Napovedovanje optimalnega povezanega spletnega članka z MAB pristopom

Cilj magistrske naloge je bil zasnovati in preizkusiti inovativen pristop za izbiro optimalne spletne novice glede na trenutno brano novico. Pomemben faktor pri izbiri našega pristopa je bilo dejstvo, da velikokrat v realnem scenariju nimamo podatkov o uporabnikih in njihovih preferencah. Zaradi tega smo za reševanje omenjenega problema uporabili Multi-armed bandit pristop. Primerjavo različnih MAB algoritmov smo naredili na problemu Jackpot, ki je bil zasnovan v sklopu Celtrinega programerskega izziva. Osnovni problem pa smo simulirali in evalvirali na realnih podatkih enega izmed ponudnikov priporočil povezane vsebine. Obstoječ priporočilni sistem smo poskušali izboljšati tako, da smo prerazporedili podobne novice trenutno brani novici. Na ta način so predhodno najboljše priporočene novice bile priporočene najvišje, hkrati pa smo vseskozi raziskovali potencialno še boljše novice. V magistrski nalogi smo preverili, ali je smiselno upoštevati pod katero novico je priporočena novica prikazana, ali je možno aproksimirati stopnjo odzivnosti s pomočjo vsebinsko sorodnih priporočilnih tokov in ali dosežemo boljše rezultate, če z Bayevsovskim pristopom aproksimiramo začetne vrednosti parametrov beta distribucije. Naš najbolj obetaven pristop je povprečno pozicijo priporočila izboljšal za približno 40% glede na naključno razporeditev vsebinsko sorodnih priporočil. The main purpose of this master thesis was to develop and evaluate innovative approach for selection of optimal related web article. We often do not have access to data about user's profiles and their preferences. Therefore, these settings have an important influence to our research. Consequently, we decided to use Multi-Armed bandit approach to deal with described settings. Comparison of different MAB algorithms has been done on Jackpot problem, which was designed for Celtra programming challenge. The main research problem has been simulated and evaluated on a real data obtained from a provider of related web content. We tried to improve existing recommendation system with reordering similar news to currently read one. Using this approach, the most interesting news have been recommended on top positions and the most promising news have been explored as well. In this master thesis, we tried to answer to the following questions: does it make sense to take into account statistics of recommended news in context of every news stream separately is it possible to approximate click-through rate using content-related recommendation streams could we achieve better results, if we approximate initial input parameters of beta distribution using Bayesian approach. Our most promising method has achieved more than 40% average position improvement regarding random selection strategy of content-related recommendations.