摘要
推荐系统(RS)帮助用户从互联网上提供的众多项目中挑选相关项目。物品可能是电影、食物、书籍等。推荐系统利用从用户获取的数据来生成推荐。通常，这些评级可能是显性的或隐性的。显式评级是一般在1到5之间的绝对评级，而隐式评级则来自购买历史、点击率、查看历史等信息。偏好关系是表示用户对项目兴趣的另一种方式。最近很少有研究表明，偏好关系比绝对评级产生更好的结果。此外，在RS中，矩阵分解(MF)等潜在因子模型给出了精确的结果，特别是在数据稀疏的情况下。欧几里得嵌入(EE)是另一种潜在因子模型，其结果与MF相似。在这项工作中，我们提出了一个欧几里得嵌入偏好关系的推荐系统。而不是使用项目与用户潜在因素的内积，而是使用项目与用户潜在因素之间的欧氏距离来预测评分。矩阵分解的偏好关系(MFPR)比传统的矩阵分解产生了更好的推荐。在这项工作中，我们提出了一个名为EEPR的协作模型。在两个真实数据集MovieLens-100K和x-1m上实现和测试,验证了该方法的有效性。在推荐系统中，我们使用了流行的评价指标precision@k。实验结果表明，该模型优于现有的MF、EE和MFPR模型。
Abstract
Recommender systems (RS) help users pick the relevant items among numerous items that are available on the internet. The items may be movies, food, books, etc. The Recommender systems utilize the data that is fetched from the users to generate recommendations. Usually, these ratings may be explicit or implicit. Explicit ratings are absolute ratings that are gener- ally in the range of 1 to 5. While implicit ratings are derived from information like purchase history, click-through rate, viewing history, etc. Preference relations are an alternative way to represent the users' interest in the items. Few recent research works show that preference relations yield better results compared to absolute ratings. Besides, in RS, the latent fac- tor models like Matrix Factorization (MF) give accurate results especially when the data is sparse. Euclidean Embedding (EE) is an alternative latent factor model that yields similar results as MF. In this work, we propose a Euclidean embedding with preference relation for the recommender system. Instead of using the inner product of items and users' latent factors, Euclidean distances between them are used to predict the rating. Preference Relations with Matrix Factorization (MFPR) produced better recommendations compared to that of tradi- tional matrix factorization. We present a collaborative model termed EEPR in this work. The proposed framework is implemented and tested on two real-world datasets, MovieLens-100K and Netflix-1M to demonstrate the effectiveness of the proposed method. We utilize popular evaluation metric for recommender systems as precision@K. The experimental outcomes show that the proposed model outperforms certain state-of-the-art existing models such as MF, EE, and MFPR.
NETL
NSTL