Models Coupling Urban Growth and Transportation Network Growth : An Algorithmic Systematic Review Approach

Abstract: A broad bibliographical study suggests a scarcity of quantitative models of simulation integrating both network and urban growth. This absence may be due to diverging interests of concerned disciplines, resulting in a lack of communication. We propose to proceed to an algorithmic systematic review to give quantitative elements of answer to this question. A formal iterative algorithm to retrieve corpuses of references from initial keywords, based on text-mining, is developed and implemented. We study its convergence properties and do a sensitivity analysis. We then apply it on queries representative of the specific question, for which results tend to confirm the assumption of disciplines compartmentalisation.

• The paper introduces an Algorithmic Systematic Review (ASR) method using text-mining to investigate literature structure and finds low lexical proximity between research communities studying urban growth and transport networks separately.
• Applying the ASR, the study provides quantitative evidence supporting the hypothesis that disciplinary compartmentalisation hinders the development of models that endogenously couple urban land-use and transportation network growth.
• This quantitative evidence of disciplinary silos offers a plausible explanation for the scarcity of truly integrated co-evolutionary models and highlights the need for increased cross-disciplinary synthesis.

This paper investigates the documented paucity of quantitative simulation models that endogenously integrate the co-evolution of urban land-use patterns and transportation network infrastructure (1605.08888). The study posits that this deficiency may arise from disciplinary fragmentation, where research communities studying land-use dynamics and network growth operate with limited interaction. To quantitatively explore this hypothesis, the authors developed and deployed an Algorithmic Systematic Review (ASR) methodology predicated on text-mining techniques.

Research Motivation and Background

The interdependence and co-evolutionary dynamics between urban spatial structure (land use) and transportation networks are widely acknowledged phenomena in urban studies, geography, economics, and physics. Urban development patterns influence travel demand, which in turn drives transportation infrastructure expansion. Conversely, new or improved transportation links alter accessibility, shaping subsequent land-use changes and urban growth trajectories. Despite this recognized coupling, a preliminary review of the literature indicated a striking absence of models that simultaneously simulate both processes as endogenous variables.

Existing modeling paradigms often address only one side of the feedback loop:

• Land-Use Transportation Interaction (LUTI) Models: Predominantly employed in urban planning, these models typically simulate changes in land use (e.g., residential location choice, employment distribution) as a response to accessibility changes derived from a fixed or exogenously defined transportation network. They generally do not model the evolution of the network itself.
• Network Growth Models: Found more often in physics and economics literature, these models simulate the topological or functional evolution of transportation networks, often driven by simplified assumptions about the underlying urban form (e.g., population density fields) which is typically treated as static or exogenous.

While some hybrid models explicitly attempting to couple both dynamics exist, the paper notes they are relatively scarce. The central hypothesis explored is that this scarcity stems from disciplinary "compartmentalisation," where distinct research communities focus on either land use or network growth, employing different terminologies, methodologies, and publication venues, thereby hindering the development of integrated frameworks.

Algorithmic Systematic Review Methodology

To investigate the structure of the scientific literature surrounding this topic and test the compartmentalisation hypothesis, the authors developed an ASR approach. This method diverges from traditional bibliometric analyses based on citations or co-authorship, instead focusing on reconstructing the semantic landscape using text-mining.

The core of the methodology is an iterative algorithm:

1. Initialization: The process begins with a small, manually curated set of initial keywords ($K_0$) relevant to the research question.
2. Reference Retrieval: In iteration $i$, the current keyword set ($K_{i-1}$) is used to query a bibliographic database (specifically, the Mendeley API) to retrieve a set of relevant references ($R'_i$).
3. Corpus Update: The retrieved references are added to the cumulative corpus: $C_i = C_{i-1} \cup R'_i$.
4. Keyword Extraction: NLP techniques are applied to the abstracts and titles within the updated corpus $C_i$ to extract a new set of the $N_k$ most relevant keywords ($K_i$). $N_k$ is a key parameter of the algorithm.
5. Iteration/Termination: Steps 2-4 are repeated. The algorithm terminates when the corpus size stabilizes (i.e., few new relevant references are found) or a predefined maximum number of iterations is reached.

This iterative process aims to progressively refine the corpus and the associated keyword set, exploring the semantic space related to the initial query. The implementation utilized Java for the main algorithm structure, the Mendeley API for data retrieval, and a Python script leveraging the NLTK library for the NLP-based keyword extraction phase. The final output consists of corpuses in RIS format, suitable for further analysis.

Algorithm Behavior: Convergence and Sensitivity

The authors empirically evaluated the behavior of the ASR algorithm. Formal proof of convergence is challenging due to the complex interactions with the external database and the NLP procedures.

• Convergence: Empirical tests using various queries related to urban and network growth themes demonstrated that the algorithm generally converges effectively. The size of the corpus typically stabilized within approximately 10 iterations, indicating that the iterative refinement process reached a reasonably stable state.
• Sensitivity to $N_k$: The total number of references retrieved (final corpus size) was found to be highly sensitive to the parameter $N_k$ (the number of keywords extracted at each iteration). This sensitivity is considered logical, as varying $N_k$ effectively alters the breadth and focus of the semantic search. Smaller $N_k$ values tend to produce more focused, smaller corpuses, while larger values explore a wider semantic field, potentially retrieving more tangentially related documents.
• Lexical Consistency: The internal lexical consistency of the generated corpuses was also analyzed. As expected, corpuses generated with smaller $N_k$ values exhibited higher internal consistency (meaning the documents within the corpus shared more common terminology). While consistency decreased as $N_k$ increased (indicating broader topic coverage), the variability remained within acceptable bounds for the analysis.

Testing the Disciplinary Compartmentalisation Hypothesis

The primary application of the ASR was to test the hypothesis of disciplinary compartmentalisation. Five distinct initial keyword sets ($K_0$) were defined, each representing a different disciplinary perspective or common phrasing identified during the initial manual literature review (e.g., "land use transport interaction," "transportation network urban growth," "complex network spatial," "city system network," "urban morphogenesis accessibility").

The ASR algorithm was run independently for each initial keyword set, using a relatively large $N_k = 100$ to minimize constraints imposed by the algorithm itself and allow for broader exploration from each starting point. This resulted in five distinct final corpuses ($C_{final}$).

To quantify the degree of separation or overlap between these disciplinary viewpoints, the lexical proximity between each pair of final corpuses was computed. This proximity was likely based on the Jaccard index or a similar measure applied to the sets of extracted keywords characterizing each final corpus.

The results showed significantly low lexical proximity values between the corpuses generated from different initial disciplinary queries (detailed in Table 1 of the paper). For instance, the proximity between the "LUTI" corpus and the "Network Urban Growth" corpus was low, suggesting limited overlap in the terminology and, by extension, the core literature referenced by researchers starting from these different conceptual points.

Implications and Conclusion

The low lexical proximity observed across corpuses derived from different initial query sets provides quantitative support for the hypothesis of disciplinary compartmentalisation in the study of coupled urban and transportation network growth. The distinct terminologies and literature bases suggest that research communities focusing on land-use dynamics, network science perspectives, and traditional transport planning often operate in relative isolation.

This compartmentalisation likely acts as a significant barrier to the development and proliferation of integrated models capable of simulating the co-evolution of cities and their transport systems endogenously. The lack of shared vocabulary and cross-referencing hinders the synthesis of knowledge and methodologies required for such integrative modeling efforts.

The authors position this ASR approach as a complementary tool to qualitative reviews, providing a quantitative, data-driven perspective on the structure of scientific discourse. They suggest future work could involve constructing citation networks from the generated corpuses to further validate the findings, comparing the observed lexical modularity with structural modularity in the citation graph.

In conclusion, the paper introduces a novel algorithmic systematic review technique based on text-mining and applies it to investigate the structure of research on coupled urban and transport network growth. The findings strongly suggest disciplinary compartmentalisation, quantitatively evidenced by low lexical proximity between literature corpuses representing different facets of the field, offering a plausible explanation for the observed scarcity of truly integrated co-evolutionary models.