基於矩陣分解的雙向推薦系統

Abstract

推薦系統通常基於使用者的喜好而推薦物品給使用者，雙向推薦系統卻是將使用者推薦給另一名使用者，而且希望能使他們互相滿意對方而達到高的配對率。實際上的應用有線上求職或線上交友系統，這些網站嘗試幫他們的顧客配對但卻無法使用傳統的推薦系統來達成目的，因為傳統的推薦系統只能推薦使用者可能喜歡的列表，無法保證使用者也會被喜歡。為了解決這個問題，我們提出了一個基於矩陣分解模型的新方法並且實驗在兩個真實的資料集上：分別是線上交友網站(libimseti)以及線上求職網站(104人力銀行)。我們模型中的核心概念是調整MF模型中找區域最佳解的學習方式(SGD函式)，試著讓可能配對的雙方都出現在對方的推薦列表中。根據實驗結果，我們的模型在AUC的評估準則中表現的比傳統MF模型以及一些目前發展最好的雙向推薦系統還要好。換句話說，我們的模型可以藉由推薦讓使用者更容易找到適合的配對。

Recommendation systems usually recommend items based on user’s preference, but reciprocal recommendation systems recommend people to people and try to make them match. Applications like online-job-hunting or online-dating. They match their customers but can’t profit from traditional recommendation system, because we don’t know whether the recommendation terms like us or not. To solve this problem, we provide a Matrix Factorization (MF) based model and experiment with two real-world dataset, one is from online-dating website (libimseti), one is from an online employment system (104 human resource bank). The core concept of our model is to adjust the learning strategy of MF (SGD function), trying to re-rank the recommendation list of two potential match users. By experiment result, our model gets better AUC (area under curve) than traditional MF model and some state-of-the-art reciprocal recommenders. In other words, we can make higher match ratio from recommendations.