Route Recommendation Based on POIs and Public Transportation

Agata Palma

1 a

, Pedro Morais

1 b

and Ana Alves

1,2 c

Polytechnic University of Coimbra, Rua da Miseric

ordia, Lagar dos Cortic¸os, S. Martinho do Bispo, 3045-093 Coimbra,

Portugal

CISUC, LASI, University of Coimbra, Polo II, Pinhal de Marrocos, 3030-290,

Coimbra, Portugal

{a2023113935, a21280686, aalves}@isec.pt

Keywords:

GIS, Information Retrieval, Ambient Intelligence, Clustering, Route Recommendation, POIs.

Abstract:

With the rapid advancement of technology in today’s interconnected world, Ambient Intelligence (AmI)

emerges as a powerful tool that revolutionizes how we interact with our environments. This article delves

into the integration of AmI principles, Python programming, and Geographic Information Systems (GIS) to

develop intelligent route recommendation systems for urban exploration. The motivation behind this study

lies in the potential of AmI to address challenges in urban navigation, personalized recommendations, and

sustainable transportation solutions. The objectives include optimizing travel routes, promoting sustainable

transportation options, and enhancing user experiences. This research will contribute to advancing AmI tech-

nologies and their practical applications in improving urban living standards and mobility solutions.

1 INTRODUCTION

Ambient Intelligence (AmI) represents a new

paradigm in computing that aims to embed intelli-

gence into everyday environments. It involves the in-

tegration of computational capabilities into ordinary

objects, allowing them to interact with users and each

other in a natural and intelligent manner. AmI em-

phasizes the presence of humans alongside smart in-

terfaces that can adapt to human emotions, behaviors,

and expectations. This concept envisions the creation

of smart environments, such as smart homes, smart

healthcare facilities, and smart cities, where everyday

objects are seamlessly connected and capable of en-

hancing daily living experiences. AmI is seen as a

signiﬁcant societal and cultural shift, with the poten-

tial to transform the way people interact with technol-

ogy and their surroundings (Thankachan, 2023).

Since AmI takes advantage of sensors and Internet

of Things (IoT) devices to gather information about

the surrounding environment, it is a useful tool to

make inferences based on proximity, intent, and be-

havioral patterns. This facilitates personalized ex-

periences, as for example, receiving location-based

alerts when reaching points of interest (POIs) in a new

https://orcid.org/0009-0009-4450-700X

https://orcid.org/0009-0003-5962-0386

https://orcid.org/0000-0002-3692-338X

city(Mahmood et al., 2023).

Ambient Intelligence enhances the accuracy and

relevance of environmental data by incorporating Ge-

ographic Information Systems (GIS) and Informa-

tion Retrieval techniques. GIS offers spatial analy-

sis and mapping to understand user interactions ge-

ographically, while Information Retrieval efﬁciently

extracts relevant data for context-aware recommen-

dations. Clustering techniques group similar data

points to identify patterns in user behavior, aiding

route recommendation systems by predicting optimal

paths based on historical data and preferences. These

technologies enable AmI to create intelligent environ-

ments that anticipate and respond to user needs, pro-

viding seamless and enriched interactions.

The motivation behind this study lies in the prac-

tical application of advanced geospatial technologies

and algorithms to enhance urban navigation. The

aim is to develop an intelligent system that can pro-

vide efﬁcient, customizable routing solutions tailored

to individual preferences, particularly in urban envi-

ronments where efﬁcient navigation and personalized

experiences are crucial to navigate busy zones and

discover POIs. Therefore, these intelligent systems

can offer context-aware recommendations and opti-

mize routes tailored to user preferences and sustain-

able transport options, ultimately promoting efﬁcient

travel and contributing to sustainable urban mobility.

Palma, Á., Morais, P. and Alves, A.

Route Recommendation Based on POIs and Public Transportation.

DOI: 10.5220/0013006800003838

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 1: KDIR, pages 391-398

ISBN: 978-989-758-716-0; ISSN: 2184-3228

391

This work explores the integration of AmI prin-

ciples, Python programming, and GIS to develop in-

telligent route recommendation systems for urban ex-

ploration based on POIs and available public trans-

portation. Given a city and, optionally, a category,

the system will reply with a list of POIs and a sug-

gested route to visit the largest number of POIs in

the shortest route possible using public transporta-

tion. The base of this work is a variant of the Trav-

elling Salesman Problem (TSP), which involves ﬁnd-

ing the shortest possible route that visits a set of given

locations exactly once and then returns to the start-

ing point (

Ozcan and Kaya, 2018). The challenge in

this research is similar to the one on TSP: determine

the most efﬁcient route between multiple POIs while

minimizing the total travel distance.

The structure of the article includes a review of re-

lated works and AmI principles and their relevance in

urban navigation, followed by a discussion of techni-

cal aspects such as data integration, route optimiza-

tion algorithms, and visualization techniques using

platforms like Quantum GIS (QGIS).

Practical implications, potential extensions, and

the broader impact of AmI-driven solutions on urban

mobility and city planning are also addressed in the

discussion and conclusions sections. This work aims

to contribute to the advancement of technology that

enhances user experiences and promotes sustainable

and efﬁcient mobility solutions in urban settings.

2 RELATED WORK

In this section, a brief literature review is presented,

focusing on existing applications, systems, and stud-

ies that share similar objectives or themes related to

AmI and intelligent route recommendation systems

for urban exploration based on POIs.

Based on the TSP, (

Ozcan and Kaya, 2018)

aimed to create a new tourist guide app using Open-

StreetMap (OSM). To achieve this, the study involved

various tasks using OSM tools, libraries, and frame-

works. These tasks included real-time area drawing

on OSM, path computation, selection of POIs, and

map understanding. The app intends to determine the

shortest route between user-selected destinations, op-

timizing travel time and displaying the route visually

on the map. The Hill Climbing Algorithm (HCA),

known for its memory efﬁciency and local search ap-

proach, was used for the TSP.

On the itinerary recommendation variant, (Pana-

giotakis et al., 2022) proposed a method to person-

alize itinerary recommendation (PIR) with POIs cat-

egories, for tourists tours. The authors’ method was

based on the Expectation Maximization (EM) algo-

rithm, and solves, sequentially, the PIR problem by

selecting POIs that maximize a suitable objective

function, such as user satisfaction, user time bud-

get, POIs opening hours, POIs category and spatial

constraints. In a similar scope, (Lou, 2022) focused

on categorizing POIs but with an improved k-means

algorithm to be applied to intelligent tourism route

planning. The proposed scheme considers tourists’

preferences and aims to ﬁnd the shortest route be-

tween desired locations within a selected area.

(Mahdi et al., 2023) also redirected their research

focus towards POIs. They applied regression mod-

els to analyze the data obtained from Google Popu-

lar Times (GPT) to predict the amount of time people

would spend at POIs. With this contribution, a sim-

ilar process would be possible to improve the route

generation plan when time constraints are a variable.

Besides the prediction of the time spent at a POI,

when planning a route based on public transportation,

it is also crucial to take into account the time spent

from one point to another. (Zhang et al., 2022) state

the importance of improving travel time prediction.

The study highlights the importance of real-time, ac-

curate, reliable and low-cost multi-source data for bet-

ter predictions. The authors afﬁrm that the traditional

methods for predicting travel time are deﬁcient and a

new approach based on intelligent technology would

improve the prediction accuracy. In order to accom-

plish this, a prediction model based on the Kalman

ﬁlter - high accuracy in one-step prediction - was de-

signed. For this model, two sub-modules were cre-

ated: the Route Travel Time Prediction Model - pre-

dicts travel time for an entire bus route - and the Stop

Dwell Time Prediction - predicts the time spent at bus

stops. In this study, the data sources used included

GPS (Global Positioning System), AFC (Automatic

Fare Collection), and IC (Integrated Circuit) and the

models were validated using Automatic Vehicle Lo-

cation (AVL) from real world scenarios. The results

indicate the prediction model meets accuracy require-

ments for travel time prediction.

(Sarridis et al., 2022) proposed a personalized

route recommendation system that balances the trade-

off between distance and POIS using hypergraph

models. Their framework considers tourist satisfac-

tion and leverages both visual and geographical data

to optimize the shortest path algorithm through POI

images embedded in a hypergraph. Similarly, (Karan-

taidis et al., 2021) applied multi-stage optimization

learning in hypergraph structures for image and tag

recommendations, dynamically updating hypergraph

structures and hyperedge weights to achieve higher

accuracy in POI ranking and recommendations.

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

392

The contribution of (Li et al., 2021) to this work

is based on a solution for another problem. Instead

of the common questions such as “ﬁnd the k near-

est POIs around me” or “give me the bus plan from

s to d”, the authors proposed a method to answer the

“give me k POIs that I can reach earliest within one

transfer by bus”. In the public transportation network

(PTN), the users’ primary concern is “which POI can

be reached with the least travel time under some spec-

iﬁed transfer numbers considering the different depar-

ture time and frequency of buses”. To answer the pro-

posed question, the k-nearest neighbor (kNN) query

should be applied. Given a set of information—POIs,

a PTN, a location, departure time, and a transfer num-

ber constraint—the kNN query returns the k POIs that

meet these conditions.

Another possible method for location-based sys-

tems, besides kNN, is the Multi-Cost Transportation

Network-constrained skyline query (MCTN-CSQ).

(Gong et al., 2020) implemented the CSQ System,

the ﬁrst of its kind, as a web application supporting

constrained skyline query on multi-cost transportation

networks. Users input query points and receive sky-

line answer-objects reachable via transportation net-

works, superior on at least one dimension. For exam-

ple, lets assume a user needs to book a room for the

night but he has more constraints about the desired

room: it can be reached by taking public transporta-

tion, and the transportation fare and the travel time

should be reasonable - the query processing compo-

nent of the CSQ System handles the query execution.

“The system is implemented as a web application,

which allows users to input a query point from a web

interface, get the skyline result by using several algo-

rithms, and display the result on the web interface”

(Gong et al., 2020).

3 SYSTEM ARCHITECTURE

The application is designed to provide a comprehen-

sive solution for route planning and analysis within

the QGIS environment. The system architecture is

composed of several elements, interconnected to fa-

cilitate preprocessing, route generation, spatial analy-

sis, visualization, and user interaction.

The user interface, implemented using the QGIS

interface in the ﬁrst phase and a plugin in QGIS in

the second phase, serves as the entry point for users

to select the city and visualize the suggested route, as

well as to specify POI categories. The route genera-

tion engine employs an algorithm to compute the op-

timal route, connecting different POIs based on user-

deﬁned parameters. The aim is to ﬁnd the shortest

path in the road network integrating public transport

routes and stops.

In the ﬁrst phase, a proof of concept is achieved

by working with the available processing tools and

plugins in QGIS, such as ORS (OpenRouteService)

tools. The system uses data provided for the devel-

opment of this project, speciﬁcally POIs and roads in

Portugal and public transportation in Coimbra. The

spatial visualization is handled by the QGIS environ-

ment, taking advantage of HeatMaps and route over-

lays to visually present analysis results, as well as to

produce a georeferenced PDF.

For the second phase, a custom plugin is devel-

oped to provide the user with a more friendly and in-

tuitive interface. Using the user’s input for a location

and category of POIs, and the processing of POIs with

machine learning methods, a route is drawn using the

shortest path possible across the region with the most

of these POIs.

4 DATA SOURCES

To populate the application with pertinent data con-

cerning POIs, the primary source relies on Open-

StreetMap for the second step, retrieved through the

“osmnx” python package.

For the ﬁrst step, data containing POIs and roads

of Portugal in shapeﬁles format were provided in the

context of this project, as well as public transporta-

tion data for Coimbra, with routes and stops for SM-

TUC (Servic¸os Municipalizados de Transportes Ur-

banos de Coimbra). Additionally, the application in-

tegrates (1) the QGIS plugin QuickOSM to retrieve

the boundaries of Coimbra city and (2) the module

Quick Map Services (QMS) to procure a standardized

raster layer of OSM.

5 MACHINE LEARNING

To enhance the performance of the application, Clus-

tering is applied, which allows the identiﬁcation of

groups of POIs that are geographically close to each

other. Through this unsupervised learning method,

the most concentrated area of POIs is identiﬁed and a

route is established within that zone. A density-based

cluster analysis algorithm, the Density-Based Spatial

Clustering of Applications with Noise (DBSCAN), is

applied due to its robustness and effectiveness in han-

dling spatial data. As noted in (Lou, 2022), DBSCAN

has the great advantage of clustering dense datasets of

any shape and is “sensitive to the selection of initial

values, but insensitive to noise points and has certain

Route Recommendation Based on POIs and Public Transportation

393

noise immunity”. Another advantage is the unneces-

sary need to predeﬁne the number of clusters. The

initial DBSCAN parameters are:

• eps (ε). The maximum distance between two

points to be considered as part of the same neigh-

borhood. This parameter deﬁnes the radius of the

neighborhood around each point.

• minPts. The minimum number of points required

to form a dense region. A point is considered a

core point if it has at least minPts within its eps

radius.

These parameters are crucial for the performance

of the DBSCAN algorithm. In this study, eps and

minPts are ﬁne-tuned based on the spatial distribution

of POIs in the dataset.

The clustering process entails loading the road net-

work graph using OSMnx, projecting the POIs to align

with the Coordinate Reference System (CRS) of the

graph, and generating a distance matrix based on net-

work distances. Subsequently, the DBSCAN algo-

rithm is employed to detect clusters, and the outcomes

are assessed using metrics like Silhouette Score and

Davies-Bouldin Index.

The Silhouette Score measures how similar a

point is to its own cluster compared to other clus-

ters. Higher values indicate well-deﬁned clusters with

clear separation between them. The Davies-Bouldin

Index assesses the average similarity ratio of each

cluster with its most similar cluster, where lower val-

ues indicate better-deﬁned clusters with less overlap.

These metrics provide a quantitative evaluation of the

clustering quality, ensuring that the clusters formed

are meaningful and accurate.

6 VISUALISATION OF DATA

Throughout the ﬁrst and second phases, different ap-

proaches are utilized for collecting, treating, and dis-

playing the results. Both approaches are addressed,

demonstrating the evolution and reﬁnement of the

methods to achieve a more automated response in cus-

tom route generation.

6.1 Phase I

Taking advantage of the already present module in

QGIS, QMS, the standardized raster layer is retrieved,

providing a comprehensive and detailed map back-

ground in EPSG:4326. This CRS, also known as

WGS 84, is widely used in geographic coordinate sys-

tems and is the one that the ORS API expects in the

requests.

To commence data analysis, the shapeﬁles of Por-

tugal’s POIs and roads are imported, along with the

SMTUC General Transit Feed Speciﬁcation (GTFS)

containing route information and bus stops, facilitated

by the GTFS GO plugin. Furthermore, the polygon

delineating the region of Coimbra is imported using

the QuickOSM plugin. Additionally, a new polygon

is drawn within the Coimbra region to delimit the

analysis area.

A new layer, named Coimbra POIS is created

through the extraction by location of elements that in-

tersect or are contained within the area of the poly-

gon. This layer is subsequently utilized as the foun-

dation for generating a HeatMap, providing a visual

representation of the concentration of POIs within the

delimited area. Since meters are preferred over de-

grees for measurements, the layers are re-projected to

EPSG:3763. This adjustment enables the proper con-

ﬁguration of parameters for the DBSCAN algorithm

(ε: 200 meters; minPts: 4), using the Coimbra POIS

layer as the data source. The outcomes demonstrate

a clear separation of clusters, indicating a satisfactory

ﬁt, as shown in Figure 1. To enhance visualization

and delineate cluster regions more distinctly, concave

hulls are employed for each cluster. A concave hull

is a shape that closely wraps a set of points, capturing

the boundaries of the points more accurately. The re-

sult provided a collection of polygons encompassing

the points within each cluster, as depicted in Figure 2.

Figure 1: Heatmap with DBSCAN clustered POIs.

In this phase, the simulation involves a user who

wishes to travel from point A to point B, as depicted in

Figure 2, utilizing the shortest path and public trans-

portation services. With this goal in mind, a man-

ual approach is adopted for route construction. Us-

ing the ORS tools, a few points are manually selected

as coordinates to create two custom routes, employ-

ing the shortest path and driving-car preferences. Af-

ter re-projecting both custom and SMTUC routes and

stops, each route is segmented into sections of ap-

proximately 500 meters, resulting in the creation of

two new layers: (1) SMTUC sections intersecting the

custom routes and (2) SMTUC stops along the cus-

tom route. Additionally, leveraging ORS tools, a new

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

394

layer with isochrones is created, as seen in Figure 3.

An isochrone is a line or boundary on a map that con-

nects points representing equal travel time or distance

from a particular location. This new layer provided a

visual analysis of the area accessible from each SM-

TUC stop along the route within 2, 5, and 7-minute

thresholds.

Figure 2: Route 1.

Figure 3: Route 2 with isochrones for 2 minutes.

Given the limitations of the ORS API for

isochrones, no route with them is created to the center

of the largest cluster of POIs. Nonetheless, through

visual analysis, it can be conﬁrmed that using this ap-

proach could indeed provide an effective tool for route

customization based on POI concentration and public

transportation.

6.2 Phase II

In the second phase, the system allows users to se-

lect POIs from any geographic location. These POIs

are organized into top-level categories, each contain-

ing subcategories. For example, the Tourism category

includes Hotels and Museums, the Amenity category

features Bars and Cafes, and the Shop category en-

compasses Malls.

The main objective of this stage is to provide a

more automated response to the challenge presented

in the ﬁrst phase of this project. A plugin for QGIS

has been developed - Optimal Custom Route - which

offers an intuitive and efﬁcient tool for route planning

and visualization.

The development environment includes OSMnx

for geographic data handling, OpenRouteService for

routing, gpxpy for GPS Exchange Format (GPX) ﬁle

manipulation, scikit-learn for clustering, and geopy

for geocoding. These libraries provide the necessary

tools for implementing the plugin’s core functional-

ities and can be installed using the following com-

mands:

$pip install osmnx

$pip install openrouteservice

$pip install gpxpy

$pip install scikit-learn

$pip install geopy

The next step focuses on designing and imple-

menting the user interface, developed using PyQt5.

This interface comprises two main windows:

1. The ﬁrst window allows users to input the name

of the city for which they want to plan a route.

2. The second window is dedicated to route cus-

tomization details, as shown in Figure 4.

Figure 4: Customization window with a starting point de-

ﬁned and customized DBSCAN parameters.

To handle geographic data, the plugin utilizes

“OSMnx” to collect and process map data from Open-

StreetMap. The process commences with geocod-

ing the city name to acquire geographic coordinates.

These coordinates are subsequently utilized to import

the relevant map tiles into QGIS as layers, thereby

offering users a visual representation of the area of

interest.

For selecting POIs, users can choose from various

categories, such as tourism, amenities, and shops. The

plugin dynamically generates checkboxes for each

subcategory, allowing for detailed selection. Once the

POIs are selected, the plugin retrieves the correspond-

ing data from OpenStreetMap and saves the data in a

new layer with the name, category, and subcategory

of the POI. Then, it proceeds to clustering using the

pre-deﬁned values of ε = 200 meters and minPts = 4

or custom values chosen by the user.

Clustering analysis plays a crucial role in the plu-

gin, focusing on grouping POIs based on geographic

proximity. To use the collected data, a transformation

is needed. In this phase, the coordinates of the POIs

are converted to a suitable coordinate system to en-

sure accurate distance measurements. Typically, the

Universal Transverse Mercator (UTM) projection is

used because it provides a more accurate represen-

tation of distances compared to latitude and longi-

tude. This is essential for spatial data analysis, as the

DBSCAN algorithm operates on distances between

Route Recommendation Based on POIs and Public Transportation

395

points. To ensure consistency in distance calculations,

the POIs data is projected into the same CRS as the

road network graph generated by OSMnx. This CRS

transformation is important for aligning the POIs with

the graph, allowing for accurate integration and sub-

sequent analysis.

Once the data is transformed, a road network

graph is created using OSMnx. This graph represents

the road network within the speciﬁed place, where

nodes correspond to intersections, and edges repre-

sent road segments connecting these intersections.

The road network graph is truncated to retain only the

largest connected component. This step is needed to

avoid isolated nodes that do not contribute to the main

network, ensuring a coherent and comprehensive road

network for analysis. The truncated graph provides a

strong foundation for mapping POIs and calculating

network distances.

With the road network graph prepared, the POIs

are projected into the same CRS as the graph to main-

tain consistency, and the nearest nodes in the road

network graph are found for each POI. This way, dis-

tances between POIs can be calculated within the con-

text of the road network.

The clustering process involves calculating a dis-

tance matrix, which is essential for applying the DB-

SCAN algorithm. Initially, a pairwise Euclidean dis-

tance matrix is calculated between the nodes repre-

senting the POIs. However, for more accurate dis-

tance measurements that account for the road net-

work, this Euclidean distance matrix is converted into

a network distance matrix using Dijkstra’s algorithm.

This algorithm computes the shortest path between

nodes based on the actual road network distances. By

using the network distance matrix, the DBSCAN al-

gorithm can accurately cluster POIs based on real-

world distances, rather than straight-line distances.

The DBSCAN algorithm is then applied to this net-

work distance matrix. The eps parameter, which is

used in meters, deﬁnes the maximum distance be-

tween two points for them to be considered part of

the same cluster. The minPts parameter speciﬁes the

minimum number of points required to form a dense

region. The metric precomputed is used to indicate

that the distance matrix has already been calculated.

After the clustering is performed, the results are

processed to extract meaningful clusters. Noise

points, which are points labeled as -1 by DBSCAN,

are excluded from further analysis and the largest

and densest clusters are identiﬁed. Subsequently, the

outcomes are assessed using Silhouette Score and

Davies-Bouldin Index.

These clusters are then visualized within the QGIS

environment, providing an intuitive and comprehen-

sive view of the spatial distribution of POIs. This vi-

sualization aids in identifying key areas of interest, as

shown in Figure 5 and supports the generation of an

optimal route.

Figure 5: Clustered POIs.

Based on the clustered POIs, the route generation

process initiates a request to the ORS API to com-

pute the optimal route. Users can specify the start-

ing point either manually or by utilizing the center of

the largest cluster (the default by omission). Subse-

quently, the plugin selects waypoints from the largest

cluster, applies a greedy TSP solver to determine the

optimal order of waypoints, and generates the route

using ORS. The resulting route is then visualized in

QGIS (see Figure 6), providing users with an interac-

tive map display. To enhance the user experience, the

plugin includes a route animation feature, which reads

GPX data and animates the movement along the route

on the QGIS map canvas. Finally, it also supports ex-

porting the created route and layers to a PDF, as an

image.

Figure 6: Route visualization in QGIS.

The integration of clustering analysis and ad-

vanced route generation techniques in the second

phase represents a signiﬁcant advancement in devel-

oping intelligent route recommendation systems. By

allowing users to select POIs from a variety of cate-

gories and subcategories, the system provides highly

personalized and efﬁcient routing solutions. The use

of DBSCAN for clustering POIs based on real-world

distances ensures accurate and meaningful groupings,

while the ORS API facilitates the generation of op-

timized routes. The inclusion of a user-friendly in-

terface, interactive map displays, and features such

as route animation and export options enhances the

overall user experience, making the system a power-

ful tool for urban exploration and navigation.

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

396

7 EVALUATION

To assess the success of this work, the system is eval-

uated on functionality, user experience, performance,

clustering effectiveness, and accuracy. The system’s

ability to accurately recommend routes based on user-

selected cities and POI categories is assessed, as

well as the effectiveness of the route generation en-

gine in optimizing factors like distance (for phase

II) and available public transportation options (for

phase I), in real-time route recommendations. Each

project phase meets expectations, demonstrating ﬂex-

ibility and efﬁciency in using user inputs to rec-

ommend routes within concentrated areas of inter-

est. This aligns with the motivations described by

(Mahmood et al., 2023), who highlights the impor-

tance of context-aware recommendations and opti-

mized routes tailored to user preferences.

Regarding performance evaluation, the system

demonstrates high efﬁciency under varying loads. In

Phase I, a signiﬁcant number of POIs are retrieved

without delay. In Phase II, although fewer POIs are

processed, more complex operations are executed in

sequence (API calls followed by clustering analysis,

display, and data export) within a few seconds. This

real-time response capability underscores the practi-

cal application of advanced geospatial technologies

and algorithms in urban navigation, as discussed in

the introduction.

The reliability of clustering effectiveness in iden-

tifying concentrated areas of POIs and generating op-

timized routes within those zones is also assessed. To

ensure clustering effectiveness, two metrics are used

for validation: the Silhouette Score and the Davies-

Bouldin Index. These metrics provide quantitative

evaluations of clustering quality, conﬁrming that the

system effectively identiﬁes meaningful clusters of

POIs. However, some challenges are encountered in

clustering effectiveness in phase II, particularly with

results suggesting an overlap of clusters when using

the same parameters as in phase I. This can be visual-

ized in the layers and in the Silhouette Score with val-

ues ranging from -0.7 to -0.4 and the Davies-Bouldin

Index with values from 1 to 2 or 3, depending on the

parameter values. This cluster overlap could be at-

tributed to the presence of various sources and cat-

egories for the POIs. In the ﬁrst step, all the re-

trieved POIs are used for the clustering process, and in

the second step, only a few categories are processed.

Nonetheless, the overall results are promising. This

ﬁnding corroborates the work of (Lou, 2022), who

emphasizes the importance of accurate clustering for

intelligent tourism route planning.

The system’s clustering process beneﬁts from the

use of the DBSCAN algorithm, known for its robust-

ness in handling spatial data and noise, as noted by

(Lou, 2022). The application of DBSCAN, along

with the conversion of Euclidean distance matrices

into network distance matrices using Dijkstra’s algo-

rithm, allows for accurate clustering based on real-

world distances. This approach is consistent with the

ﬁndings of (Zhang et al., 2022), who highlights the

importance of accurate distance measurements and

intelligent technology in improving travel time pre-

dictions and route optimization.

In terms of user experience, the development of

a custom QGIS plugin

,“Optimal Custom Route”,

provides an intuitive and efﬁcient tool for route plan-

ning and visualization. The user interface, designed

using PyQt5, offers a seamless and interactive expe-

rience for selecting POIs and generating routes. The

integration of clustering analysis, route optimization,

and visualization within the QGIS environment en-

hances the system’s usability and practicality, align-

ing with the envisioned AmI principles of creating

smart environments that enhance daily living experi-

ences (Thankachan, 2023).

In conclusion, the application achieves its primary

goals of generating optimal routes that connect dif-

ferent POIs within selected cities, demonstrating high

accuracy in visualization and spatial analysis results.

By identifying areas with high concentrations of POIs

(in both phases) and public transportation coverage

(in Phase I), the system successfully provides person-

alized and efﬁcient navigation solutions. Future work

will focus on enhancing clustering techniques, inte-

grating real-time data, optimizing performance, incor-

porating user feedback, and expanding the range of

POIs categories to further improve the system’s func-

tionality and applicability.

8 CONCLUSIONS

This study successfully integrates AmI principles,

Python programming, and GIS to develop intelligent

route recommendation systems for urban exploration.

The developed systems optimize travel routes, pro-

mote sustainable transportation options, and enhance

user experiences. The research demonstrates the po-

tential of AmI to address challenges in urban naviga-

tion and personalized recommendations, contributing

to sustainable urban mobility.

The development process is divided into two

phases. In the ﬁrst phase, existing QGIS tools

and plugins are utilized to manually create and ana-

https://github.com/AgataPalma/OptimalCustomRoute

Route Recommendation Based on POIs and Public Transportation

397

lyze routes based on POIs and public transportation

data. This phase demonstrates the feasibility of us-

ing geospatial technologies to optimize urban naviga-

tion. The second phase involves the creation of a cus-

tom QGIS plugin, “Optimal Custom Route,” provid-

ing an automated and user-friendly interface for route

planning and visualization. This phase leverages ad-

vanced machine learning techniques, speciﬁcally DB-

SCAN clustering, to identify dense areas of POIs and

generate optimized routes.

The system’s performance is evaluated based on

functionality, user experience, performance, cluster-

ing effectiveness, and accuracy. The results indi-

cate that the system accurately recommends routes

based on user-selected cities and POI categories, ef-

ﬁciently handles varying loads, and generates well-

deﬁned clusters of POIs. However, some challenges

are encountered, particularly in clustering effective-

ness when dealing with different sources and cate-

gories of POIs, which will need further reﬁnement.

8.1 Future Work

While the current system shows promising results,

several areas for future work can enhance its func-

tionality and applicability:

• Enhanced Clustering Techniques. Future re-

search could explore more advanced clustering al-

gorithms and parameter tuning to improve cluster-

ing effectiveness, particularly when dealing with

diverse categories of POIs.

• Integration with Real-time Data. Incorporating

real-time data from public transportation systems,

trafﬁc conditions, and user location can enhance

the system’s ability to provide dynamic and real-

time route recommendations.

• Extended POI Categories: Expanding the range

of POI categories and integrating additional data

sources can provide more comprehensive and per-

sonalized route recommendations.

• Mobile Application Development. Developing a

mobile app of the system can make it more acces-

sible to users on the go, providing seamless and

interactive route recommendations.

• Sustainability Metrics. Incorporating sustain-

ability metrics, such as carbon footprint reduc-

tion and energy efﬁciency, into the route optimiza-

tion process can further promote sustainable ur-

ban mobility solutions.

In conclusion, this research demonstrates the sig-

niﬁcant potential of integrating AmI, Python pro-

gramming, and GIS in developing intelligent route

recommendation systems. By addressing the iden-

tiﬁed challenges and exploring future research di-

rections, ongoing advancements in AmI technologies

and their practical applications can continue to im-

prove urban living standards and mobility solutions.

REFERENCES

Gong, Q., Liu, J., and Cao, H. (2020). Csq system: A sys-

tem to support constrained skyline queries on trans-

portation networks. In Proc. of Intl. Conf. on Data

Engineering, volume 2020-April, pages 1746–1749.

IEEE Computer Society.

Karantaidis, G., Sarridis, I., and Kotropoulos, C. (2021).

Adaptive hypergraph learning with multi-stage opti-

mizations for image and tag recommendation. Signal

Processing: Image Communication, 97.

Li, J., Zhang, L., Ni, C., An, Y., Zong, C., and Zhang, A.

(2021). Efﬁcient k nearest neighbor query processing

on public transportation network. In Proc. of 20th Intl.

Conf. on Trust, Security and Privacy in Computing

and Communications, TrustCom, pages 1108–1115.

IEEE.

Lou, N. (2022). Analysis of the intelligent tourism route

planning scheme based on the cluster analysis algo-

rithm. Computational Intelligence and Neuroscience,

2022:3310676.

Mahdi, A. J., Tettamanti, T., and Esztergar-Kiss, D. (2023).

Modeling the time spent at points of interest based on

google popular times. IEEE Access, 11:88946–88959.

Mahmood, M. R., Kaur, H., Kaur, M., Raja, R., and Khan,

I. A. (2023). Ambient intelligence and internet of

things: An overview, chapter 1, pages 1–32. John Wi-

ley & Sons, Inc., 1 edition.

Panagiotakis, C., Daskalaki, E., Papadakis, H., and

Fragopoulou, P. (2022). The tourist trip design

problem with poi categories via an expectation-

maximization based method.

Sarridis, I., Karantaidis, G., and Kotropoulos, C. (2022).

Image driven optimal personalized route recommen-

dation. In IEEE IVMSP 2022. Institute of Electrical

and Electronics Engineers Inc.

Thankachan, D. (2023). Introduction to Ami and IoT, chap-

ter 12, pages 361–381. John Wiley & Sons, Inc., 1

edition.

Zhang, X., Lauber, L., Liu, H., Shi, J., Xie, M., and Pan, Y.

(2022). Travel time prediction of urban public trans-

portation based on detection of single routes. PLoS

ONE, 17:e0262535.

Ozcan, S. and Kaya, H. (2018). An analysis of travelling

salesman problem utilizing hill climbing algorithm for

a smart city touristic search on openstreetmap (osm).

In 2nd Intl Symposium on Multidisciplinary Studies

and Innovative Technologies (ISMSIT), pages 1–5.

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

398