AskNearby: Urban POI & Proximity Framework

Updated 9 December 2025

AskNearby is an advanced location-based framework that provides context-aware neighborhood information and personalized point-of-interest recommendations.
It features a three-tier architecture integrating mobile LBS, scalable spatial/textual indexing, and real-time route mapping for efficient retrieval.
The framework leverages deep learning, cognitive-map scoring, and fairness-aware ranking to enhance recommendation accuracy and social equity.

AskNearby is an advanced location-based application framework designed for context-aware neighborhood information retrieval, personalized point-of-interest (POI) recommendations, and proximity-driven service matching within urban or community environments. Emerging from work on Local Life Information Accessibility (LLIA), AskNearby systematizes active retrieval and cognitive-map-based recommendation, integrating algorithmic advances from mobile LBS architectures, LLM pipelines, scalable proximity search, and POI recommendation models.

1. Foundations and System Architecture

AskNearby is architected atop mobile LBS and Android application frameworks, building on principles established for location-centric services (Sari et al., 2016). The system comprises:

Three-tier design: Client (mobile/app), stateless server farm, and backend spatial/textual index.
Core modules:
- Location acquisition: Android LocationManager using GPS_PROVIDER/NETWORK_PROVIDER, periodically emitting $(\varphi, \lambda)$ .
- Data connectivity: Requires persistent Internet for Google Maps tiles, Directions API, and POI JSON fetches from RESTful endpoints.
- Presentation layer: Activities, Fragments, and MapView, supporting ListViews, filtered search, and interactive map rendering.
Map and route rendering: Google Maps Android API (MarkerOptions, PolylineOptions), allowing marker placement and real-time route calculation via the Directions API.
Data flow:

App pulls current $(\varphi_0, \lambda_0)$ .
Issues HTTP GET: /get_pois_nearby?lat=\varphi_0&lng=\lambda_0.
Server returns POIs in JSON (ID, name, $(\varphi_i, \lambda_i)$ , description, image).
Client parses, computes distances, sorts/filters, populates UI.
On user POI selection, detailed info and routing are invoked.

These modular interfaces extend trivially to arbitrary POIs, enabling universal proximity-driven access for restaurants, ATMs, crowd-service providers, etc.

2. Nearest Neighbor Search and Proximity Computation

AskNearby employs scalable neighbor search methodologies:

Haversine metric for spatial proximity (Sari et al., 2016):

$d = 2R \arcsin\left( \sqrt{ \sin^2 \left( \tfrac{\Delta\varphi}{2} \right) + \cos(\varphi_1)\cos(\varphi_2)\sin^2 \left( \tfrac{\Delta\lambda}{2} \right) } \right )$

where $R \approx 6\,371$ km.

Database-side index optimization:
- R-tree or QuadTree spatial indexes (SQL/PostGIS) prune POIs by bounding box or radius before client-side retrieval.
- For very large POI sets or high-dimensional data, scalable neighbor-graph search via beam search metaheuristics and parameter-free logarithmic neighbor selection yields high recall with sublinear search cost and memory overhead (Tellez et al., 2017).
Active search paradigm:

In 2D domains, an image-based zoom-and-count strategy adaptively grows/shrinks a search circle until $k$ neighbors are enclosed, matching $O(k)$ runtime, independent of dataset cardinality (Um et al., 2019).

3. Retrieval-Augmented Generation and Cognitive-Map Recommendation

Recent advances formalize the LLIA objective as returning items $I_u$ most relevant to user $u$ 's semantic query $Q$ , location $s_u$ , time $t_u$ , and cognitive anchors $p_u$ (Niu et al., 2 Dec 2025):

LLIA operator:

$A(u; Q, K) = \mathrm{Retrieve}(Q, s_u, K) \cup \mathrm{Recommend}(s_u, t_u, p_u, K)$

Three-layer RAG pipeline:
- GraphRAG: Neighborhood expansion on semantic concept graph; candidates scored by short-path or personalized PageRank.
- VectorRAG: Semantic-vector retrieval using transformer encoders and HNSW ANN indexing, optimizing nearest neighbor matching for query and item embeddings.
- GeoRAG: Geographic retrieval by spatial window, leveraging R-tree indexes and precise haversine evaluation.
- Candidates are merged and ranked by an LLM for final user delivery.
Cognitive-map scoring:

Final recommendation $\Psi(u, i)$ combines user-location semantic relevance, proximity, and public familiarity:

$\Psi(u, i) = [f_{\mathrm{sem}}(u, i)]^\alpha \cdot [f_{\mathrm{dist}}(u, i)]^\beta \cdot [f_{\mathrm{pop}}(i)]^\gamma$

with hyperparameters for per-user or global tuning.

Empirical evaluation demonstrates AskNearby's superiority over standard LLMs and keyword engines in precision, NDCG, spatiotemporal grounding, and hallucination rates (Niu et al., 2 Dec 2025).

4. POI Recommendation Algorithms: Deep Learning and Geographic Awareness

AskNearby leverages state-of-the-art POI recommendation models incorporating nonlinear user-POI interactions and neighbor-aware geographic influence (Ma et al., 2018):

Self-attentive autoencoder model (SAE-NAD):
- SAE: Multi-aspect attention produces user embedding by weighting past POIs.
- NAD: Incorporates geographical context via radial basis function (RBF) kernels, enhancing proximity impact on recommendations.
- Training minimizes weighted reconstruction with regularization, using a confidence matrix to encode check-in frequency.

Performance benchmarks on real datasets (Gowalla, Foursquare, Yelp) show MAP@10 improvements (14–16%) over prior best baselines, validating deep geographic-aware personalization.

5. Exposure Bias and Fair Ranking in Proximity Services

Location-based retrieval exposes inherent biases: position and landmark popularity skew exposure away from meritocratic fairness (Banerjee et al., 2020):

Metrics:
- Relevance $V(l, s)$ : Mean rating.
- Exposure $E_j = \sum_i \omega(\rho_i(s_j))$ , with position-bias decay $\omega(t)$ .
- Disparity $B = \sum_j |V_j - E_j|$ : Aggregate exposure deficit.

Empirical findings show systematic under-exposure of high-rated businesses located away from frequent-queries, and over-exposure of low-rated ones near popular spots, especially in platforms like Yelp and Google. Remedies include composite ranking by distance and rating, position-bias reweighting, exposure-equalizing optimization, and providing transparency dashboards for business owners.

A plausible implication is that integrating multi-factor and fairness-aware ranking into AskNearby is essential to mitigate exposure bias and ensure equitable access.

6. Group Proximity Queries: Trip-Based Meetup Optimization

For multi-user coordination, AskNearby integrates trip-based group nearest neighbor (T-GNN) search (Zaman et al., 29 Aug 2025):

Formal model: Given $n$ users each with a trip $L_i = \langle l_i^1, ..., l_i^{m_i} \rangle$ , seek POI $p^* \in P$ and detour indices minimizing total group overhead:

$\Delta_{\text{group}}(p) = \sum_{i=1}^n \Delta_i(p)$

where

$\Delta_i(p) = \min_j [d(l_i^j, p) + d(p, l_i^{j+1}) - d(l_i^j, l_i^{j+1})]$

Efficient algorithm (EA_T-GNN):
- Indexes POIs in R*-tree.
- Pruning by centroids and geometric regions, updating three interlocking circles $S_1, S_2, S_3$ derived from user-trip properties and current best solution.
- Best-first traversal terminates when the known area covers all candidates that could improve the optimum.

Experiments on large city networks (New York, Chicago) confirm linear or sublinear scaling to group and POI count, and reductions in real-world query time by 60–80× over brute-force baselines.

7. Mobile Crowd-Service and Context-Aware Matching

AskNearby generalizes to instant service discovery and provision among local users by implementing spatial-textual crowd-service infrastructure (Wu et al., 2017):

Provider/client registry is maintained with real-time location and skill updates.
Scalable spatial-textual index (BIG-tree; quadtree + inverted lists) enables region-aware search with lazy refinement and prior termination.
Contextual task recommender uses matrix factorization over explicit feedback and context tuples (time, location, skill-domain).
Matching integrates spatial ( $S_\ell$ ), textual ( $S_\psi$ ), temporal ( $S_t$ ), and contextual ( $R(\cdot)$ ) signals in a composite scoring function and delivers sub-second real-time candidate matching and notifications.

Precision@5 exceeds 0.90, and recommendation components surface latent-skilled providers otherwise missed, expanding system utility.

AskNearby synthesizes graph-based, neural, and geometric proximity search with cognitive modeling in urban spaces, implements fairness-aware ranking to counter exposure bias, and supports both individual and group-level real-time proximity queries. The system advances not only location-based query precision and recommendation but also social equity and actionable local information accessibility in “15-minute city” and mobile crowd-service applications.