Journal Archive

We first collected existing academic papers related to the transportation decarbonization from the SCOPUS database. The search string process was based on a two-level assembly structure, with level one containing transportation-related keywords (e.g., “transportation” OR “transport”) and level two containing decarbonization- related keywords (e.g., “decarboniz(s)ation” OR “decoupling” OR “decarboni*” OR “carbon emission”). The publication period was set to indefinite for all related articles published by August 1, 2022. Only articles written in English were selected.

The initial set included 1,342 academic papers that matched at least one of the keywords either in the title, abstract, or keywords. The screening process excluded non-peer-reviewed publications (e.g., conference papers, letters, book chapters), insufficient papers (e.g., articles without abstracts or keywords), irrelevant papers (e.g., articles unrelated to transportation), and duplicated papers. After the screening process, a total of 779 articles were included in the final dataset.

We divided the final dataset into two regions to compare the transport decarbonization articles from Asian and non-Asian countries. To do this, we first divided the article (classification of the collaborative article was based on the first author's country) by country according to the author's affiliation and country table provided by SCOPUS. During this process, the final dataset contained 777 articles, except for two that missed country information.

As shown in <figure 2(a)>, the total number of publications has significantly increased from 1990 to 2022. Particularly, academic interest in transport decarbonization has grown sharply since 2015; it seems to positively reflect the Paris Agreement's establishment of the goal of achieving net zero emissions. Between 1990 and 2022, 65 different countries (Asia=21, non-Asia=44) published articles related to transportation decarbonization. The top five publishing countries were China (32%), the United States (19%), Germany (9%), the United Kingdom (7%), and Australia (5%) (see <Figure 2(b)>).

<Figure 2> 
Publications trends in research of transport decarbonization

Next, we classified the articles into five regions based on the United Nations (UN) region classification criteria: America (22%), Asia (53%), Europe (20%), Africa (3%), and Oceania (2%) (see <Figure 2(c)>). As a result of regional classification, articles from Asian countries accounted for 53% of the total articles compared to other regions, so we combined all other regions into non-Asian countries.

Finally, 410 articles (53%) from Asian countries and 367 articles (47%) from non-Asian countries were included in the final dataset. A comparison of publication trends in the two regions is shown in <Figure 2(d)>. Before 2009, non-Asian countries published more than Asian countries, but Asian publications increased dramatically after 2010 and began to dominate in 2015. Overall, the number of publications in Asia and non-Asian increased significantly between 1990 and 2022.

(1) Data pre-processing: The procedures for cleaning and preprocessing textual data using Natural Language Processing (NLP) are as follows. First, Tokenization process was used to split the textual data into tokens (simple units). Next, Parts of speech (POS) tagging was set to label each token with a tag (NN, VB, DT, etc.) based on the word type (noun, verb, adverb, etc.), and we extracted only nouns or noun phrases for analysis. Then, Stemming was set to process several words with similar meanings into a single word (e.g., transportation, transport > transportation). Next, Lemmization was set to that complex words (e.g., green, house, gas > greenhouse gas) were not extracted separately. Lastly, Stopwords was set to exclude unnecessary words (e.g., study, purpose) frequently occurring in the document corpus.

(2) Text mining: Following data pre-processing, we used the bag of words (BOWs) to extract the keywords with the high frequency of common occurrence in the documents of each region. Then, in preparation for topic modeling analysis, we generated a term-document matrix (TDM) and a document-term matrix (DTM). We also estimate the term frequency-inverse document frequency (TF-IDF) and word distance weights for weight settings. The TF-IDF is a combination of two indicators: term frequency and inverse document frequency, and is computed by multiplying the term frequency by the inverse document frequency (Bai et al., 2021). In the TF-IDF equation shown in <Figure 1>, TF(t,d) represents the term frequency of term t within document d, and IDF represents the inverse document frequency of term t.

(3) Topic modeling: The Latent Dirichlet Allocation (LDA) topic model was employed to reveal key latent topics and determine their knowledge structure in the literature corpus by each region. LDA is a probabilistic-based classification technique that extracts content from unsupervised data (Boyer et al., 2017). That is, it is an appropriate method for exploring unobserved information and identifying latent themes and their semantic structures in large volumes of textual data (Bayarsaikhan et al., 2022; Cao et al., 2009).

For inference of the LDA, we set the TF-IDF threshold (λ) = 0.05 (mean value), iterations = 1000, and hyperparameters alpha (α) = 0.1 and beta (β) = 0.01. Both α and β are parameters for determining the topic modeling probability (Griffiths and Steyvers, 2004). In order to determine the optimal number of topics and to derive the probability distribution of the terms in each topic, we used the perplexity of the Gibbs sampling algorithm (Griffiths and Steyvers, 2004). The perplexity of Gibbs Sampling is a sampling method from a joint distribution when conditional distributions of topics and words can be computed efficiently (Boyer et al., 2017). The generating process of LDA is presented in <Figure 1>.

(4) Machine learning: The K-Nearest Neighbor (KNN) algorithm was utilized to evaluate the accuracy of latent topics extracted from LDA and to predict the fit of the topic classifier by supervised data. In the process of KNN, we first applied each dataset from Asia and non-Asian countries by dividing them into training datasets (i.e., classified data) and test datasets (i.e., non-classified data). We set training data to 80% and text data to 20% for our machine-learning experiments. That is, the standard for the training-to-test ratio is 80% to 20%, as it is more effective when the training dataset ratio is higher than the test dataset ratio (Casillo et al., 2021).

We also applied the Euclidean distance parameter to calculate document-topic level similarity. The Euclidean distance indicates a higher degree of similarity between the documents based on their distance value (Gallego et al., 2018). We set the topic similarity threshold (λ) = 0.05, iterations = 10, and the number of topics in the trained model of each region to the same as in the LDA result. In KNN, the confusion matrix can present the proportion of topics between the predicted and actual as well as the correct and incorrect classification of the classified topics (Casillo et al., 2021). Therefore, the confusion matrix is used to check the accuracy of the classified topics based on the document-topic matrix generated by LDA. The standard parameters AUC–ROC, classification accuracy (CA), F-criteria (F1), precision, and recall (Casillo et al., 2021) are applied to evaluate the accuracy of classified topics.

As a result of the perplexity of the Gibbs Sampling iterations by setting the number of topics in the range of 2 to 15, the most optimal number of topics in Asia was six (perplexity log=66.735), while the most suitable number of topics in non-Asia was five (perplexity log=57.797). As shown in <Table 1> and <Figure 3>, six topics were derived from the Asian dataset.

<Figure 3> 
Latent topic structure of Asia

Topic 1 accounted for about 16% of the total dataset and includes core terms (i.e., sub-theme) such as “effect,” “carbon neutrality,” “infrastructure,” and “urban.” Among these core terms, the topic probability value of the term “effect” is the highest, indicating that this includes research on the impact of carbon emissions in the transportation sector. Thus, the label of topic 1 was set as “drivers of transport CO2 emissions.” Topic 2 indicated the second most covered topic in Asian literature, which included approximately 18% of all papers. Topic 2 consists of core terms such as “energy saving,” “reduction,” “urban,” “strategy,” and “public transport,” indicating that it mainly deals with topics related to reducing transportation carbon emissions. Topic 3 accounted for 14% of the total dataset, and the topic probability of terms such as “freight,” “road,” “trade,” “cost,” and “material flow” was the highest.

Therefore, it can be inferred that this topic deals with decarbonization in terms of road freight transport. Topic 4 accounted for 13%, the lowest proportion compared to other topics, which focused on electric vehicles. It shows that in this topic, key terms such as “fuel consumption”, “Greenhouse Gas (GHG),” “energy use,” and “mitigation” are connected to electric vehicles. Topic 5 accounted for about 15%, and the topic probability value of “economic growth,” “decoupling,” “industry,” and “climate change” terms was the highest. Finally, topic 6 accounted for roughly 18% of the dataset, indicating that it is the most discussed topic compared to other topics. This topic consists of major related terms such as “energy use,” “transportation system,” “policy,” and “conservation” and includes studies dealing with energy consumption in the transport sector.

Meanwhile, the optimal number of topics derived from non-Asian datasets is five (see <Table 2> and <Figure 4>). Topic 1 accounted for about 16% of the total dataset and includes core terms such as “road,” “energy efficiency,” “climate change,” and “passenger.” The topic probability value of the term “road” is the highest, indicating that includes research related to road transport decarbonization. Topic 2 indicated the second most covered topic in Asian literature, which included approximately 24%. Topic 2 consists of core terms such as “GHG,” “alternative fuels,” and “technology;” it implies that it mainly deals with topics related to energy consumption in the transport sector.

<Figure 4> 
Latent topic structure of non-Asia

Topic 3 accounted for 11% of the total dataset, the lowest proportion compared to other topics, focused on reducing transport CO2. The topic probability of terms such as reduction, “Environmental Kuznets Curve (EKC),” “decoupling,” “mode choice,” “energy efficiency,” “passenger,” and “economic growth” was the highest. Topic 4 accounted for 22%, and the topic probability of terms such as “fuel consumption,” “transportation system,” “gas,” “public transport,” “hydrogen,” and “oil” was the highest. It shows that this topic deals with fuel consumption issues in transport decarbonization. Finally, topic 5 accounted for about 25% of the dataset, which was the most covered topic compared to the other topics. This topic addressed electric vehicles related to the terms “carbon capture and storage (CCS),” “pollution,” and “supply.”

This section shows the analysis results of evaluating potential topics extracted from LDA and predicting the suitability of the proposed topic model as follows.

1) Distribution of the document-topic probability

For the analysis, the Asian dataset (n=410, x²=1892.26, p=0.000) was divided into six categories, whereas the non-Asian dataset (n=367, x²=959.53, p=0.000) was divided into five categories. The Euclidean distance scores range from 0 to 1, with lower values indicating high accuracy of the classification. The smallest Euclidean distance value indicates the closest distance between documents (Gallego et al., 2021). That is, the lowest score means that document similarity is high and can be grouped into one topic. Thus, the document-topic distribution of the two regions measured by Euclidean distance was at an appropriate level.¹⁾

<Figure 5> shows the distribution of documents by topic extracted from the LDA. In Asia, similarities between documents accounted for the highest proportion of topics as “decoupling of transport carbon emissions,” “freight road transport,” and “energy use of transport” in the order of 89%, 82%, and 76%, respectively. On the other hand, the similarity between documents distributed in the rest topics of “transport CO₂ drivers,” “transport CO₂ reduction,” and “electric vehicles” was slightly lower, which seems to be mixed with documents from all other topics. In non-Asia, similarities between documents were high in topics of “fuel consumption,” “road transport,” and “transportation energy use,” and the distribution of documents by topic accounted for the highest proportion in the order of 82%, 79%, and 72%, respectively.

<Figure 5> 
The distribution of documents similarity in the topic

By contrast, CO₂ reduction and electric vehicle topics were mixed in with others, as were Asian topics. The similarity between the documents distributed on the topics of “reduction of transport CO₂” and “electric vehicles” was slightly lower, and shows that they are mixed with documents of others similar to the results of the Asian topic distribution. Therefore, it can be concluded that these topics, which consist of mixed-documents, appear to be dominated by articles dealing with incorporated combination research, or are areas of research that have not yet been sufficiently addressed as core topics in this field.

2) Evaluation of the accuracy of classified topics extracted from LDA.

We evaluated the accuracy of classified topics extracted from LDA using standard parameters of machine learning algorithms. The score range of performance parameters is 0 to 1, which means that the closer the value is to 1, the higher the classification accuracy (Casillo et al., 2021). <Table 3> and <Figure 6> present the classification evaluation results for the proposed topic model by LDA. The AUC was 0.925 in Asia and 0.892 in non-Asia.

<Table 3> 
The classification evaluation on the proposed topic model

KNN model	N	AUC	CA	F1	Precision	Recall
Asia	410	0.925	0.888	0.884	0.889	0.888
Non-Asia	367	0.892	0.660	0.608	0.743	0.660

<Figure 6> 
The confusion matrix of KNN

The classification accuracy (CA) was 0.888 in Asia and 0.660 in non-Asian regions. The F1, precision, and recall values also achieved high accuracy for the topic classification of each region. In other words, these results indicated that the topic classifications of both Asian and non-Asian datasets were accurately classified. Thus, it implies that the proposed model with latent topics can significantly improve the performance of unsupervised textual data classification.

The confusion matrix can present the proportion of topics between the predicted and actual as well as the correct and incorrect classification of the classified topics (Kulkarni et al., 2020). Therefore, the confusion matrix is used to observe which specific topics were misclassified and/or how they were classified. <Figure 6> presents the confusion matrix obtained from each dataset (i.e., Asia = 6×6, non-Asia = 5×5).

In the confusion matrix, the rows represent the actual numbers of the document samples, while the columns reflect the estimates of the model. The sum of the total probabilities of each document shows a slight difference between the actual and predicted. The purple diagonal also displays the correct classification ratio, and the rest shows the ratio of incorrect classification displayed in pink. Among Asian topics, 64%~93% were correctly classified, whereas 57%~89% of non-Asian topics were correctly classified. Regarding misclassification, however, Topic 1 in Asia and Topic 2 in non-Asia had slightly higher error classifications.

3) Histogram results for document-topic probability

<Figure 7> depicts the histogram results of the document-topic probability. Asian document flows show that topics related to transportation decarbonization were primarily addressed in terms of energy in the transportation sector between 1990 and 2000. Then, in the early 2000s, investigations about reducing transportation carbon emissions began. Since 2010, transportation decarbonization issues have been debated in research areas such as “transportation CO₂ emission drivers,” “transportation decoupling,” and “electric vehicles.”

<Figure 7> 
Trend histograms for document-topic probabilities

By contrast, non-Asian literature has discussed transport decarbonization topics in “fuel,” “road transport,” and “transport energy use” research since 1990, but the probability of distributing this body of literature has decreased since 2008. It implies that this is due to the emergence of new research areas rather than a decrease in relevant research topics. Since the mid-2000s, there has been an increase in both Asian and non-Asian literature on “transport CO₂ reduction” and “electric vehicles.” Notably, the topic of transportation decarbonization has been overgrown in the non-Asian electric vehicle-related literature since 2008. It may reflect the rapid expansion of international academic interest in this transportation decarbonization research area.