izpis_h1_title_alt

Uporaba singularnega razcepa v priporočilnih sistemih : delo diplomskega seminarja
ID Furlan, Nika (Author), ID Plestenjak, Bor (Mentor) More about this mentor... This link opens in a new window

.pdfPDF - Presentation file, Download (2,33 MB)
MD5: A58DBEB0DB07C644BE798AD6650EBBA4

Abstract
Leta 2006 je Netflix razpisal tekmovanje, katerega cilj je bil izboljšanje takratnega algoritma za napoved ocen, s katerimi bi uporabniki ocenili filme. Podatke, ki so jih o uporabnikih, filmih in ocenah dobili tekmovalci, predstavimo z matriko, katere elementi predstavljajo uporabnikovo oceno filma. Tekmovalci so popularizirali numerično orodje, ki se uporablja za zmanjšanje obsežnih podatkov, imenovano singularni razcep. Ta uporabnike in filme preslika v skupni navidezni prostor razsežnosti $k$. Cilj priporočilnega sistema je, da uporabniku priporoča filme oziroma napove s kakšno oceno bi jih vrednotil. Oceno uporabnika $u$ za film $i$ izračunamo s skalarnim produktom med vektorjema $p_u$ in $q_i$. Prvi predstavlja uporabnikove preference za določene navidezne lastnosti, kot so na primer, kako všeč mu je določen žanr, igralci, lokacija snemanja, itd. Drugi pa v kolikšni meri $i$-ti film poseduje posamezne navidezne lastnosti -- na primer kateremu žanru pripada, kateri igralci so v njem igrali in kje je bil film posnet. Kot mero uspešnosti predlaganih algoritmov so na tekmovanju uporabili koren povprečne kvadratne napake, pri čemer je napaka definirana kot razlika med dejansko in napovedano oceno. Slednjo dobimo prek singularnega razcepa. V optimizacijski problem minimizacije korena povprečne kvadratne napake uvedemo regularizacijski element, s katerim preprečimo preprileganje. Problem rešimo s stohastičnim gradientnim spustom ali alternirajočimi najmanjšimi kvadrati.

Language:Slovenian
Keywords:Singularni razcep, priporočilni sistemi, koren povprečne kvadratne napake, stohastični gradientni spust, alternirajoči najmanjši kvadrati
Work type:Final seminar paper
Typology:2.11 - Undergraduate Thesis
Organization:FMF - Faculty of Mathematics and Physics
Year:2022
PID:20.500.12556/RUL-141295 This link opens in a new window
UDC:519.6
COBISS.SI-ID:125079043 This link opens in a new window
Publication date in RUL:28.09.2022
Views:951
Downloads:87
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:English
Title:Usage of singular value decomposition in recommender systems
Abstract:
In 2006, Netflix launched a contest to improve accuracy of movie recommendation algorithm. We represent the user, movie and rating data by a matrix whose elements represent the user's rating of the movie. The competitors popularised a numerical tool used to reduce large-scale data, called singular value decomposition. It is used to map users and movies into a common latent factor space of dimensionality $k$. The goal of the recommender system is to recommend movies to the user or to predict how they would rate them. The rating of a user $u$ for a movie $i$ is computed by the dot product between the vectors $p_u$ and $q_{i}$. The former represents the user's preferences for certain latent features of the movie, while the latter represents the extent to which the movie possesses those features. The root mean square error was used as a measure of the performance of the proposed algorithms, where the error is defined as the difference between the actual and the predicted rating. The latter is obtained via a singular value decomposition. The optimisation problem of minimising the root mean square error is solved by stochastic gradient descent or alternating least squares.

Keywords:Singular value decomposition, recommender systems, root mean square error, stochastic gradient descent, alternating least squares

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back