MALLS: Urban, Retail & Computational Infrastructure

• MALLS are multifaceted infrastructures that include shopping malls functioning as urban public spaces and social nodes, as well as machine learning benchmarks for logic translation and active learning.
• They are analyzed using mobility models, indoor graph representations, and real-world data such as cell phone and Twitter flows to map spatial accessibility and customer dynamics.
• Operational techniques in malls span indoor navigation, robotics, 3D imaging, and digital implementations, demonstrating their dual role as physical destinations and computational testbeds.

In contemporary research, malls are studied in two distinct senses. The primary sense is the shopping mall: a multi-store environment “under one roof” with shared space, centralized management, and common services, but also a contested urban form that can function simultaneously as retail infrastructure, mobility attractor, social node, indoor sensing environment, and navigable graph (Khalil, 2015, Liu et al., 2018). A second, acronymic sense appears in machine learning and formal logic, where MALLS denotes both a natural-language-to-first-order-logic dataset and a method for active learning under label shift (Vossel et al., 26 Sep 2025, Zhao et al., 2020). Across these literatures, malls are treated less as a single object than as a family of spatial, computational, and socio-technical systems.

1. Urban centrality, publicness, and social mixing

In urban studies, shopping malls are described as pseudo-public spaces: formally private environments that nonetheless operate as meeting grounds, leisure sites, and everyday destinations. One line of work states that shopping malls offer “an open, safe and democratic version of the public space,” while also noting a longstanding critique that malls target customers in subtle ways and may promote social exclusion (Liu et al., 2018). In a medium-sized Chilean city, malls are explicitly grouped with CBDs, leisure areas, university campuses, transport hubs, and parks as “spaces of social interaction” where “many social activities outside homes occur” (Salas-Olmedo et al., 2016).

The strongest city-scale evidence in the supplied literature comes from Santiago de Chile, where mobility of 387,152 cell phones around 16 large malls during one month was modeled using XDR data (Liu et al., 2018). The influx of people to malls is represented by a gravity model,

$$\mathbf{E}[F_{ij}] = \exp\left[\log(G) + \alpha \log(M_i) + \beta \log(M_j) - \gamma \log(D_{ij})\right],$$

with estimated parameters $\alpha = 0.5240$, $\beta = 0.4944$, and $\gamma = 1.1586$ (Liu et al., 2018). The model predicts customer profile distributions in terms of mall location, population distribution, and mall size, and the paper reports that a social-attraction term is negligible in mall choice once those variables are controlled (Liu et al., 2018).

This body of evidence does not support a simple claim that malls either dissolve or reproduce segregation. In Santiago, social mixing arises “only in peripheral malls located farthest from the city center, which both low and middle class people visit,” while co-visitation patterns show that people “choose a restricted profile of malls according to their socio-economic status and their distance from the mall” (Liu et al., 2018). A plausible implication is that mall inclusiveness is not an intrinsic property of the building type; it is mediated by metropolitan accessibility, location, and surrounding socio-spatial structure.

2. Mobility patterns, attraction fields, and ordered routes

Research on mall access often models malls as destinations with measurable catchment areas. In Concepción, Chile, geolocated tweets from 1 January – 31 March 2016 were used to map flows from census districts to public spaces, including three malls: Trebol mall, Centro mall, and Mirador mall (Salas-Olmedo et al., 2016). Residence was inferred from tweets between 22:00 and 07:59 on regular weekdays (Monday–Thursday), and destination relations were derived from tweets falling inside custom mall polygons (Salas-Olmedo et al., 2016). For each origin district $i$ and mall $j$, the raw flow was defined as

$$F_{ij} = \text{number of distinct Twitter users who reside in district } i \text{ and tweeted at } j,$$

and normalized impact was interpreted as the “proportion of Twitter users from one district to a particular public space” (Salas-Olmedo et al., 2016).

The resulting mall patterns are sharply differentiated. Trebol is described as “the largest in size and shops,” “the furthest away from the city centre,” and “the easiest to reach by car,” and it shows “the largest impact on most districts, including the city centre” (Salas-Olmedo et al., 2016). Centro mall shows “a sharp decrease in its impact outside the inner city boundaries,” while Mirador “extends most of its impact along the northwest–southeast axis, which is coincident with the railway and bus network” (Salas-Olmedo et al., 2016). The paper’s most direct claim is that mall impacts follow accessibility structure: “One clear example is the case of the malls, which impact’s spatial pattern is similar to the transport network of the mode that provides higher accessibility in each case” (Salas-Olmedo et al., 2016).

A different formalization appears in route-planning research, where malls are points of interest in top-k optimal sequenced routes. In “Finding Top-k Optimal Sequenced Routes -- Full Version” (Liu et al., 2018), a road or transportation network is modeled as

$G = (V, E, F, W),$

with vertex categories such as gas stations, restaurants, and shopping malls, and a query $(s,t,C,k)$ asks for the top-$k$ least-cost routes from source $\alpha = 0.5240$0 to destination $\alpha = 0.5240$1 that visit required categories in the order $\alpha = 0.5240$2 (Liu et al., 2018). The paper proposes PruningKOSR and StarKOSR, explicitly allowing general graphs whose edge weights may not satisfy the triangle inequality, and treats “shopping mall” as a standard category in ordered itineraries (Liu et al., 2018). This suggests that mall research spans both descriptive mobility analysis and prescriptive routing, with malls functioning either as empirically observed attractors or as explicit constraints in path search.

3. Indoor representation: floorplans, localization, and corridor-use prediction

At indoor scale, malls are modeled as layered spatial graphs derived from floorplans. “Prediction of Usage Probabilities of Shopping-Mall Corridors Using Heterogeneous Graph Neural Networks” constructs vector layers for corridors, shops, and entrances, converts them into heterogeneous graphs, and predicts the probability of usage (PoU) of each corridor edge from shop area, shop usage categories, and graph structure (Barakathullah et al., 10 Apr 2025). The dataset comprises 66 real Chinese malls, using only the first floor of each mall, and generates 200 synthetic samples per mall by varying shop area and usage categories before computing corridor usage from a probability model and shortest paths (Barakathullah et al., 10 Apr 2025). The prediction target is edge-level, and edge features are built by averaging and multiplying hidden node vectors before passing them to an MLP decoder (Barakathullah et al., 10 Apr 2025).

A related but earlier line of work addresses direct user localization. “Indoor Navigation on Google Maps and Indoor Localization Using RSS Fingerprinting” overlays indoor floor plans on Google Maps and estimates position from Wi-Fi signal strength on Android smartphones (Ramani et al., 2014). The method uses standard offline/online RSS fingerprinting and shortest-path routing with Dijkstra’s algorithm, and the accompanying technical description reports claimed accuracy around 1.5 m in the tested building-scale setting (Ramani et al., 2014). In mall terms, this corresponds to georeferenced indoor overlays, floor switching, fingerprint matching, and graph-based routing between stores, corridors, stairs, and similar nodes.

For full 3D reconstruction, “Multi-Sensor Integration for Indoor 3D Reconstruction” presents Scannect, described as the first joint static-kinematic indoor 3D mapper (Chow, 2018). The system fuses a FARO Focus3D S 120 terrestrial laser scanner, two Microsoft Kinect RGB-D cameras, and a low-cost Xsens MTi IMU on a trolley platform (Chow, 2018). The mapping strategy is explicitly “joint static–kinematic”: TLS is used in stop-and-go mode for accurate, wide-field anchors, while Kinect+IMU provides continuous kinematic mapping between scan stations (Chow, 2018). In a 120 m office-floor experiment with four TLS scans, the reported check-target RMSE was approximately 10.0 cm in $\alpha = 0.5240$3, 10.8 cm in $\alpha = 0.5240$4, and 9.0 cm in $\alpha = 0.5240$5 (Chow, 2018). A plausible implication is that malls can be represented at multiple resolutions simultaneously: as 2D georeferenced overlays for wayfinding, as typed graphs for learning and routing, and as dense 3D point clouds for asset and safety applications.

4. Operational technologies in malls: robotics, wireless systems, and acoustic sensing

Malls also appear as instrumented indoor environments for robotics and communications. In human-robot interaction, “Mobile Robot Yielding Cues for Human-Robot Spatial Interaction” studies doorway and bottleneck encounters in “public spaces such as shopping malls, airports, and urban sidewalks” (Hetherington et al., 2021). The robot’s nominal linear speed is

$\alpha = 0.5240$6

and five yielding cues are compared: Stop, Decelerate, Retreat, Tilt, and Nudge (Hetherington et al., 2021). In an online study with 102 participants, Retreat was the most socially acceptable cue and was interpreted as yielding 85% of the time, higher than the approximately 75% reported for Stop and Nudge (Hetherington et al., 2021). The result is operational rather than architectural: malls require robots whose local trajectories communicate intent legibly, not merely collision avoidance.

For wireless infrastructure, “Indoor Millimeter-Wave Systems: Design and Performance Evaluation” treats the 3GPP eMBB Indoor Hotspot scenario as explicitly including shopping malls (Kibiłda et al., 2020). The paper uses experimentally validated indoor channel models and system-level analysis to evaluate coverage, spectral efficiency, and area traffic capacity. It cites downlink targets of approximately 15 Tb/s/km² for area traffic capacity and 1 Gbit/s for 5th-percentile experienced data rate, and reports body-blockage offsets in indoor measurements ranging from roughly 7 dB to 20 dB depending on environment and device posture (Kibiłda et al., 2020). The design conclusion is consistent: narrow beams and short serving distances are beneficial, and malls are treated as natural initial deployment environments for indoor mmWave systems (Kibiłda et al., 2020).

Acoustic sensing adds another layer of operational analytics. “LSTM-CNN Network for Audio Signature Analysis in Noisy Environments” frames malls as one of several target environments for estimating the number of simultaneous speakers and their gender composition from sound (Damacharla et al., 2023). The model uses 19,000 audio samples, assumes a maximum of 10 speakers, and reports training/validation MSE values of about 0.019/0.017 for count and gender detection (Damacharla et al., 2023). Mall recordings are included among the environmental noises added at approximately 10 dB SNR (Damacharla et al., 2023). This suggests that malls are increasingly treated not only as places to be navigated, but as complex sensory fields in which mobility, speech activity, wireless propagation, and robot behavior must all be modeled jointly.

5. Virtual malls and the online reimplementation of the mall form

The mall concept is also reimplemented as a digital environment. “The 3D virtual environment online for real shopping” presents a 3D virtual shopping mall in which users move through a mall-like building modeled in VRML, enter stores through clickable doors, and complete transactions through ASP.NET pages backed by SQL Server (Khalil, 2015). The system reproduces the physical mall model—many stores under one roof, shared management, centralized authentication, and shared services—within a web browser (Khalil, 2015).

Its architecture has three sides: client-side rendering of the VRML world and store pages, a server-side ASP.NET application and SQL Server database, and a mall management layer that integrates multiple stores into one mall (Khalil, 2015). The mall management can authenticate clients across all participating stores with a single login, and the virtual mall “allows shoppers to perform actions across multiple stores simultaneously such as viewing product availability” (Khalil, 2015). Entities in the accompanying ER diagram include Administration, Shops, Items, Products, Offers, Customer, CreditCard, and Recommendations (Khalil, 2015).

This work treats the mall less as a metaphor than as an organizing principle. Floors, corridors, storefronts, and category-based placement are translated directly into online spatial navigation; meanwhile, checkout, payment processing, and administration remain conventional e-commerce components (Khalil, 2015). A plausible implication is that the mall persists digitally because it combines exploration, centralized services, and multi-tenant coordination in a form that flat product catalogs do not reproduce.

6. Acronymic uses: MALLS in formalization benchmarks and label-shift learning

In machine learning, MALLS names at least two independent technical objects unrelated to shopping-mall architecture except by acronym. The first is a benchmark for natural language to first-order logic translation. In “Advancing Natural Language Formalization to First Order Logic with Fine-tuned LLMs,” MALLS is one of two core datasets combined with Willow into a corpus of 49,950 examples, with 35,964 train, 3,996 validation, and 9,990 test instances, plus 62,981 unique predicates and 2,011 unique constants (Vossel et al., 26 Sep 2025). The original MALLS resource, introduced by Yang et al., contains 34K sentence-level NL–FOL pairs generated and verified using GPT-4 (Vossel et al., 26 Sep 2025). On the combined corpus, a fine-tuned Flan-T5-XXL achieves about 70% logical-equivalence accuracy when given predicate lists, and the paper’s central conclusion is that predicate extraction, not formula structure, is the main bottleneck (Vossel et al., 26 Sep 2025).

That benchmark is itself controversial. “Fixing FOLIO and MALLS: Verified Annotations and an LLM-assisted Framework to Focus Human Relabeling” audits the first 100 instances of the claimed human-checked MALLS test set and finds 36% incorrect FOL formalizations and 48% ambiguous NL sentences (Brunello et al., 1 Jun 2026). With corrected annotations, translation accuracy on this MALLS subset increases by +18.0 points for Gemma 4 31B-it, +12.9 for Qwen3-30B-A3B, and +9.0 for GPT-4o-mini on all instances (Brunello et al., 1 Jun 2026). The same paper proposes a Verdict-and-Refinement review framework and reports that, on MALLS, 13% of instances must be reviewed to reach 90% dataset accuracy under the best pipeline, compared with 72% under unguided review of the original subset (Brunello et al., 1 Jun 2026). The misconception addressed here is clear: a “human-checked” benchmark is not necessarily a reliable gold standard.

The second acronymic use is Mediated Active Learning under Label Shift, also called MALLS. In “Active Learning under Label Shift,” the setting assumes

$\alpha = 0.5240$7

and introduces a medial distribution that interpolates between source and target label marginals to balance the variance of importance weighting against the bias of class-balanced subsampling (Zhao et al., 2020). The method inherits theoretical guarantees from disagreement-based active learning and is reported to reduce sample complexity by 60% in deep active-learning tasks (Zhao et al., 2020). Here, “MALLS” has no relation to retail space; it is an algorithmic name for a bias–variance mediation strategy under label shift.

These acronymic uses are methodologically important because they turn “MALLS” into a site of benchmark governance and learning theory. One MALLS exposes failures in symbolic formalization data and predicate vocabularies; the other formalizes how active learning should proceed when class priors move across domains (Vossel et al., 26 Sep 2025, Zhao et al., 2020).

7. Synthesis across domains

Taken together, the cited work portrays malls as a research object with unusual breadth. At metropolitan scale, malls are measurable attractors whose reach follows transport accessibility and size more closely than any simple notion of “social attraction” (Salas-Olmedo et al., 2016, Liu et al., 2018). At building scale, they are graph-structured interiors whose corridors, entrances, and shop frontages can be digitized, localized, and reconstructed using Wi-Fi fingerprints, GNNs, RGB-D cameras, laser scanners, and IMUs (Barakathullah et al., 10 Apr 2025, Ramani et al., 2014, Chow, 2018). As operational environments, they are testbeds for socially legible mobile robots, indoor mmWave systems, and acoustic analytics under noise and crowding (Hetherington et al., 2021, Kibiłda et al., 2020, Damacharla et al., 2023). As computational abstractions, they reappear in e-commerce as centralized multi-store virtual environments and, independently, as acronymic benchmark and algorithm names in formal logic and active learning (Khalil, 2015, Vossel et al., 26 Sep 2025, Zhao et al., 2020).

This suggests a unifying editorial term: “mall-as-infrastructure” (Editor's term). In the supplied literature, a mall is rarely just a retail container. It is instead a managed topology in which flows of people, information, packets, robots, labels, and logical forms are organized through common interfaces, shared constraints, and measurable access structures.