Over the last decades Internet has spread dramatically and online advertising has become one of the major channels through which companies can advertise and sell their products and services. The potentialities of digital advertising are huge, and must be thoroughly studied in order to be exploited as much as possible. Disciplines such as Artificial Intelligence and Machine Learning can give a significant contribute, providing the tools to automate the management of internet advertising campaigns and optimize their performances. In this work we investigate how we can efficiently learn from logged bandit feedback in order to improve advertising campaigns' performances, and we propose three algorithms (KnapSack Tree, Lower Bound Tree Search and Multiple Split KnapSack) which are able to select the optimal target for each campaign. In particular, we focus on the choice of the optimal timing for showing an ad and we show how this choice can significantly increase the revenue obtained from a campaign. We tested our methods both on synthetic and real-world data, achieving very satisfactory results, since they lead to a consistent increment on the number of expected conversions with respect to the strategy currently adopted by advertisers.
Negli ultimi anni Internet si è diffuso enormemente e le pubblicità online sono diventate il mezzo più comune attraverso cui le aziende pubblicizzano e vendono i loro prodotti e servizi. La pubblicità digitale ha un potenziale enorme, ma richiede metodi automatici per essere sfruttata al meglio, data la complessità di Internet. Alcune delle tecniche sviluppate dai campi dell'Intelligenza Artificiale e del Machine Learning possono dare un contributo significativo in questo senso, fornendo strumenti utili per automatizzare la gestione delle campagne pubblicitarie online e per ottimizzare le loro prestazioni. In questa tesi affronteremo il problema di come sia possibile apprendere efficacemente da dati storici prodotti da algoritmi che operano in un ambiente di tipo bandit, allo scopo di migliorare le prestazioni delle campagne pubblicitarie online. Inoltre, proporremo tre algoritmi (KST, LBTS, MSKS) per selezionare il target ottimo per una campagna. In particolare, ci concentreremo sulla scelta delle fasce orarie migliori in cui mostrare gli annunci pubblicitari durante il giorno. Questi algoritmi sono stati applicati sia in esperimenti con dati sinteticamente generati sia su dati reali, producendo risultati molto soddisfacenti: l'applicazione di LBTS e MSKS ha portato ad un aumento significativo nel numero di conversioni attese rispetto al caso in cui venga adottata la strategia usuale.
Learning from logged bandit feedback techniques for targeting optimization of online advertising
GASPARINI, MARGHERITA
2016/2017
Abstract
Over the last decades Internet has spread dramatically and online advertising has become one of the major channels through which companies can advertise and sell their products and services. The potentialities of digital advertising are huge, and must be thoroughly studied in order to be exploited as much as possible. Disciplines such as Artificial Intelligence and Machine Learning can give a significant contribute, providing the tools to automate the management of internet advertising campaigns and optimize their performances. In this work we investigate how we can efficiently learn from logged bandit feedback in order to improve advertising campaigns' performances, and we propose three algorithms (KnapSack Tree, Lower Bound Tree Search and Multiple Split KnapSack) which are able to select the optimal target for each campaign. In particular, we focus on the choice of the optimal timing for showing an ad and we show how this choice can significantly increase the revenue obtained from a campaign. We tested our methods both on synthetic and real-world data, achieving very satisfactory results, since they lead to a consistent increment on the number of expected conversions with respect to the strategy currently adopted by advertisers.File | Dimensione | Formato | |
---|---|---|---|
2017_12_Gasparini.pdf
accessibile in internet per tutti
Descrizione: Testo della tesi
Dimensione
1.27 MB
Formato
Adobe PDF
|
1.27 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/10589/137330