¡Descarga Analysis of co-movement pattern mining methods for recommendation y más Tesis en PDF de Optimización de Motores de Búsqueda y Publicidad (SEO y SEM) solo en Docsity! Analysis of co-movement patternminingmethods for recommendation Extended Abstract Sergio Torrijos Universidad Autónoma de Madrid Madrid, Spain sergio.torrijos@estudiante.uam.es Alejandro Bellogín Universidad Autónoma de Madrid Madrid, Spain alejandro.bellogin@uam.es ABSTRACT Location-Based Social Networks allow users to share the Points- of-Interest they visit, hence creating trajectories throughout their usual lives – even though they are also used by tourists to explore a city. There exist several algorithms in the trajectory pattern mining area able to discover and exploit interesting patterns from trajectory data, such as which objects tend to move together (co-movement), however, to the best of our knowledge, they have not been usedwith data coming from that type of systems. In this work, we analyse the extent to which these techniques can be applied to that type of data and under which circumstances they might be useful. 1 INTRODUCTIONANDBACKGROUND The aim of Recommender Systems (RS) is to help users in finding relevant items,usuallybyfiltering largecataloguesand taking intoac- count theusers’ preferences.CollaborativeFiltering (CF) systemscan be considered as the earliest and most widely deployed recommen- dation approach [9], suggesting interesting items to users based on the preferences from “similar” or related people [14]; although other types of recommendation algorithms exist, such as content-based systems and social filtering, among the most popular ones [7, 13]. Because of the increasing number of users registered in Location- BasedSocialNetworks (LBSNs),whereusers share thevenues, places, orPoints-of-Interest (POI) theyvisit;POIrecommendationapproaches have become particularly useful and several specific techniques have been proposed in recent years. In particular, such methods tend to incorporate inherent properties of these systems, such as social, geographical, or temporal information [11, 12]. A recent trend in the literature is to exploit trajectory patternmin- ing methods for recommendation. Some examples include creating recommenders for itineraries based on geo-tagged photos [3] and, in a more general context, exploiting the GPS trajectories left by the users to suggest interesting locations [4, 20]. However, there are still several trajectory pattern mining methods that have not been used for recommendation yet. More specifically, in this extended abstract we discuss the possibility to integrate user mobility patterns for rec- ommendation, inparticular, andas afirst solution, to compute similar users (according to the detected trajectory patterns) as a straightfor- ward way to include such information into well-known recommen- dation algorithms. We do this by focusing on the techniques known as co-movement pattern methods and those derived by them. In fact, the literature on trajectory pattern mining provides sev- eral methods to compute similarities either between trajectories or, "Copyright© 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0)." Figure 1: Check-ins collected from twousers (one in blue and theother ingreen) in the cityofNewYork.Note that thereare manymore places in common in the center ofManhattan. in a more general context, moving objects, for instance, to cluster or identify them in order to track their positions or analyse their be- haviour [21]. Examples of applications and services could be animal analysis, social recommendations, and traffic planning, where data would arrive from chips for animal telemetry, wearable devices, or vehicles with GPS [6]. More specifically, many of these problems are solved under the umbrella of discovering co-movement patterns, which refers to finding groups of objects travelling together for a certain period of time, hence exploiting both the temporal and spa- tial dimensions. There exist several co-movement pattern methods, such Flock, Convoy, Swarm, or ST-DBSCAN [6, 15]. They all impose different constraints on what a group or a co-movement is when solving the problem, and hence they evidence different complexities, this is why parallel and scalable solutions are needed [6]. In the next sections, we discuss some potential applications of these methods for recommendation and how to address the scalability issues. 2 EXPLOITING TRAJECTORY PATTERNS FORRECOMMENDATION A simple approach to exploit some of the most common trajectory patternminingmethods such as querying, indexing, or retrieval (i.e., similarity metrics between trajectories) [5] is to perform some kind of filtering instead of a pure recommendation task, that is, to present the user the most similar trajectory(ies) with respect to her previous trajectories, her last query, or her inferred tastes. A paradigmatic example of this type of filtering is the well-known “customers who 1
