基於潛在協同關係之快速推薦系統

Abstract

在近年來推薦系統技術已經被廣泛應用在各類型的電子商務網站，其中協同過濾是最常見的解決方案。矩陣分解是協同過濾研究領域中最受矚目的方法，它最主要的概念是不同的人面對不同的人事物，會有不一樣的行為模式，而根據這個行為模式去學習背後的潛在因素，預測時可根據學習得來的潛在因素去判斷未來的行為模式。本研究將此概念延伸到網路分析中，探索物件與物件之間的關係，提出了 latent collaborative relations 模型。此外，每個物件一些事先俱備的背景知識應該要加入模型中，既能提升模型準確度，又可有效解決遇到新物件的推薦問題。除了準確度之外，本研究另外使用一個衡量指標:覆蓋率。覆蓋率愈高，代表推薦的種類較多樣性，促進長尾式銷售的機會也愈高。在實務上目前的潛在因素模型面臨了即時推薦的問題，本研究模型的推薦過程可轉化成最近 k 個鄰居搜尋問題，故能搭配 multi-probe locality sensitive hashing 索引，可大幅度提升預測速度，讓推薦在有效時間內計算完成。

Matrix factorization (MF) methods provide one of the most effective approaches to collaborative filtering. Its main idea is that according to the behavior patterns of people, we can learn the latent factors behind these patterns. Those factors can help up to predict behavior of people in the future. The study also extends this concept to a network analysis to explore the relationship between objects and objects, and we call this task latent collaborative relations. In addition, we can also take account of the object’s profile. Beyond accuracy, we also discuss another crucial metric: coverage and ponder the long tail phenomenon. Finally, we reduce the retrieval of recommendation in this model to a simple task of k-nearest-neighbor search via multi-probe locality sensitive hashing. We evaluate our algorithms on real-world datasets, demonstrating 5-300x speedup with respect to the naive linear search.