個人化的社群短文轉發推薦 - 使用隱語義模型以及反應時間的特徵

Abstract

隨著互聯網（Internet）的快速發展下，使用者可以使用任何形式的管道發布個人資訊以及各種言論。不論是哪一種的平台都面臨大量資訊流（Large information overload）氾濫情況，因此我們需要的是資訊流 Filter 的動作，篩選出自己真正感興趣的內容，這就是本篇研究最主要解決的目標：推薦系統（Recommendation Systems）。而往往在做推薦系統中常常會碰到資料收集不完善的問題，我們只能片面的得知服務整體的評斷分數（Overall rating），例如按讚數量、分享數量、轉發次數、留言等等資訊，最缺乏的就是擁有使用者（User）對於物品（Item）個人化評分的指標，因此我們在本研究就由此著墨，希望找出能夠代表使用者喜好程度（User Preference score）的方法。 本研究使用的是新浪微博（Sina Weibo） 的資料集，我們提出名為「LFMRT 2 」的模型，使用 Matrix Factorization (MF) Model 以及 Linear Model 的混合（Hybrid）模型，希望透過 Latent Factor Model 將我們定義的 User Preference 反應時間（Response time） 以 Preference score matrix 作為學習的目標，另外透過 Explicit Features 來做為輔助 MF Model 中 Bias 的重要影響因素。最後透過 BPR（Bayesian Personalized Ranking，基於貝氏後驗優化的個人化排序演算法） Framework 來做為求解整個模型的流程。

Under the development of Internet, users can share their information and comments by ways. No matter what platform users use, they are facing large information overload, most of which they do not have interested in. We need a way to filter large information streams, which they do have interested in. So, the target problem we aim to solve is: Recommendation Systems. However, we can collect only the overall rating or overall information about items by the platform. For example, number of likes, number of shares, number of retweets, number of comments ,etc. It is hard to gather some information about user preference, which is used to evaluate recommended results. Therefore, we apply some method to represent user preference in this research. Our experiment uses the Sina Weibo online tweets as data sets. To learn the recommendation systems, we present a model called the LFMRT 2 model in this research. We use Matrix Factorization (MF) Model and Linear Model as our hybrid method. Using our self-defined user preference to learn latent factor model, we consider not only response time as our preference score matrix but also explicit features. Moreover, we adopt BPR (Bayesian Personalized Ranking) framework as our system flow.