A Model Architecture for Public Transport Networks Using a Combination of a Recurrent Neural Network Encoder Library and a Attention Mechanism

Round 1

Reviewer 1 Report

Paper: A Model Architecture for public transport networks Using a combination of a Recurrent Neural Network Encoder Library and a Attention Mechanism was written on the basis of research. The authors conducted a literature review, although the number of bibliographic entries is small - only 23. Other works on urban transport performance in cities can be added here. 

Research problem presented correctly, outline of research methodology also, Conclusions in the paper are divided into two parts: future and conclusion. This is a pilot study that should be continued. 

After analyzing the paper, concludes that the presented paper is interesting. It is addressed to scientists looking for inspiration in further scientific research. The conclusions of the paper can deliver some information for business . So, the main scientific purpose has been fulfilled, as a new innovative contribution to the literature on the topic, as well as an application purpose for urban transport stakeholders. The conclusions relate to the research problem identified at the beginning of the article. I conclude that the work is correctly written, in terms of language and content. It provides an original point of view on the problem of networks.

Author Response

Thank you for your comments regarding the adding some additional references regarding public transport related publications. 

A paragraph has been added to the introduction expanding on the different methods used in public transport predictions. The changed paragraph can be found at line 53-65 and reads as follows:

The literature contains a wide range of approaches to predict bus ETAs. These range from more conventional methods such as historic averages [ 8 ,9], ARIMA [10] or Kalman 54
Filters [ 11 –16 ]. In general, such methods have low predictive power, and the introduction of Neural Networks (NN) drastically improved the performance of ETA predictions [ 14 –16 ]. 56
In the more recent literature, NN based approaches have taken center stage with some impressive results [ 17], however further improvements compared to NNs were achieved using hierarchical NNs [18 ]. A particular focus can be found on RNN structures due to the sequential nature of ETA prediction problems. These methods include Long Short Term Memory (LSTM) networks [19], bidirectional LSTMs [20 ] or even convolutional LSTMs [21]. However, much more complex methods have become more common and tend to use different types of models for different aspects of the prediction task [ 22 – 24 ]. As there is 
no limit to the complexity of an ETA model somewhat more exotic methods, such as the artificial bee colony algorithm are also represented in the recent literature [25].

Reviewer 2 Report

The paper presents a study on the architecture of a model to use the traffic state, the working concept of the entire transport network, to predict the estimated time of arrival (ETA) and the location of the next step. For this purpose, a dynamically changing recurrent neural network (RNN) based coding library is used. To achieve this, an attention mechanism was used that includes the states of other vehicles in the network by encoding their positions in their individual current states. According to the available information, their impact, expected accuracy can be evaluated in the subgroup according to the available data. The results of the experimental study show that the entire model with access to all network data performed better. However, the model is limited to model vehicles, and the same line before the target was the best-performing model, suggesting that the inclusion of additional data may have a negative impact on the prediction accuracy if it does not add any useful information. This may be due to poor data quality, but also to interactions between the included line and the target line. The technical aspects of this study are complex, so the results were very inefficient. We point out several areas where the method we present needs to be improved in order to make it a viable alternative to currently used methods. In conclusion, this study should be viewed as potentially useful for future development of the method architecture. Thus, this is a stepping stone for future research to improve public transport predictions if network operators have access to high-quality datasets. The article is characterized by a clear structure, a logical teaching sequence, the latest suitable literature sources are used for it, the illustrations are understandable and informative, the solution algorithm is presented, the further direction of research is indicated, in my opinion, the conclusions should be summarized a little more broadly.

Author Response

Thank you for your constructive comment. As suggested the conclusion was summarised more broadly and now reads as follows (550-565): 

To summarise, we have demonstrated the potential benefit of a novel model architecture using network-wide data to make predictions in public transport networks. The method developed is based on a library of GRU models for each individual line within a network to embed the temporal information of the individual vehicles. The combination of this embedding method with a transformer model allows the presented method to expand its information range to other vehicles within a bus network. The results are especially promising if vehicles ahead of the target vehicle are included. As this is a pilot study and was designed to test whether the described architecture is a promising area of research, more work is required to compare this architecture to more conventional methods. We have also highlighted several areas where improvements to our presented method are required to make it a viable alternative to current methods. As such, it is a stepping stone for future research to improve public transport predictions if network operators provide high-quality datasets. Furthermore, the developed method could also be used to simulate specific traffic scenarios, such as delays of specific vehicles and the effect on the rest of the bus network. Therefore, the presented method could not only result in better predictions, but could be a valuable simulation tool for network operators.