不同相似度計算對於圖推薦系統之影響

Abstract

圖推薦模型非常適用於Top-N推薦系統，因為它們能夠捕獲實體之間的潛在關係。然而，大多數現有方法僅使用一個全域的物品圖(Item Graph)，被所有使用者共用，並未考慮到不同使用者之間偏好的差異。先前有研究者設計出了一種新穎的圖推薦模型，名為GLIMG(Global and Local IteM Graphs)，它可以同時捕獲全域和局部使用者的品味，實現更個性化的推薦。傳統GLIMG模型使用餘弦相似度來評估物品之間的相似性，由於不同相似度計算對於圖推薦系統之有所影響，因此在本研究中，我們在GLIMG模型原有的餘弦相似度計算基礎上探索是否存在更好的相似度計算方法來提升推薦系統的性能。 為保持實驗公平性，我們同樣使用Movielens-1M和Yelp2018資料集，並替換了GLIMG模型中的餘弦相似度計算方法，引入了三種不同的相似度計算方式，在模型中進行實驗，包括：歐氏距離、曼哈頓距離和調整後的餘弦相似度(Adjusted Cosine Similarity)。通過比較使用這些方法之後的各項推薦系統評估指標數值，我們發現，加上評分偏置項之後的調整後的餘弦相似度在Top-50的推薦任務上取得了最好的推薦效果。

Graph-based recommendation models are highly suitable for Top-N recommendation systems as they can capture latent relationships between entities. However, most existing methods only utilize a global item graph shared by all users, neglecting differences in preferences among different users. A novel graph recommendation model named GLIMG (Global and Local IteM Graphs) has been previously designed to concurrently capture the tastes of both global and local users, achieving more personalized recommendations. The traditional GLIMG model uses cosine similarity to assess item similarity, a crucial metric for graph-based recommendation models. In this study, we explore whether there are better similarity calculation methods to enhance the performance of the recommendation system, building upon the original cosine similarity calculation in the GLIMG model. To maintain experimental fairness, we employ the Movielens-1M and Yelp2018 datasets. We replace the cosine similarity calculation method in the GLIMG model with three different similarity calculation approaches – Euclidean distance, Manhattan distance, and adjusted cosine similarity. Through comparing various evaluation metrics of the recommendation system after incorporating these methods, we find that the adjusted cosine similarity, especially with the addition of rating bias, achieves the best recommendation performance in Top-50 recommendation tasks.