Source-linked AI summary
Combining time-series and textual data for taxi demand prediction in event areas: a deep learning approach
Filipe Rodrigues, Ioulia Markou, Francisco Pereira
TL;DR
Routine-focused forecasting may miss mobility demand in eventful areas, while web information about events is often unstructured text. The paper combines event text and historical taxi demand with deep learning, and reports reduced forecast error and stronger performance than state-of-the-art methods.
Problem
Routine trends and recent observations can fail to model mobility demand accurately in eventful areas, while relevant event information is often unstructured text.
Source: §1 · pp. 1–2Method
Two deep learning architectures combine event descriptions and historical taxi demand using word embeddings, convolutional layers, and attention.
Source: §1 · p. 2Results
On New York taxi and web-event data, the proposed approaches outperform other state-of-the-art methods and reduce forecast error.
Source: §1 · p. 2Takeaways & Limitations
The authors make the experimental code and datasets public as a data-fusion benchmark for further methods in this application domain.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Accurate time-series forecasting is vital for numerous areas of application such as transportation, energy, finance, economics, etc. However, while modern techniques are able to explore large sets of temporal data to build forecasting models, they typically neglect valuable information that is often available under the form of unstructured text. Although this data is in a radically different format, it often contains contextual explanations for many of the patterns that are observed in the temporal data. In this paper, we propose two deep learning architectures that leverage word embeddings, convolutional layers and attention mechanisms for combining text information with time-series data. We apply these approaches for the problem of taxi demand forecasting in event areas. Using publicly available taxi data from New York, we empirically show that by fusing these two complementary cross-modal sources of information, the proposed models are able to significantly reduce the error in the forecasts.
1. Introduction
The paper targets taxi-demand prediction in eventful areas, where routine trends and recent observations may miss demand surges. It combines event text with historical demand and reports better forecasts than state-of-the-art methods.
1. Introduction
- Routine-trend and recent-observation approaches can miss demand in event-prone areas, where surges make anticipation important.Source: §1 · pp. 1–2
- Web event information may contextualize time-series patterns, but is commonly unstructured text, making cross-domain data fusion central to the problem.Source: §1 · p. 2
- The proposed architectures combine word embeddings, convolutional layers, attention, event descriptions, and historical taxi demand to forecast future demand.Source: §1 · p. 2
- The proposed deep learning approaches outperform other state-of-the-art methods using taxi trips and event data from New York.The dataset contains 1.1 billion taxi trips.Source: §1 · p. 2
2. Related work
Prior work addresses event-related travel, text–time-series fusion, and deep learning for transportation, but the reviewed approaches leave event text and its characteristics underused. The paper positions joint text and demand modeling as a way to address these gaps.
2.1. Urban mobility and special events
- Survey-based event models predict travel outcomes, but often omit event characteristics even though event categories can inform public-transport demand prediction.Source: §2.1 · p. 3
2.2. Data fusion of time-series and text
- Prior text–time-series work includes Twitter indicators associated with stock-market events, while this paper's convolutional layers learn text patterns jointly rather than relying on predefined lexicons.Source: §2.2 · pp. 3–4
- The proposed models avoid bag-of-words representations, preserving word order to capture multi-word patterns in event descriptions.Source: §2.2 · p. 4
- The paper places its models in data-fusion category (iii), which feeds different datasets into different parts of a model simultaneously.Source: §2.2 · p. 4
2.3. Deep learning in transportation
- Earlier deep-learning transportation approaches reviewed here do not account for event effects, while the proposed work combines event text and demand observations to improve predictions.Source: §2.3 · p. 5
3. Proposed methodology
The methodology prepares time-series and text inputs, then fuses their learned representations in two neural architectures for forecasting. It removes recurring mobility trends before prediction and models text with embeddings, convolutions, and optional attention.
3.1. Time-series detrending
- Historical averages for each time-of-day and weekday pair are removed as recurring trends, so models forecast residuals and add the trend back to predictions.The averages are computed from training data; detrending is reported to improve predictive performance.Source: §3.1 · p. 5, §3.3 · p. 11
3.2. Text data preprocessing
- Removing words that appear only once reduces vocabulary size and neural-network complexity without reducing predictive performance.The text pipeline also includes conventional preprocessing transformations.Source: §3.2 · p. 6
3.3. Neural network architecture
- The two proposed architectures differ in how they model time-series data: one uses LSTMs and the other fully connected layers.The authors report that these proposed variants performed best among the architectures they tried.Source: §3.3 · pp. 7, 9
- Text is represented by word embeddings, then convolution and max-pooling layers detect patterns at different abstraction levels; optional attention weights text features using time-series context.The convolutional architecture can capture variable-length patterns in input sequences.Source: §3.3 · pp. 7–9
- DL-LSTM encodes the time series with an LSTM that also receives event indicators, then combines its latent representation with the text representation for prediction.The final dense layer can also receive other relevant inputs, such as weather or event details.Source: §3.3 · pp. 9–10
- DL-FC models lagged observations through fully connected layers, combines the resulting representation with the text representation, and adds back the removed recurrent trend.The lag vector contains observations from t through t − L.Source: §3.3 · pp. 10–11
4. Experiments
The experiments evaluate taxi-demand forecasting in two New York event areas using historical pickups, weather, and Web-derived event information. Adding event text reduces forecast error in both areas, with the full DL-FC model improving MAPE over its L+W version.
4. Experiments
- The proposed architectures combine time-series and textual information, and their implementations and result-reproduction code are provided as supplementary material.The experiments focus on next-day taxi pickups, using previous days' pickups, weather, and Web-derived event information.Source: §4 · p. 11, §4.1 · p. 11
4.1. Dataset and case studies
- The study analyzes taxi pickups within bounding boxes of ±0.003 decimal degrees (roughly 500 meters) around Barclays Center and Terminal 5, using 1.1 billion New York taxi trips from January 2009 to June 2016.Barclays Center event records came from its official website, while Terminal 5 records came from the Facebook API; both included event title, date, time, and description.Source: §4.1 · pp. 11–12
4.2. Experimental setup
- Models are evaluated on separate time-based train, validation, and test sets, with independent models for each area and incremental inputs from lags through event text.The compared inputs add weather, event presence, and event descriptions in sequence; SVR and Gaussian processes are also included as forecasting baselines.Source: §4.2 · pp. 13–14
4.3. Results
- At Barclays Center, event information produced the largest source-specific gain, and adding event text further improved DL-FC and DL-LSTM predictions.For DL-FC, event information reduced MAE by 20.8% and RMSE by 16.0%; text then reduced MAE by 2.5% and raised average R2 from 0.630 to 0.643. Text also reduced error on event and non-event days.Source: §4.3 · pp. 15–16, Table 2 · p. 15, Table 3 · p. 16
- At Terminal 5, LSTM performed best using lags alone, while adding event presence and text further reduced errors for the deep-learning models.For DL-FC, text reduced MAE from 168.1 to 152.6 and RMSE from 242.7 to 232.1; text also reduced errors on event and non-event days.Source: §4.3 · pp. 16–17, Table 4 · p. 16, Table 5 · p. 17
- MAPE fell from 15.6 to 12.3 at Barclays Center and from 19.6 to 16.1 at Terminal 5 when DL-FC L+W was extended to L+W+E+T.The comparisons report the full model against DL-FC L+W in both study areas.Source: Table 2 · p. 15, Table 4 · p. 16, §4.3 · p. 18
5. Conclusion
The paper proposes two deep-learning architectures that fuse time-series and unstructured text, and demonstrates reduced taxi-demand forecasting error in event areas. Future work would extend the approach to city-wide spatio-temporal models using information about events across the city.
5. Conclusion
- The two architectures combine temporal data with unstructured event text to improve time-series forecasts.They use word embeddings and convolutional layers for text, and model time series with either LSTMs or fully connected layers.Source: §5 · p. 18
- In taxi-demand forecasting for event areas, fusing New York taxi trips with event data mined from the Web significantly reduces forecasting error.The models also significantly outperform popular time-series methods that do not account for event information.Source: §5 · p. 18
- The authors make the code and datasets publicly available as a baseline for developing cross-modal data-fusion approaches.Source: §5 · p. 18
- Future work will explore city-wide spatio-temporal models that use event information, including titles and descriptions, to distinguish impactful from negligible events.Source: §5 · p. 18