基於多核學習之球鞋喜愛度預測方法

Abstract

在此研究中，我們藉由多個不同類型的資訊來描述球鞋產品，以達到預測消費者對於球鞋喜愛度之目的。透過使用Sole Collector球鞋網站上所提供之資料建置了1913雙球鞋之資料庫，資料包含了球鞋圖片、名稱、及價錢等，並利用其當作訓練集與測試集。對於球鞋產品，本工作做了不同面向的描述，擷取多種特徵，並利用機器學習中的多核學習(Multiple Kernel Learning, MKL)方法結合多個適合各自特徵空間的前計算核(Pre-computed Kernel)，藉由這些前計算核的線性組合訓練出球鞋喜愛度分數預測模型，透過新球鞋之特徵資訊當作輸入，輸出球鞋受消費者喜愛度的預測分數。實驗部分則提供了多核學習與支持向量回歸(Support Vector Regression, SVR)兩方法比較，結果顯示相同核數下，使用後期融合方法(MKL)較前期融合方法(SVR)在預測球鞋喜愛度問題上，有較佳的關聯性。而使用前計算核廣泛上擁有較徑向基函數核(Radial basis function kernel)更好的表現。In this paper, we explore using different sources of information to describe sneakers based on which we predict the popularity of sneakers. In particular, we utilize Multiple Kernel Learning (MKL) techniques to train a popularity prediction model consisting of multiple pre-computed kernels that were suited to their own feature space. We further evaluate two regression models: support vector regression (SVR) and MKL to predict a popularity score from several features. In the experiment, we collected 1913 sneakers images from Sole Collector for building a dataset for both training the model and evaluating the prediction accuracy of the approach. We show that a late fusion method (i.e. MKL) has a better performance than an early fusion method (i.e. SVR) in predicting the popularity of sneakers. The pre-computed kernels have better result than radial basis functionkernel generally.