CoChat
Sign Up
Cited by 12Traffic Prediction and Management TechniquesTransportation Planning and OptimizationUrban Transport Systems Analysis

BusTime: Which is the Right Prediction Model for My Bus Arrival Time?

Open access
Liu, Dairui, Sun, Jingxiang, Wang, Shen·2020·2020 5th IEEE International Conference on Big Data Analytics (ICBDA)
DOI 10.1109/icbda49040.2020.9101265
Open this paper in CoChat — free accountOpen PDF

Abstract

With the rise of big data technologies, many smart transportation applications have been rapidly developed in recent years including bus arrival time predictions. This type of applications help passengers to plan trips more efficiently without wasting unpredictable amount of waiting time at bus stops. Many studies focus on improving the prediction accuracy of various machine learning and statistical models, while much less work demonstrate their applicability of being deployed and used in realistic urban settings. This paper tries to fill this gap by proposing a general and practical evaluation framework for analysing various widely used prediction models (i.e. delay, k-nearest-neighbor, kernel regression, additive model, and recurrent neural network using long short term memory) for bus arrival time. In particular, this framework contains a raw bus GPS data pre-processing method that needs much less number of input data points while still maintain satisfactory prediction results. This pre-processing method enables various models to predict arrival time at bus stops only, by using a kd-tree based nearest point search method. Based on this framework, using raw bus GPS dataset in different scales from the city of Dublin, Ireland, we also present preliminary results for city managers by analysing the practical strengths and weaknesses in both training and predicting stages of commonly used prediction models.

AI snapshot

Written from the paper’s full text.

TL;DR

This paper proposes a practical evaluation framework for bus arrival-time prediction, centered on a stop-based GPS interpolation method that predicts only at bus stops using kd-tree nearest-point search. Using Dublin bus line 46A data, it compares delay, k-NN, kernel regression, additive, and RNN-LSTM models in training cost, prediction cost, and accuracy.

Key findings

  • The stop-based interpolation reduces computation versus conventional distance-based interpolation, with training time up to about 5 times less. Under 4311 trips, k-NN prediction drops from about 1 minute to less than 15 seconds, and KR becomes feasible under full data (less than 1.5 minutes) whereas distance-based KR exceeds 1.5 minutes beyond 1500 trips.
  • Accuracy ranking is LSTM > KR > k-NN > Delay > Additive. LSTM achieves the highest accuracy, but a larger dataset does not guarantee higher LSTM accuracy because it requires hyperparameter tuning.
  • Efficiency rankings are: training Delay > KR > Additive > k-NN > LSTM; predicting Additive > Delay > k-NN > LSTM > KR.
  • The stop-based method does not compromise much accuracy and can even improve it in some cases, e.g. 1000 trips using LSTM. Interpolation deviation from nearest GPS point to bus stop averages 33.53 m (std 111.72 m), with 99% under 244 m, corresponding to less than 30 seconds travel-time error at 30 km/h.

Methods

  • Data: Dublin, Ireland bus line 46A outbound (Phoenix Park to Dun Laoghaire, nearly 20 km), using GTFS schedules and bus GPS data from November 2012 and January 2013 with about 30-second update frequency. After preprocessing, 4311 trips were used.
  • Preprocessing: trip segmentation with coarse filtering, de-duplication, splitting by 900-second gaps, direction judgment using first 30 GPS points and cosine similarity, and completeness checks within 300 m and half of predefined length/duration. Then stop-based interpolation selects the nearest GPS point to each bus stop via kd-tree, yielding 59 points per trip versus 191 points per trip for distance-based interpolation every 100 m.
  • Evaluation: five models—delay, k-NN, KR, additive (basic additive model), and RNN-LSTM. Trips were sorted chronologically and divided into 9 groups of first 500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, and 4311 trips, with 80% training and 20% testing. Metrics were MAE, MAPE, and RMSE. LSTM was tested only on 1000 and 3000 trips, implemented in TensorFlow 2.0 with Adam optimizer and learning rate 0.01.

Limitations

  • The paper states that the additive model is only a basic version due to limitations of the Python pyGAM library, and it uses a new Python library rather than R as in prior literature.
  • The paper states that LSTM has limited data accessibility, uses only trip departure time, travel time, travel distance, and distance to the next point, and was not tested on GPU/TPU machines. It was tested only on 1000 and 3000 trips.
  • The paper states the results are preliminary and based on one bus line (46A outbound) in Dublin, so generalization to other routes or cities is not established. Evident from the text: no weather, hours-of-day, or other features are included for most models, and hyperparameter tuning time for LSTM is not included in training-time comparisons.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/icbda49040.2020.9101265

Cites

11 of 11 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 10 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    The proposed raw bus GPS data pre-processing method requires much fewer input data points than conventional methods while still maintaining satisfactory prediction results.

    “In particular, this framework contains a raw bus GPS data pre-processing method that needs much less number of input data points while still maintain satisfactory prediction results.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The proposed stop-based trip interpolation method reduces training computation time compared with the distance-based trip method, by up to about 5 times less.

    “The results also show that with the increasing size of input data, our”
    Quote in the paperNot a substantial claim
  • Findingnone given

    The proposed pre-processing method enables various models to predict arrival time at bus stops only, using a kd-tree based nearest point search method.

    “This pre-processing method enables various models to predict arrival time at bus stops only, by using a kd-tree based nearest point search method.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    The proposed efficient bus GPS data pre-processing method enables the comparison of many widely used prediction models.

    “An efficient bus GPS data pre-processing method that enables the comparison of many widely used prediction models.”
    Quote in the paperSubstantial claim
  • Findingown data

    The prediction computation cost of KR increases sharply with the growing size of input data.

    “The prediction computation cost of KR increases sharply with the growing size of input data.”
    Quote in the paperSubstantial claim
  • May combine two claims. 19
  • Classified as probably not a substantial claim. 15
  • The quote only partly states this claim. 2

  • May combine two claims.

    Findingnone given

    The delay model requires the least computation cost and can still give acceptable prediction results.

    “Delay: A simple and efficient model to start with.It requires the least computation cost and still can give acceptable prediction results.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Bus line 46A is the most frequent bus line in the dataset and is a long bus trip of nearly 20 km that gives enough points to study.

    “We pick bus line 46A (outbound, from Phoenix Park to Dun Laoghaire, also used in the literature [3]) to study as it is the most frequent bus line in our dataset and it is a long bus trip (i.e.nearly 20km) that gives us enough points to study.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingnone given

    k-NN can generate better prediction results than the delay model, but it requires hand tuning for parameter k.

    “k-NN: It can generate better prediction results than delay model, but it requires hand tuning for parameter k.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    After pre-processing, 4311 trips were selected for evaluation.

    “After pre-processing, we have selected 4311 trips that we can use for evaluation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The bus GPS data contains about two months of real-time bus status, with timestamps at an update frequency of about 30 seconds.

    “Bus GPS data contains about two months (i.e.November 2012 and January 2013) real-time status of buses such as timestamps (i.e.update frequency about 30 seconds), location, and relevant bus trip information.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    The prediction accuracy and computation performance of the five models in training and predicting stages vary under different big data sizes.

    “to demonstrate how their prediction accuracy and computation performance in training and predicting stages varies under different big data size.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given

    The stop-based trip method does not compromise a lot of accuracy loss, and in some cases it can even improve accuracy, for example with 1000 trips using LSTM.

    “Our”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The delay model requires the least amount of training time, 2 to 4 times faster than KR and at least 10 thousand times faster than the additive model and LSTM.

    “As shown in Table II, the delay model requires the least amount of training time, which is 2 to 4 times faster than KR, and at least 10 thousand times faster than additive model and LSTM shown in Table III.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given
    § BusTime: Which is the Right Prediction Model for My Bus Arrival Time?

    The distance-based trip contains 191 points per trip, while the stop-based trip includes 59 points per trip.

    “The”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    KR is very sensitive to big data: the distance-based trip method makes it almost infeasible to use when input data size is greater than 1500, while the stop-based method makes it feasible under full data settings with less than 1.5 minutes to predict.

    “KR is very sensitive to the big data.Traditional”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Under the full data setting of 4311 trips, k-NN needs about 1 minute for prediction, making it less practical.

    “Under the full data setting 4311 trips, k-NN needs about 1 minute, which makes it less practical”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Under the full data setting of 4311 trips, the stop-based trip method reduces k-NN prediction time to less than 15 seconds, whereas k-NN with distance-based trips needs about 1 minute.

    “Under the full data setting 4311 trips, k-NN needs about 1 minute, which makes it less practical, but with our”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    KR can provide very good prediction results in all accuracy metrics and can be trained very fast.

    “KR: It can provide very good prediction results in all accuracy metrics and it can be trained very fast.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    The evaluation uses a realistic dataset at different scales from Dublin, Ireland, to demonstrate how prediction accuracy and computation performance in training and predicting stages vary under different big data sizes.

    “To make this evaluation practical for city managers, we use a realistic dataset at different scales from the city of Dublin, Ireland, to demonstrate how their prediction accuracy and computation performance in training and predicting stages varies under different big data size.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    GPS data pre-processing has well-known challenges that the proposed method needs to deal with.

    “There are some well-known challenges in GPS data pre-processing [7] that our method needs to deal with.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [7]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    k-NN can provide predicted results with mean absolute error less than about 3 minutes when the bus travelled at least around a quarter of its whole journey.

    “This early research show that k-NN can provide predicted results with the mean absolute error less than about 3 minutes when the bus travelled at least around a quarter of its whole journey.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [2]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    Raw GPS data is often noisy and has irregular update frequency.

    “due to the fact that the raw GPS data is often noisy and have irregular update frequency”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    The additive model's poor prediction accuracy might be because only a simple version was implemented and a new Python library was used rather than R used in the literature.

    “The accuracy of additive model prediction is the worst but it might because that we only implemented a simple version, and we use a new Python library rather than R used in literature.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The proposed stop-based trip interpolation method outperforms the conventional distance-based trip method by significantly reducing training and prediction computational cost while maintaining satisfactory prediction accuracy.

    “Besides, we can summarize that our proposed”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    RNN-LSTM captures long-term dependencies of bus arrival times at different location points, which further improves prediction results.

    “Last year, Pang et al. [5] adopted recurrent neural network with long short-term memory block (RNN-LSTM) to capture long-term dependencies of bus arrival times at different location points, which further improves the prediction results.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [5]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisereasoning

    A performance analysis study using big data is missing for practitioners deploying bus arrival time prediction models into realistic urban scenarios, considering the trade-off between accuracy and computation/storage requirement.

    “a performance analysis study using big data is missing for practitioners deploying them into realistic urban scenarios, by considering the trade-off between their accuracy and computation/storage requirement.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    State-of-the-art bus arrival time predictions have become more accurate and stable.

    “Although the state-of-the-art bus arrival time predictions have been achieved more accurate and stable results”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    A larger dataset does not guarantee a higher accuracy for LSTM because it requires deep learning experts to tune many hyper-parameters for a given dataset.

    “But a larger dataset does not guarantee a higher accuracy for LSTM as it requires deep learning experts working on tuning many hyper-parameters for a given dataset.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Kernel regression improves k-NN by automatically selecting the proper number of most similar past trajectories for predictions, resulting in much more stable arrival time estimations.

    “A year later, the kernel regression (KR) [3] model is applied to improve k-NN by automatically selecting the proper number of most similar past trajectories for predictions, which have resulted in much stable arrival time estimations.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Additive models are a more flexible tool that can combine multiple features and do not need GPS point interpolations to align different trips by fixed travelled distance.

    “Kormáksson et al. [4] proposed to use additive models, a more flexible tool that can combine multiple features and does not need GPS points interpolations to align different trips by fixed travelled distance.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [4]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    RNN-LSTM is a very accurate prediction model but requires deep learning experts for tuning hyper-parameters and a large amount of data to train.

    “RNN-LSTM: Very accurate prediction model.But it requires deep learning experts for tuning hyper-parameters and large amount of data to train.”
    Quote in the paperSubstantial claim
26 of 26

  • The authors state this as an assumption.

    Bus passengers only care about arrival times at bus stops, not about interpolated points at locations with a fixed distance gap.

    “we propose a stop-based trip interpolation method based on a practical view that bus passengers only care about arrival times at bus stops, rather than many interpolated points at locations with a fixed distance gap.”

  • The paper relies on this without support.

    City managers or engineers can only use historical data for training models to predict using real-time bus location updates, and models are re-trained after a certain time, normally few weeks or months.

    “In practice, city managers or engineers can only use historical data for training models to predict using real-time bus location updates, and these models will be re-trained after a certain time (normally, few weeks or months, as opposed to few seconds or minutes) using new data.”

  • The paper relies on this without support.

    The raw GPS data fields vehicle and vehicle journey are normally not reliable.

    “Although these data fields are normally not reliable, we can use them to coarsely narrow down the range of a bus trip to be filtered.”

  • The paper relies on this without support.

    In the realistic city scenario, prediction demand occurs every second for many bus trips for the city servers where the prediction is deployed.

    “As in the realistic city scenario, prediction demand occurs every second for many bus trips for the city servers where the prediction is deployed, the prediction service running time presented here is a aggregated one which have included all testing trips per testing group at all stops, rather than the average prediction time per trip per stop or point.”
4 of 4

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Reported RMSE is lower than MAE

    Major

    For the 500-trip delay model, Table V reports MAE values of 238.23 and 220.31, while Table VI reports smaller RMSE values of 219.94 and 213.78 for the same two trip methods. RMSE cannot be lower than MAE when both are calculated from the same prediction errors; this pattern also appears in other delay, k-NN, and KR rows.

  • Ten-thousand-fold training-speed claim fails at 500 trips

    Minor

    For 500 stop-based trips, Table II gives delay training time as 2.17 × 10−4 seconds and additive training time as 1.35 seconds. That makes delay about 6,200 times faster than additive, not at least 10,000 times faster.

2 of 2

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Liu, D., Sun, J., & Wang, S.
Affiliations
University College Dublin (IE)
ORCID
Wang, Shen
Venue
2020 5th IEEE International Conference on Big Data Analytics (ICBDA)
Publisher
IEEE
Published
2020-05
Pages
180-185
DOI
10.1109/icbda49040.2020.9101265
OpenAlex
W3031435341
Semantic Scholar
1ce611630ca02c980b0065c89a20153b237dd861
Cited by
11
References
11
Citations by year
2026: 1 · 2024: 3 · 2022: 3 · 2021: 4
FWCI
2.89
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE International Conference on Big Data Analytics (ICBDA) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Engineering › Building and Construction › Traffic Prediction and Management Techniques
  • Social Sciences › Transportation › Transportation Planning and Optimization
  • Physical Sciences › Engineering › Civil and Structural Engineering › Urban Transport Systems Analysis
Keywords
bus arrival time prediction, public transport, k-nearest neighbors, kernel regression, additive models, LSTM, model comparison, urban transport, Dublin

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-11-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-11-05.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review