推薦系統之方法組合比較–以歌曲資料為例

現在網路商店的設計越來越精細，時常在網頁的一旁能夠看見商家推薦給消費者的商品，但推薦的商品時常僅是當下較熱門的商品或是使用者曾經點選瀏覽過的商品，無法精準的推薦消費者需要的商品。本研究利用歌曲收藏資料，比較集群分析搭配推薦系統、關聯規則搭配推薦系統，將使用者與歌曲之間的關係利用集群分析與關聯規則將其找出，協助後續的推薦系統能夠提出較準確的推薦依據。用戶收藏資料分析結果顯示，關聯規則搭配推薦系統優於集群分析搭配推薦系統，且在關聯規則搭配推薦系統中，基於用戶的推薦（UBCF）是最適當的推薦方法，且只推薦前五首歌曲時準確度最高。另外關聯規則也可以解決推薦系統中冷啟動(Cold start)、資料稀疏性(Sparsity)的問題，例如:較少用戶收聽的新歌曲，或收聽率高但是無收藏、評分紀錄的歌曲。

關鍵字

推薦系統；關聯規則；集群分析

並列摘要

Nowadays, the design of online stores is becoming more and more sophisticated, and it is often possible to see the products recommended by merchants to consumers on the side of the web page. However, the recommended products are often only the current popular products or the products that users have clicked and browsed, without accurately recommending the products that consumers need. This study uses the song collection data to compare the cluster analysis collocation recommendation system and association rules collocation recommendation system in order to assist the subsequent recommendation system to put forward more accurate recommendation basis. Analysis results of user collection data show that association rule collocation recommendation system is superior to cluster analysis collocation recommendation system, and in association rule collocation recommendation system, user-based recommendation (UBCF) is the most appropriate recommendation method, and the accuracy is the highest when recommending only the first five songs. Association rules can also solve the problems of cold start and sparsity problem in recommendation systems, such as new songs that are listened to by fewer users, or songs with high attendance but no collection of score records.