CityHood: Quantifying Urban Identity

• CityHood is a quantitative framework that defines urban identity through measurable indicators such as street patterns, mobility flows, and cultural symbols.
• It integrates structural, behavioral, and perceptual methodologies to differentiate cities and neighborhoods in a reproducible manner.
• The approach informs smart urban planning and recommendation systems by linking multimodal data with actionable insights on urban distinctiveness.

CityHood denotes a family of quantitative approaches for characterizing the identity of cities and neighborhoods from observable traces. In the cited literature, this identity is read from street-block morphology, multiscale urban hierarchy, soundscapes, mobility networks, service-request composition, street names, street-view perceptions, and user-facing recommendation interfaces. Taken together, these works suggest that CityHood is not a single ontology but a research program: the attempt to operationalize what makes urban places structurally distinct, behaviorally legible, symbolically meaningful, and computationally actionable (Louf et al., 2014, Luo et al., 2024, Bancilhon et al., 2021, Elizalde et al., 2016, Wang et al., 2016, Santos et al., 24 Jul 2025).

1. Conceptual scope and definitions

CityHood is used at multiple spatial and analytical scales. In some work it refers to the identity of whole cities, as in audio-only city identification from Flickr videos or quantitative typologies of street patterns; in other work it refers to neighborhoods, overlapping communities, or ZIP-code-level recommendation targets (Elizalde et al., 2016, Louf et al., 2014, Santos et al., 24 Jul 2025). A common feature is that identity is treated as measurable rather than purely descriptive.

One major strand defines CityHood through structural identity. Louf and Barthelemy frame it as the distinctive geometry of a city’s street network, captured by block area and shape rather than ambiguous street entities (Louf et al., 2014). The dominance-tree approach similarly treats a city’s role as a multiscale dominance relation among spatially distributed settlements, where importance depends jointly on population and spatial context (Louail et al., 2022). The “Nature of Human Settlement” work pushes this further toward a global morphology space, where cities are mixtures of neighborhood types derived from standardized block and color features across 1,667 cities (Nice et al., 2019).

A second strand defines CityHood through activity and interaction. In overlapping-community analysis, a place belongs simultaneously to multiple latent communities inferred from mobility flows, formalized through affiliation vectors $F_i$ and edge probabilities $P(m,k)=1-\exp(-F_m \cdot F_k^T)$ (Luo et al., 2024). In comparative mobile-traffic work, a city is a spatio-temporal pattern of calls, SMS, and data requests, summarized as normalized weekly signatures (Grauwin et al., 2014). Twitter-based and 311-based studies likewise construct location signatures from temporal tweeting patterns or complaint compositions, then use them to classify neighborhoods and model socioeconomic structure (Kats et al., 2017, Wang et al., 2016).

A third strand treats CityHood as symbolic or perceptual identity. Streetonomics interprets honorific street names as a quantitative record of commemorative values, gender bias, professional prestige, and cosmopolitanism (Bancilhon et al., 2021). Street-view perception modeling reconstructs within-city landscapes of perceived walkability and neighborhood quality from image pair judgments (Muller et al., 2022). Audio-based city identification argues that “cityhood” can be partially encoded in soundscape composition alone, without images, tags, or geo-tags (Elizalde et al., 2016).

This diversity corrects a common simplification: CityHood is not reducible to administrative boundaries or a single sensing modality. The literature instead suggests that urban identity is multi-layered, with different methods exposing different components of the same phenomenon.

2. Morphological and hierarchical CityHood

In street-pattern research, the basic units of city identity are blocks, defined as polygonal faces enclosed by roads. Louf and Barthelemy show that block-area distributions $P(A)$ are broadly similar across cities, with an approximate power-law tail $P(A)\sim 1/A^\tau$ and $\tau \approx 2$, so area alone is weakly discriminative (Louf et al., 2014). Their core contribution is a city “fingerprint” built from the conditional distribution of the shape factor $\Phi$ across area bins, effectively $P(\Phi \mid A)$. Using hierarchical clustering on fingerprints from 131 cities, they identify four large families—Buenos Aires–type, Athens–type, New Orleans–type, and Mogadishu–type—and recover a quantitative separation between most European and North American cities within the largest family. They also show that New York’s boroughs have distinct fingerprints and that a city fingerprint can be seen as the sum of neighborhood-level ones.

The “Nature of Human Settlement” work develops a different morphology pipeline at neighborhood scale. Each neighborhood is represented by a 320×320 pixel Google Static Map rescaled to 335 m × 335 m on the ground, from which block-size histograms, block-irregularity histograms, and color fractions are extracted into a 35-dimensional feature vector (Nice et al., 2019). Across almost 1.7 million tiles from 1,667 cities, these vectors are organized by a 100×100 self-organising map. City fingerprints are then kernel densities over SOM space. This representation is not merely descriptive: typology-weighted city averages correlate with moving vehicles fraction ($r=0.97$), impervious surfaces ($r=0.86$), sky fraction ($r=0.75$), building fraction ($P(m,k)=1-\exp(-F_m \cdot F_k^T)$0), mean AOD ($P(m,k)=1-\exp(-F_m \cdot F_k^T)$1), and mean NO$P(m,k)=1-\exp(-F_m \cdot F_k^T)$2 ($P(m,k)=1-\exp(-F_m \cdot F_k^T)$3) (Nice et al., 2019). A plausible implication is that CityHood, in a morphological sense, is predictive of environmental and transport performance rather than being a purely visual classification.

A third structural formulation comes from the dominance tree of urban systems. Here cities are points with marks $P(m,k)=1-\exp(-F_m \cdot F_k^T)$4 given by population, Voronoi neighbors define local competition, and recursive elimination of non-maxima yields a rooted tree whose node height encodes multiscale dominance (Louail et al., 2022). Height is not equivalent to size rank: the empirical Kendall $P(m,k)=1-\exp(-F_m \cdot F_k^T)$5 between population rank and height is small in France and the US, and the authors report that height is slightly less sensitive than rank to different statistical definitions of cities, with $P(m,k)=1-\exp(-F_m \cdot F_k^T)$6 versus $P(m,k)=1-\exp(-F_m \cdot F_k^T)$7 for French municipalities under morphological and functional delineations (Louail et al., 2022). CityHood in this formulation is relational and hierarchical: a city’s identity includes the basin of smaller cities it dominates and the scales at which it remains a local maximum.

Across these studies, morphology is not limited to map appearance. It is formalized through block geometry, typological mixtures, and recursive dominance relations, yielding a structural account of CityHood that is comparable across cities and across scales.

3. Behavioral, mobility, and networked CityHood

Mobility-based work shifts CityHood from form to interaction structure. In the Twin Cities overlapping-community study, the urban system is a graph whose nodes are 2,591 census block groups and whose edges are trip counts aggregated from 10.1 million PlaceIQ records from 166,850 devices over 7 days in July 2021 (Luo et al., 2024). The Geospatial Graph Affiliation Generation Model uses node2vec embeddings as GCN inputs and infers a community affiliation matrix $P(m,k)=1-\exp(-F_m \cdot F_k^T)$8, with edge probabilities $P(m,k)=1-\exp(-F_m \cdot F_k^T)$9. The learned structure is explicitly overlapping: locations can have multiple positive community memberships. The paper reports that the overlap index correlates with POI entropy at $P(A)$0 on weekdays and $P(A)$1 on weekends, and interprets this as showing that 95.7% of urban functional complexity stems from overlap during weekdays (Luo et al., 2024). It also finds strong correlations between overlap and income, and between overlap and racial composition, indicating that CityHood as overlapping urban belonging is deeply stratified.

The relational model of neighborhood mobility extends this logic to amenity and culture. Neighborhoods are nodes, flows are co-visits or residential moves, and dyadic fixed-effects negative binomial models estimate how geographic similarity, amenity mix similarity, and scene similarity shape connection strength (Silva et al., 12 Dec 2025). In the US, a 1-SD increase in amenity similarity is associated with $P(A)$2 more co-visits, and a 1-SD increase in scene similarity with $P(A)$3 more co-visits, net of geographic, racial, educational, political, and rent controls (Silva et al., 12 Dec 2025). In Canada, the corresponding effects on residential moves are $P(A)$4 and $P(A)$5, with a strong positive amenity × scene interaction (Silva et al., 12 Dec 2025). This suggests a CityHood defined by “soft infrastructure”: symbolic cues and amenity ecologies that connect neighborhoods even when distance or demographics would not fully predict those ties.

Hoodsquare offers an earlier neighborhood-detection formulation built from Foursquare check-ins relayed via Twitter. Cities are partitioned into 100 m × 100 m cells, each represented by a 310-dimensional vector comprising 298 place-type features, 10 time features, and 2 local/tourist features (Zhang et al., 2013). Local homogeneity is quantified by

$P(A)$6

and connected high-homogeneity cells are assembled into contiguous neighborhoods via a thresholded depth-first search (Zhang et al., 2013). In recommendation experiments for Manhattan, the small-neighborhood variant achieves Accuracy@10 of about 54% with AreaCost@10 of about 2.41 km², illustrating a CityHood notion tuned for geographically precise recommendation rather than only descriptive mapping (Zhang et al., 2013).

Other activity-signature approaches reach similar conclusions through different data. Twitter Activity Timeline Signatures normalize each ZIP code’s weekly tweet distribution as

$P(A)$7

with $P(A)$8 15-minute bins, and then cluster New York ZIP codes into functional zones such as commercial downtown, nightlife/entertainment, high-rise residential, low-rise residential, mixed-use, and recreational areas (Kats et al., 2017). Mobile-traffic comparison across New York, London, and Hong Kong builds analogous normalized weekly signatures for calls, SMS, requests, and data, then shows that business cores across all three cities collapse into a common transversal cluster while residential clusters remain city-specific (Grauwin et al., 2014). In both cases, CityHood appears as a temporal rhythm.

4. Semantic, symbolic, and perceptual CityHood

Some of the most explicit formulations of CityHood rely on human-interpretable semantics rather than latent geometry alone. In audio-only city identification, each city soundtrack is projected onto a flat taxonomy of 10 urban sounds from UrbanSound8K—air conditioner, car horn, children playing, dog bark, engine idling, gun shot, jackhammer, siren, drilling, and street music—using

$P(A)$9

The resulting semantic weights and reconstructed signals are used to classify 1,080 Flickr videos from 18 cities in the MediaEval Placing Task (Elizalde et al., 2016). The best result is 24.2% EER with an MLP on the reconstructed signal, improving on the 25.3% EER Lei baseline. The paper also shows that using the full 10 bases outperforms smaller taxonomies: 30.4% EER for pairs, 28.5% for quintuples, 30% for octuples, and 24.2% for all 10 (Elizalde et al., 2016). CityHood here is an acoustic fingerprint with semantic evidence: sirens, car horns, dog barks, or children playing become interpretable reasons for city attribution.

Streetonomics translates symbolic commemoration into urban identity. Across 4,932 honorific streets in Paris, Vienna, London, and New York, the method records denomination date, honoree gender, occupation, country of origin, and lifespan, then derives indices for female representation, foreignness, occupational rankings, and temporal focus (Bancilhon et al., 2021). The findings are strongly city-specific: Paris has 4% of today’s honorific streets named after women; Vienna reaches 54% in the 2010s; London peaks around 40% in the 1980s; New York reaches 26% in the 2010s (Bancilhon et al., 2021). Foreign-honoree shares also diverge sharply: Vienna 44.6%, London 14.6%, Paris 10.9%, and New York 3.2–5% (Bancilhon et al., 2021). This yields a symbolic CityHood grounded in memory politics, gendered public recognition, and cosmopolitan orientation.

Perceptual CityHood is built from street-view imagery and human judgments. In London, 25,154 street-view images are paired for the question “On which street would you prefer to walk?”, producing 25,987 usable games from 207 raters after quality filtering (Muller et al., 2022). Image scores are inferred with TrueSkill and predicted with a ResNet‑101 model; city-wide predictions are then aggregated to Output Areas. For walkability, the best model obtains MSE around 2.52–2.58 and Pearson correlation around 0.15–0.16, lower than Place Pulse models for other perceptions such as safety or wealth (Muller et al., 2022). Interpretability analysis using DeepLab segmentation, CenterNet detection, and logistic regression shows that sidewalk has the strongest positive association with walkability, while truck and bus are negative (Muller et al., 2022). CityHood in this sense is a perceptual field of preferred, safe, lively, beautiful, or depressing streets.

The 311-signature literature supplies a civic-administrative semantic layer. For an area $P(A)\sim 1/A^\tau$0, the signature is

$P(A)\sim 1/A^\tau$1

the vector of complaint-type shares (Wang et al., 2016). K-means clustering of tract-level signatures with $P(A)\sim 1/A^\tau$2 yields neighborhood types that align strongly with income, education, race, and poverty. Out-of-sample $P(A)\sim 1/A^\tau$3 for tract-level socioeconomic prediction reaches 0.70 for income per capita in New York, 0.85 for African-American share in Chicago, and 0.63 for below-poverty share in Boston (Wang et al., 2016). At ZIP level, 311 signatures also predict relative housing prices, with Extra Trees $P(A)\sim 1/A^\tau$4 of 0.79 in New York, 0.90 in Chicago, and 0.83 in Boston (Wang et al., 2016). This makes CityHood legible through the composition of reported nuisances and service failures.

5. CityHood as recommendation interface and intelligent urban system

The term also appears directly as a deployed system in "CityHood: An Explainable Travel Recommender System for Cities and Neighborhoods" (Santos et al., 24 Jul 2025). Here CityHood is an interactive recommender that operates at two scales: cities represented by CBSAs and neighborhoods represented by ZIP codes. The behavioral substrate is a filtered Google Places corpus of 245M reviews by 4.6M users, derived from an original dataset of 666M+ geo-tagged reviews over about 5M U.S. venues and 113M users (Santos et al., 24 Jul 2025). User interest in a region is approximated by the number of reviews written there; top and bottom regions are identified by dense ranking and used as implicit positive and negative signals. Separate LightGBM binary classifiers are trained at city and neighborhood scale, and local explanations are generated with LIME via

$P(A)\sim 1/A^\tau$5

Offline evaluation reports city-level recall between 0.66 and 0.79 and F1-score between 0.56 and 0.65 for $P(A)\sim 1/A^\tau$6, with neighborhood-level recall between 0.59 and 0.77 and F1-score between 0.47 and 0.75, outperforming popularity and item-based collaborative filtering baselines (Santos et al., 24 Jul 2025).

This system-level framing has antecedents in Hoodsquare’s map-based tool for neighborhood exploration and in the broader smart-city platform literature. "City-Scale Intelligent Systems and Platforms" argues that urban intelligence requires integrated layers of measurement, connectivity, data access, controls, and communication to users, rather than isolated applications (Nahrstedt et al., 2017). The paper emphasizes open platforms, distributed sensing, edge computing, privacy, resilience, and cross-sector coordination, using examples such as Array-of-Things, connected autonomous vehicles, and transportation electrification (Nahrstedt et al., 2017). A plausible implication is that CityHood systems can be read as neighborhood- or city-scale applications running on top of a larger cyber-physical urban platform.

Recommendation and platform work make a methodological shift. Earlier CityHood studies are primarily descriptive or inferential: they classify, fingerprint, or explain. Recommender and platform studies make CityHood operational: they expose identities to end users, allow preference elicitation, and embed explanations in decision support.

6. Limits, open problems, and future directions

A recurring limitation is partial observability. Audio cityhood is constrained by a 10-sound flat taxonomy, non-orthogonal bases, and dataset mismatch between UrbanSound8K clips and noisy MediaEval videos (Elizalde et al., 2016). Street-pattern typology uses only block area and shape, not their spatial arrangement, road hierarchy, or social context (Louf et al., 2014). The SOM-based global morphology approach relies on stylized map tiles and coarse environmental overlays (Nice et al., 2019). Mobility-based studies note bias in mobile phone, review, and tax-record data, as well as the use of ZIP codes or FSAs as imperfect neighborhood proxies (Luo et al., 2024, Silva et al., 12 Dec 2025). Walkability modeling in London is limited by a low games multiplier, rater composition, and the temporal mismatch between 2018 imagery and 2022 perception collection (Muller et al., 2022). Twitter-based neighborhood signatures face sparsity and platform-specific demographic bias (Kats et al., 2017).

A second limitation is non-causality. Most models are strong on pattern extraction and prediction but weaker on identification. The amenity-and-scene mobility paper is explicit that fixed effects do not yield causal estimates (Silva et al., 12 Dec 2025). 311-based socioeconomic and housing models are predictive rather than mechanistic (Wang et al., 2016). Indicator-group work shows that high-income-profile and high-income-low-age-profile visitors anticipate rent increases, but does not claim that their visits cause those increases (Steentoft et al., 2017). This suggests that CityHood is often best treated as an observational construct: a high-dimensional signature of urban state, not a closed-form causal theory.

A third limitation concerns boundary and overlap assumptions. Hoodsquare constructs contiguous neighborhoods by design (Zhang et al., 2013), whereas overlapping-community work argues that urban communities are intrinsically overlapping rather than mutually exclusive (Luo et al., 2024). Street-pattern work shows that a city fingerprint can be decomposed into borough-level components, implying internal heterogeneity even within a single named city (Louf et al., 2014). The literature therefore warns against equating CityHood with a single crisp partition of urban space.

Future directions are correspondingly plural. Several papers explicitly call for larger or richer taxonomies, whether of urban sounds (Elizalde et al., 2016), street-name semantics (Bancilhon et al., 2021), or neighborhood scene and amenity features (Silva et al., 12 Dec 2025). Others point to temporal modeling, causal identification, multimodal integration, and multi-city comparison (Santos et al., 24 Jul 2025, Muller et al., 2022, Nahrstedt et al., 2017). Taken together, these directions suggest that the next phase of CityHood research will likely be multimodal, multiscale, and more explicitly socio-technical: combining morphology, mobility, perception, civic data, and explainable interfaces while retaining attention to privacy, governance, and inequality.

The literature therefore supports a broad but technically precise understanding of CityHood. It is the measurable distinctiveness of urban places, whether encoded in block geometry, sound mixtures, complaint compositions, temporal activity rhythms, commemorative canons, overlapping mobility communities, or user-conditioned recommendation models. What varies across studies is the sensorium and the mathematical machinery; what persists is the attempt to turn urban identity into a reproducible object of analysis.