融合电影流行性与观影时间的协同过滤算法
网络安全与数据治理
钱泽俊,刘润然
(杭州师范大学阿里巴巴商学院,浙江杭州311121)
摘要: 相似度评估作为协同过滤推荐算法的核心,尽管研究人员对其不断改进,却仍难以在各个维度上充分利用评价数据。针对这一挑战,首先以用户与电影之间的相互影响方式作为切入点,对二者间可能存在的自洽逻辑进行探究,提出了电影流行度计算公式用于对电影进行加权;接着以用户观影时间作为研究对象,探究用户观影喜好的转变与观影时间顺序之间的联系,并结合肯德尔相关系数提出了观影顺序一致性度量公式;最后将以上研究内容与传统相似度算法融合,并基于Netflix Prize数据集与豆瓣电影评价数据集对改进后的相似度算法进行验证。实验结果表明改进后的相似度算法拥有更高的推荐准确度。
中圖分類號:TP3913文獻(xiàn)標(biāo)識碼:ADOI: 10.19358/j.issn.2097-1788.2024.02.009
引用格式:錢澤俊,劉潤然.融合電影流行性與觀影時(shí)間的協(xié)同過濾算法[J].網(wǎng)絡(luò)安全與數(shù)據(jù)治理,2024,43(2):54-63.
引用格式:錢澤俊,劉潤然.融合電影流行性與觀影時(shí)間的協(xié)同過濾算法[J].網(wǎng)絡(luò)安全與數(shù)據(jù)治理,2024,43(2):54-63.
Collaborative filtering algorithm combining movie popularity and viewing time
Qian Zejun,Liu Runran
(Alibaba Business School, Hangzhou Normal University, Hangzhou 311121, China)
Abstract: As the core of the collaborative filtering recommendation algorithm, similarity evaluation is still difficult to fully utilize evaluation data in all dimensions, despite researchers constantly improving it. In this paper, aiming at this challenge, the mutual influence between users and movies is taken as the starting point, the possible self consistent logic between the two is explored, and a formula called Movie Popularity Weight (MPW) calculation formula is proposed to calculate the weight of movies. Then, taking the viewing time of users as the research object, the relationship between the change of viewing preference and the viewing time sequence of users is explored, and combined with the theory of Kendall correlation coefficient, a formula called Consistency in Viewing Sequence (CVS) calculation formula is proposed. Finally, the traditional similarity algorithm is improved by using the above research content, and the improved similarity algorithm is validated by using two datasets, one is the Netflix Prize dataset, while the other one is built based on publicly available data from Douban.com called Douban Movie K5 dataset. The experimental result shows that the improved similarity algorithm has higher recommendation accuracy.
Key words : recommendation algorithm; collaborative filtering; similarity algorithm; movie popularity; viewing time
引言
推薦系統(tǒng)[1]是人們借助計(jì)算機(jī)系統(tǒng)的高計(jì)算能力,為解決用戶在面對信息過載時(shí)獲取有效信息的效率低下問題而設(shè)計(jì)的輔助系統(tǒng),其準(zhǔn)確性極大程度上依賴于所采用的推薦策略。在推薦系統(tǒng)的眾多策略中,“協(xié)同過濾”是其中廣泛使用的一種策略[2],它以用戶的興趣偏好作為推薦依據(jù),并假設(shè)每個(gè)用戶未來的行為更有可能與該用戶過去的行為類似。因此,以協(xié)同過濾策略為基礎(chǔ)的推薦系統(tǒng),會基于與目標(biāo)用戶相似的其他用戶對一些物品的評價(jià)來向目標(biāo)用戶推薦物品[3],具有良好的可解釋性。協(xié)同過濾策略的關(guān)鍵步驟是計(jì)算用戶間的相似度,但由于傳統(tǒng)的相似度算法很容易受到冷啟動(dòng)、數(shù)據(jù)稀疏性、時(shí)間衰變等問題的影響[4],因此許多研究人員對此進(jìn)行改進(jìn)并提出了一些新的相似性度量算法。在研究物品的權(quán)值計(jì)算方面,Leskovec[5]等人對Pearson相關(guān)系數(shù)算法的改進(jìn)考慮到評價(jià)的分布具有長尾特征,即隨著時(shí)間的流逝,部分受歡迎的物品將會得到更多用戶的評價(jià),而一些不受歡迎的物品,它們得到的評價(jià)數(shù)量則一直非常有限。
作者信息:
錢澤俊,劉潤然
(杭州師范大學(xué)阿里巴巴商學(xué)院,浙江杭州311121)
文章下載地址:http://ihrv.cn/resource/share/2000005903

此內(nèi)容為AET網(wǎng)站原創(chuàng),未經(jīng)授權(quán)禁止轉(zhuǎn)載。
