty.Major junctions generally experience congestion more often than lanes.
- Auxiliary information: At 18:00, unusually high queries around Capital Gym coincided with a sudden traffic-speed drop before the Fish Leong Concert.Queries use estimated arrival times, making them potentially useful as an early warning of traffic jams.
- Dataset gap: Previous traffic-prediction research commonly relied on limited datasets, and relatively few large-scale public datasets supported model comparison.This motivates releasing a dataset containing traffic speed together with offline and online auxiliary information.
- Dataset contribution: Q-Traffic combines traffic speed with geographical and social attributes, road intersections, and online crowd queries from Baidu Map.The dataset is intended to address the obstacle of limited high-quality multimodal traffic data.
- Method: The proposed hybrid Seq2Seq model uses wide transformations for sparse offline attributes, graph convolution for neighboring-road spatial correlations, and query-impact encoding for crowd queries.These components are combined through a deep fusion within an encoder-decoder sequence-learning framework.
2 Q-TRAFFIC DATASET
The Q-Traffic dataset combines query, traffic-speed, and road-network sub-datasets to represent online crowd activity, traffic conditions, and road structure. Its processing includes grid-based arrival-time event discovery, traffic-speed smoothing, and road attributes with geographical and social information.
- Q-Traffic consists of query, traffic-speed, and road-network sub-datasets for traffic prediction.It includes offline geographical and social attributes, road-network information, and online crowd queries.
- 2.1 Query Sub-dataset: About 114 million Baidu Map queries were collected in Beijing between April 1 and May 31, 2017, recording search, location, and destination information.The dataset includes anonymised user IDs, timestamps, coordinates, query words, and query modes.
- 2.1 Query Sub-dataset: Queries are filtered, mapped onto a 72 × 68 grid, and assigned estimated arrival times using mode-specific travel speeds.Redundant queries are reduced, distant current and starting locations are removed, and arrival times are estimated for route and location searches.
- 2.1.1 Event Discovery: The arrival-time tensor supports event discovery by identifying spatiotemporal query ranges whose counts exceed usual levels.Queries are aggregated every 15 minutes, producing T = 5,856 timestamps; the tensor is sparse with density ρd = 4.
- 2.1.1 Event Discovery: Using η = 0.2, ζ = 300, ∆t = 672, and ϵ = 4, the procedure discovers 932 events, including concerts, forums, attractions, and anniversaries.More than 80% of event query counts come from the top query word, which is highly related to the event.
- 2.2 Traffic Speed Sub-dataset: The traffic-speed sub-dataset contains 15,073 road segments covering approximately 738.91 km, with speeds smoothed over 15-minute windows.The road-network sub-dataset supplies topology, geographical attributes, and social attributes such as holidays and peak hours.
3 METHODOLOGIES
The methodology extends Seq2Seq traffic forecasting with offline attributes, spatial road-network information, and online query impact. These auxiliary signals are integrated into encoder and decoder components to represent temporal, spatial, contextual, and crowd effects.
- 3.1 Problem Definition: Seq2Seq forecasts future traffic speed from a previous traffic-speed sequence using an encoder-decoder architecture.The encoder processes the input sequence, and its final hidden state is passed to the decoder for future-speed prediction.
- 3.3 Seq2Seq + Attributes: Geographical and social attributes are concatenated into the decoder while the Seq2Seq encoder remains unchanged.The attributes include road width, direction, speed limit, lane count, holidays, workdays, and peak-hour information.
- 3.4 Seq2Seq + Spatial Relation: Graph convolution embeds neighbouring road-segment traffic into the encoder to model spatial dependencies that Seq2Seq alone does not capture.Five predecessors and five successors are selected using PageRank scores in a directed local road graph.
- 3.5 Seq2Seq + Query Impact: Query impact measures how online queries influence road segments using query counts, spatial distance, and a decreasing exponential impact function.The impact factor is represented by h(x) = exp(−x/σ), where σ controls the influence decay.
1 Initialisation: QI(l,t) ←0
The query-impact sequence is encoded and combined with the other auxiliary domains in the hybrid traffic-prediction model.
- 3.5 Seq2Seq + Query Impact: An RNN with LSTM encodes the query-impact sequence before its final hidden state is combined with the Seq2Seq decoder.The hybrid model simultaneously incorporates attributes, spatial relation, and query impact.
4 EXPERIMENTS
Experiments compare conventional regressors and several Seq2Seq variants on whole-period and event-specific traffic prediction. The hybrid model performs best overall, while query impact provides particularly strong gains during events.
- 4.1 Compared Methods: The evaluation compares RF, SVR, Seq2Seq, three single-auxiliary variants, and the hybrid model.The deep models are trained with Adam, hidden-state dimension 128, and a first-month training split followed by second-month testing.
- 4.2 Experimental Setting: The task uses one day of 15-minute traffic-speed history to predict the subsequent 2 hours, corresponding to input length t = 96 and output length t′ = 8.Random forests and SVR require prior forecasts for subsequent predictions during testing.
- 4.3 Evaluation Metrics: MAPE is evaluated across the whole testing set (ErrT) and during events (ErrE), using actual and predicted traffic speeds.The event-specific metric isolates performance under event-related traffic conditions.
- 4.4 Results and Discussion: 8.63% overall MAPE and 9.78% 2-hour forecasting MAPE are achieved by the hybrid model on the whole testing set.These are reported as the best performances in the quantitative comparison.
- 4.4 Results and Discussion: 9.22% event-period overall MAPE is achieved by the hybrid model, compared with 9.34% for Seq2Seq + Query Impact.Event traffic is more difficult to predict, especially beyond one hour.
- 4.4 Results and Discussion: Appropriate integration of the three auxiliary domains improves prediction during both event periods and the whole testing period.The authors report query impact as more effective than the other two auxiliary domains.
5 RELATED WORK
Related work covers parametric and non-parametric traffic prediction, including classical regressors, neural networks, and models using auxiliary or spatial information. Prior approaches face limitations involving stationarity, nonlinear urban traffic, sparse sensors, and limited spatial coverage.
- 5.1 Traffic Prediction: Traffic-prediction methods are commonly classified as parametric or non-parametric, with ARIMA, RF, SVR, Bayesian networks, and neural networks among representative approaches.ARIMA assumes a stationary process and is described as computationally unsuitable for large-scale traffic data.
- 5.1 Traffic Prediction: Many earlier deep-learning studies focus on relatively stable highway traffic, whereas urban traffic can vary greatly because of traffic lights.An LSTM-based city study is noted, but its experiments used only two points.
- 5.2 Traffic Prediction via Auxiliary Domains: Prior multimodal methods use social-media traffic indicators or detector data, but linear regression may be insufficient for nonlinear traffic and sensors cover only a small fraction of roads.Other work uses CNN-based models to incorporate spatial information among road-network nodes.
6 CONCLUSIONS
The paper addresses traffic prediction’s data and modeling challenges by releasing Q-Traffic and integrating three auxiliary information domains into sequence-to-sequence learning. A hybrid model combining these domains outperforms compared methods.
- 6 CONCLUSIONS: Q-Traffic is a large-scale Baidu Map dataset combining traffic speed data with offline and online auxiliary information.Its domains include geographical and social attributes, road intersections, and crowd map queries.
- 6 CONCLUSIONS: The framework integrates crowd map queries, road intersections, and geographical and social attributes into sequence-to-sequence traffic prediction.
- 6 CONCLUSIONS: The proposed hybrid model combines all three auxiliary information domains and is superior to compared methods.
- 6 CONCLUSIONS: The released dataset is intended to encourage further approaches for accurate and real-time traffic prediction.