基于元学习的推荐算法选择优化框架实证  

Empirical study on recommendation algorithm selection optimization framework based on meta-learning

作　　者：任义[1] 迟翠容 单菁[1] 王佳英 REN Yi;CHI Cui-rong;SHAN Jing;WANG Jia-ying(Information and Control Engineering Faculty,Shenyang Jianzhu University,Shenyang 110168,China)

机构地区：[1]沈阳建筑大学信息与控制工程学院,辽宁沈阳110168

出　　处：《计算机工程与设计》2020年第6期1610-1616,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61702345、61702346)。

摘　　要：针对给定特征的数据集,选择最佳推荐算法存在计算资源相对过高、耗时较长、正确率较低的问题,提出一种基于元学习的推荐算法选择优化框架,在常用元特征的基础上融入新的特征测度。将Donorschoose和Movielens作为实证数据集,实验分析KNN、SVD等算法的自动选择过程,通过3种元学习算法构建元模型,评估该模型的预测正确率。Stac-kingDecisionTree元模型所选择推荐算法具有较高预测性能,预测正确率分别达到86.58%和80.39%,实验结果表明了提出框架的可行性。When choosing the most suitable recommendation algorithm for the given datasets,problems emerge,such as high occupancy of computing resources,long-time consuming and low accuracy.A recommendation algorithm selection optimized framework was proposed based on meta-learning was then proposed,which integrated new characteristics into the common meta-features.Donorschoose and Movielens were taken as the empirical datasets,the automatic selection process of KNN,SVD and ot-her algorithms was analyzed.The meta-model was constructed using three kinds of meta-learning algorithms aiming to evaluate the prediction accuracy.The recommendation algorithm with high predictive performance is selected using StackingDecisionTree meta-model,and the prediction accuracy is up to 86.58%and 80.39%respectively.The feasibility of the proposed framework is demonstrated.

关 键 词：推荐算法 算法选择优化框架 元学习 元模型 预测正确率 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]