現今網際網路快速的發展，大量的數位音樂，已經廣泛在網路上傳播，造成使用者無法隨心所欲的找到想要的音樂歌曲。許多電子商務更是發展音樂推薦系統來提高顧客需求慾望。而一般線上音樂推薦系統，記錄了使用者所有歷史交易資料，並全部進行分析。因此，便增加執行時所耗費的成本、時間及是否符合使用者目前真正喜好的項目。本論文利用RFM模組來分析顧客價值，並且將相同顧客價值歸為同一群組，進而達到分群的動作。結合使用者最近習慣，提出以權重漸進探勘（Incremental Mining based on Weight）的構想，以漸進增加交易資料量的方式來探勘最近規則，而不需將全部交易資料都做分析，藉以節省計算成本、時間，並以Apriori演算法來探勘關聯式規則。而用相似向量矩陣計算使用者們之間的相似度關係，便利相似聚集。最後利用協力式推薦（Collaborative Filtering）的概念，由推薦模組將音樂推薦給使用者，做為個人化推薦方式。 實驗結果顯示，結合RFM模組及相似聚集推薦較單純只使用RFM分群方式為佳。此外，利用權重漸進與分群方式，更能夠推薦使用者喜好的音樂。而整體上，本論文的推薦準確率高達0.78，比其他推薦方法高出15％～32％，有效的達到個人化推薦的效果。

Because of the rapid development of internet network, the large amount of digital music has spread extensively on the Internet. That causes users cannot follow one’s bent to find out the music or songs they want. Many e-commerce make further efforts to develop Music Recommendation System to improve customers’ demands and desires. The general on-line Music Recommendation System records all user’s former transaction and analysis them completely. So, it increases the cost, time, and items which adapt to what users like now or not. This paper combines RFM model to analysis customers’ value and classify the same one as the same group. We combine users’ Recent Behavior to Incremental Mining based on Weight, which can mine for relations by it, not analyzing all data, to decrease to calculate cost and time. It also prospect for Association Rule Mining by the Apriori algorithms. And then, similar vector matrix is used to calculate the degree of similarity relation between users’ to assemble them conveniently. Finally, through the concept of Collaborative Filtering, we take advantage of recommendation model to be the method of individual recommendation that put music up to users. According to the experiment results, it is better to use the combination of RFM model and similar assembling than RFM classification only. What is more, it also shows that using IMW and classification model can recommend fitfully music what users like. In the whole, the accuracy of this research is up to 0.78, which is higher 15% to 32% than the others, to make up the effect of individual recommendation.