Impression-aware recommender systems (IARS) are innovative recommender systems that leverage a novel data type to learn users' preferences toward items in online systems. Such data type is called impression, a collection of items shown to a user on-screen and generated by an online system. Traditionally, recommender systems leverage user's actions with the system, such as views or clicks, to learn user preferences. IARS take a comprehensive approach to achieve the same goal. Unlike actions, impressions represent the choices the recommender system provides users at any time. IARS, then, join users' actions and shown impressions to make accurate and serendipitous recommendations. Consequently, IARS, in conjunction with impressions, enable the design and development of novel and more refined recommendation techniques. Nowadays, many open research directions fundamental for properly developing IARS still need to be answered. This Thesis addresses three directions: characterization, evaluation, and experimentation. The first direction, characterization, addresses the need for sound mathematical foundations to define and describe IARS. In the Thesis, we design and propose a theoretical and mathematical framework to define impressions and IARS. Under this framework, we present a novel classification system for IARS comprised of three taxonomies and three design properties of recommendation models that learn from impressions. Additionally, we comprehensively review existing recommendation models proposed in previous research works. The second direction, evaluation, addresses the need for standard evaluation methodologies and procedures to assess the recommendation quality of IARS. In the Thesis, we present, describe, and analyze a novel dataset with impressions suitable for evaluating IARS: ContentWise Impressions. We devise a standard evaluation framework for IARS, comprised of guidelines to process and clean datasets with impressions, to define recommendation tasks and goals, to define a set of baseline recommenders, and to search the optimal set of hyper-parameters. The third direction, experimentation, studies IARS from a practical perspective, i.e., by performing experiments and evaluating their recommendation quality. In the Thesis, we assess whether the reviewed papers provide sufficient and necessary descriptions and tools to reproduce or replicate their results. Then, we describe two evaluation studies on IARS, one aiming at replicating the results achieved by the literature, while the other aiming to incorporate impressions into graph-based recommender systems in practical and simple approaches to improve recommendation quality.
I sistemi di raccomandazione basati sulle impressioni (IARS) sono sistemi di raccomandazione innovativi che sfruttano un nuovo tipo di dati per imparare le preferenze degli utenti verso gli articoli nei sistemi online. Questo tipo di dati è chiamata impressione, una collezione di elementi mostrati all'utente sullo schermo generati da un sistema online. Tradizionalmente, i sistemi di raccomandazione sfruttano le azioni dell'utente nel sistema, come visualizzazioni o clic, per imparare le preferenze dell'utente. Gli IARS adottano un approccio più ampio per raggiungere lo stesso obiettivo perché uniscono le azioni degli utenti e le impressioni mostrate per fornire raccomandazioni. A differenza delle azioni, le impressioni rappresentano le scelte che il sistema di raccomandazione fornisce agli utenti in qualsiasi momento. Molte direzioni di ricerca fondamentali per sviluppare adeguatamente gli IARS rimangono ancora aperte. Questa Tesi tratta tre direzioni: la caratterizzazione, valutazione e sperimentazione degli IARS. La prima direzione, la caratterizzazione, risponde alla necessità di basi matematiche per definire e descrivere gli IARS. Nella Tesi proponiamo un quadro teorico e matematico per definire gli impressioni e gli IARS. In questo quadro, presentiamo un nuovo sistema di classificazione per gli IARS composto da tre tassonomie e tre proprietà dei modelli di raccomandazione. Inoltre, esaminiamo e classifichiamo i modelli di raccomandazione esistenti e proposti in precedenti lavori di ricerca. La seconda direzione, la valutazione, risponde alla necessità di metodologie e procedure standard per valutare la qualità delle raccomandazioni degli IARS. Nella tesi presentiamo, descriviamo e analizziamo un nuovo set di dati con impressioni adatto alla valutazione degli IARS: ContentWise Impressions. Anche, elaboriamo un quadro di valutazione standard per IARS, composto da orientamenti per processare e pulire set di dati con impressioni, per definire obiettivi di raccomandazione e una collezione di modelli di raccomandazione di base e per cercare degli iperparametri ottimale dei modelli di raccomandazione. La terza direzione, la sperimentazione, studia gli IARS da una prospettiva pratica, eseguendo esperimenti e valutando la qualità delle loro raccomandazioni. Nella Tesi, valutiamo se i lavori di ricerca forniscono descrizioni e strumenti sufficienti e necessari per riprodurre o replicare i loro risultati. Dopo questa valutazione descriviamo due studi che descrivono degli esperimenti sugli IARS, uno studio con l'obbiettivo a replicare i risultati raggiunti dalla letteratura, mentre l'altro con l'obbiettivo a incorporare le impressioni nei sistemi di raccomandazione basati su grafi in approcci pratici e semplici per migliorare la qualità delle raccomandazioni.
Impression-aware recommender systems
Perez Maurera, Fernando Benjamin
2023/2024
Abstract
Impression-aware recommender systems (IARS) are innovative recommender systems that leverage a novel data type to learn users' preferences toward items in online systems. Such data type is called impression, a collection of items shown to a user on-screen and generated by an online system. Traditionally, recommender systems leverage user's actions with the system, such as views or clicks, to learn user preferences. IARS take a comprehensive approach to achieve the same goal. Unlike actions, impressions represent the choices the recommender system provides users at any time. IARS, then, join users' actions and shown impressions to make accurate and serendipitous recommendations. Consequently, IARS, in conjunction with impressions, enable the design and development of novel and more refined recommendation techniques. Nowadays, many open research directions fundamental for properly developing IARS still need to be answered. This Thesis addresses three directions: characterization, evaluation, and experimentation. The first direction, characterization, addresses the need for sound mathematical foundations to define and describe IARS. In the Thesis, we design and propose a theoretical and mathematical framework to define impressions and IARS. Under this framework, we present a novel classification system for IARS comprised of three taxonomies and three design properties of recommendation models that learn from impressions. Additionally, we comprehensively review existing recommendation models proposed in previous research works. The second direction, evaluation, addresses the need for standard evaluation methodologies and procedures to assess the recommendation quality of IARS. In the Thesis, we present, describe, and analyze a novel dataset with impressions suitable for evaluating IARS: ContentWise Impressions. We devise a standard evaluation framework for IARS, comprised of guidelines to process and clean datasets with impressions, to define recommendation tasks and goals, to define a set of baseline recommenders, and to search the optimal set of hyper-parameters. The third direction, experimentation, studies IARS from a practical perspective, i.e., by performing experiments and evaluating their recommendation quality. In the Thesis, we assess whether the reviewed papers provide sufficient and necessary descriptions and tools to reproduce or replicate their results. Then, we describe two evaluation studies on IARS, one aiming at replicating the results achieved by the literature, while the other aiming to incorporate impressions into graph-based recommender systems in practical and simple approaches to improve recommendation quality.File | Dimensione | Formato | |
---|---|---|---|
FBPM - Impression-Aware Recommender Systems - PhD Thesis.pdf
accessibile in internet per tutti
Descrizione: PhD Thesis
Dimensione
1.54 MB
Formato
Adobe PDF
|
1.54 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/221112