以評價為基礎之協同過濾

Abstract

以商品為基礎的協同過濾系統是一種有名且效能傑出的協同過濾系統。但以商品為基礎的協同過濾系統在遇到打愈多種不同分數的使用者的情況下效能會愈差；在遇到訓練資料不足的情況下效能也會愈差。為了解決此問題，我們提出了一個以評價為基礎的新方法。在此方法的第一步中，對於每一個使用者，將被評價為相同分數的商品群聚在一起。接著，對於每一種評價都為其建立一個預測模型。最後，我們計算出每一個模型之機率期望值，並且藉此來預測評價。藉由此種以評價為基礎的方法，預測之評價是利用每一個評價之模型產生，而非單純使用要被預測商品之相似商品來預測。我們的此種方法在MovieLens的一百萬筆評價資料集中表現得非常好。實驗結果也呈現出我們的方法比傳統的以商品為基礎的協同過濾的方法好，且有達到統計之顯著性。

Item-based collaborative filtering (CF) recommender system is one of famous and well-performed collaborative filtering recommender system. Item-based CF suffers from the problem of various ratings, while users give more different ratings. It also suffers from the problem of insufficient training data. In order to deal with these problems, we propose new methods called rate-based. In the first step, for each user, cluster items with the same rating. Then, build a model for each rating. Finally, make predictions of ratings by calculating the expectation value of models. Through our rate-based methods, predictions are made by utilizing the models of each rating rather than the neighbors of items which are going to be predicted. Our rate-based methods perform great on the million dataset of MovieLens. The experiment results show that our methods outperform the conventional item-based CF with statistically significant.