基於時序與風格的語音節目推薦系統研究

An investigation of spoken program recommendation systems based on time and style

蘇品維(Su Pin-Wei)

指導教授 ： 杜雨儒

國立政治大學/商學院/資訊管理學系/碩士(2022年)

https://doi.org/10.6814/NCCU202201103

摘要

因為網際網路以及串流技術的蓬勃發展，在相關的市場上開始衍伸出許多新媒體服務，例如:影音串流(YouTube) 語音節目(Podcast)。其中又因為語音節目的一些自身的特性，以及在台灣市場當中發生了在短時間內湧入大量新用戶以及新產品的問題，使得在語音節目平台中建置推薦系統顯得相當困難，特別是在推薦新產品或是推薦產品給新用戶的議題上。換句話說，要如何推薦適合的語音節目給新的使用者 或是如何推薦新的語音節目給使用者等議題顯得十分重要且具有挑戰性。這些議題在過去的研究中這些問題也被稱作「冷啟動問題」。
而過去對於語音節目推薦上的相關研究十分稀少。而本研究基於在過去少量的語音節目推薦文獻上所使用之文字描述和類別的特性，結合兩個特有的特性—「時序」與語音節目的「敘事風格」，並搭配混合推薦的方法，去解決在語音節目上所發生的冷啟動問題。本研究也透過網路爬蟲的方式蒐集 APPLE Podcast 的評分資料作為推薦系統的訓練資料。除了線下測試外，也實作出推薦系統介面實際讓使用者進行使用者測試。根據我們的線下測試以及使用者測試之結果可以得知在時序與風格屬性的幫助下，在整體以及新使用者的推薦效果都有顯著的提升。

關鍵字

推薦系統 ； 冷啟動問題 ； Podcast ； 機器學習 ； 時間 ； 敘事風格

並列摘要

With the rapid development of the Internet and streaming technology, many new media services have begun to spread to related markets, such as video streaming (YouTube) and spoken program services (podcasts). Due to the characteristics of spoken programs and the influx of new users and products in the Taiwan market in a short time, it is challenging to build a recommendation system in the spoken program platform, especially on the issues of recommending new items or recommending items to new users. In other words, the challenges of recommending suitable spoken programs to new users or recommending new spoken programs to users remain open research questions. These issues are called “cold-start problems” in past studies. Related research on spoken program recommendations was scarce in the past. The present study, based on the characteristics of text descriptions and categories used in a few recommended literatures of spoken programs in the past, combined two unique characteristics—listening time and speaking styles of spoken programs—in a hybrid recommendation method to solve the cold-start problems in spoken programs. The study also collected the rating data of Apple Podcasts through the web crawler, and podcast ratings were used as training data for the recommendation system. In addition to offline testing, we implemented a recommended system interface for users to conduct user testing. According to the results of our offline and user tests, with the addition of time and speaking styles, the performance of the recommendation in overall and new users was significantly improved.

並列關鍵字

: Recommendation systems ； Cold-start problem ； Podcast ； Machine Learning ； Listening Time ； Speaking Style

參考文獻

1. An, H. S. 2017. "Breaking the Pop Music Market in the Us: Effects of Online PrePurchase Influences and Acculturation of Asian Consumers on Their Responses

to Asian Artists’ Contemporary Pop Music," Proceedings of the Northeast

Business & Economics Association).

2. Arif, I., Aslam, W., and Siddiqui, H. 2020. "Influence of Brand Related UserGenerated Content through Facebook on Consumer Behaviour: A StimulusOrganism-Response Framework," International Journal of Electronic Business

(15:2), pp. 109-132.