Primerjava metode Gradient boosting in nevronskih mrež z Uberjevo arhitekturo za napovedovanje prodaje

DEBELAK, JAN (author), Možina, Martin (mentor)

Diplomska naloga govori o razlikah med metodo Gradient boostinga z regresijskimi drevesi (GBRT) in Uberjevimi nevronskimi mrežami pri napovedovanju prodaje. Napovedovanje prodaje je pomembno zaradi vedno večjega števila izdelkov in omejenega prostora v skladiščih. Obstaja že dolga zgodovina napovedovanja prodaje, ki se razdeli na več dob. Najnovejša med njimi je doba napovedovanja s strojnim učenjem, kamor spadata tudi obravnavani metodi diplomske naloge. Metode so bile testirane na podatkih tekmovanja iz spletne strani Kaggle, kjer je bil cilj tekmovanja napovedati prodajo za 10 različnih Wallmartovih trgovin za obdobje 28 dni.Na splošno je bil bolj uspešen model, ki je uporabljal metodo GBRT, ki se je izkazala bolje na volatilnem in stabilnem obdobju in napovedala bolje, kot Uberjeva arhitektura LSTM (ULSTM). ULSTM se je izkazala bolje v krajših časovnih obdobjih in v enem, kljub splošni prevladi metode GBRT napovedala bolje. Pri večjih naborih atributov se je metoda GBRT izkazala za veliko hitrejšo od ULSTM, prav tako je bila velika razlila pri vplivu naključnega semena na rezultate, kar se je izkazalo za večji problem pri ULSTM. Rezultati testiranja so pokazali, da obstaja veliko razlogov zakaj je Gradient boosting veliko bolj uporabljen pri napovedovanju časovnih serij, kot LSTM.

• Language: Slovenian
• Keywords: LSTM, Gradient boosting, napoved prodaje, Uber
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2021
• PID: 20.500.12556/RUL-130172
• COBISS.SI-ID: 77313283
• Publication date in RUL: 10.09.2021

Secondary language

• Language: English
• Title: Comparison between Gradient Boosting and Neural Networks with Uber Architecture for Sales Forecasting

Abstract

The diploma thesis talks about the differences between Gradient boosting with regression trees (GBRT) and Uber neural networks in sales forecasting. Sales forecasting is important because of the ever-increasing number of products and limited space in warehouses. There is a long history of sales forecasts, which is divided into several periods. The latest among them is the period of forecasting through machine learning, which also includes the methods discussed methods. The methods were tested on data from a competition on Kaggle, where the goal of the competition was to predict sales for 10 Wallmart stores for a period of 28 days. In general, the more a successful method was GBRT method, which overall predicted better than Uber’s LSTM architecture (ULSTM). ULSTM performed better in shorter time periods and in one, despite the general dominance of the GBRT method, predicted better. For larger sets of attributes, the GBRT method proved to be much faster than ULSTM and there was also a large difference between the impact of random seed on the results, which proved to be a major problem with ULSTM. Test results showed that there are many reasons why Gradient boosting is much more used in forecasting time series such as LSTM.

• Keywords: LSTM, Gradient boosting, sales forecasting, Uber