Gručenje oglasov s pomočjo globokih nevronskih mrež

Džubur, Benjamin (author), Demšar, Jure (mentor)

V oglaševalski industriji je razumevanje različnih parametrov oglaševalskih kampanj ključnega pomena za optimizacijo delovnega procesa. Eden izmed ključnih parametrov je število uporabljenih glavnih dizajnov za pripravo slikovnih oglasov, na podlagi katerega lahko sklepamo o kompleksnosti kampanje. Oglasi, ki pripadajo istemu glavnemu dizajnu, pogosto vsebujejo podobne tipografije besedila in grafične elemente, včasih pa tudi kompozicije. V eksperimentalnem delu diplomskega dela za namen napovedovanja števila uporabljenih glavnih dizajnov v množicah slikovnih oglasov razvijemo dva napovedna modela. Oba temeljita na konvolucijskih nevronskih mrežah za pridobitev značilk iz slik in na algoritmih za gručenje podatkov. Razlikujeta se predvsem po načinu določanja podobnosti med posameznimi oglasi. Oba razvita napovedna modela dosežeta boljše rezultate od izhodiščnega pristopa, ki na podlagi porazdelitve podatkov naključno napove število glavnih dizajnov. Napovedna modela na vzorcu 50 kampanj dosežeta 5,2% oz. 1,2% izboljšavo v klasifikacijski točnosti. Drugi napovedni model, ki temelji na podobnosti regij med oglasi, dosega kvalitativno boljše rezultate od prvega, ki temelji na enostavnih primerjavah značilk celotnih oglasov.

• Language: Slovenian
• Keywords: konvolucijska nevronska mreža, globoko gručenje, segmentacija slik, gručenje slik, nenadzorovano učenje
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2020
• PID: 20.500.12556/RUL-117764
• COBISS.SI-ID: 24100611
• Publication date in RUL: 24.07.2020

Secondary language

• Language: English
• Title: Clustering adverts with support of deep neural networks

Abstract

In the advertising industry the understanding of different advertising campaigns' parameters is key for workflow optimization. One of these parameters is the number of master designs used to prepare image based adverts, which is a crucial for determining the complexity of a campaign. Adverts which originate from the same master design typically use similar typographies, graphical elements and compositions. In the experimental part of this thesis, we develop two predictive pipelines for the task of predicting the number of master designs in sets of image based adverts. Both pipelines use convolutional neural networks for feature extraction and clustering algorithms. The main difference between the two is in the way that the similarity between individual adverts is computed. Both developed models achieve better results than our baseline approach which, based on the distribution of data, randomly predicts the number of master designs. Our predictive models achieve a 5.2% and 1.2% classification accuracy improvement respectively over the baseline when tested on a sample of 50 campaigns. Our second model, which is based on the similarity of regions between adverts, achieves qualitatively better results than our first model, which is based on simple comparisons of the adverts' features.

• Keywords: convolutional neural network, deep clustering, image segmentation, image clustering, unsupervised learning