利用隱含式回饋基於情境資訊之餐廳推薦

Abstract

對大部份的人而言，找尋一間符合當下情境的餐廳是一個困難的問題。除了考慮餐廳本身的資訊(例如：食物類型、價位、氣氛)外，一個好的餐廳推薦系統也必須考慮使用者當下的情境資訊，例如：聚餐的目的、與會人員之間的關係。雖然目前已經有許多考慮情境資訊的餐廳推薦研究，但大部份都只包含使用者當下的位置資訊而已。本研究致力於結合更多有用的情境資訊，以產生更好的推薦結果。

不同於大部份利用明確式回饋(explicit feedback)的餐廳推薦系統，本研究利用使用者的餐廳訂位紀錄來預測其偏好，視為隱含式回饋(implicit feedback)。每筆訂位紀錄包含該次用餐的情境資訊，例如：用餐日期、聚餐人數。在本研究中，我們提出 PITF-BPR 演算法以建立考慮使用者偏好、及當下情境資訊的模型(model)。我們利用台灣最大的線上餐廳訂位系統 EZTable 三年多的訂位紀錄來操作實驗，並呈現使用 event contexts 後的模型可使推薦結果進步。

Selecting a good restaurant for an event is a difficult problem for most people. In addition to intrinsic factors of restaurants (e.g. food style, price, and taste), a good recommendation system should also consider users' contextual information such as the purpose of gathering and the type of companions. Although there are many context-aware restaurant recommenders, most of them include location information only. This research aims to incorporate more useful contexts into the recommendation process. Unlike extensive work on explicit user ratings of restaurants, this study utilizes users' restaurant booking history to recommend restaurants. Each booking record contains the dining contexts, e.g. dining date and number of companions, a user exits the system when he reserves the table of a restaurant. In this study, we proposed to use the PITF-BPR algorithm to model the context contained in restaurant booking history. Experiments were conducted using the three-year booking history from EZTable, the largest online restaurant booking service in Taiwan, to see the improvements of recommender after we modeled the event contexts.