- 學位論文
建構擴散模型以預測行銷影片在 YouTube 上的人氣
Diffusion Model Learning for Predicting Popularity of Marketing Videos on YouTube
摘要
近年來，隨著社群媒體的發展，人們和品牌的互動逐漸從以實體廣告與行銷活動為主的方式，轉為以線上為主。人們逐漸習慣了透過社群媒體與網際網路瀏覽品牌的最新資訊。因此，對品牌的行銷團隊而言，如何能夠及早的預測一個行銷影片在發佈到社群媒體上之後的影響範圍與觸及人數，將有效協助他們評估一份計畫的價值與效益。換言之，要能有效的在行銷影片的製作階段就評估它上線後的觀看人數變動情況，對於品牌而言是重要且日漸迫切需要的。 我們將透過分析影片與品牌相關資訊（包含：影片標題、影片長度、影片的類別、品牌類別等）並萃取出其中的重要訊息，以預測一段影片在發佈後的擴散情況。我們認為創新擴散模型將有效的描述影片發佈後的觀看人數變動曲線。然而由於擴散曲線的參數並非直接透過搜集觀看人數資料就可取得，我們設計了一個包含兩階段的以深度學習為基礎的擴散模型學習方法。首先是對每日累積觀看人數的資料做曲線擬合，並將擬合結果作為第二階段--擴散模型預測的正確答案。在第二階段的模型中，我們更採用深度學習與注意力機制的技術去萃取出文字中的重要訊息。 除此之外，我們更利用 YouTube 官方所提供的 API 以及 Google Cloud Platform 開發一個可以在每天固定時間自動更新資料集的系統，用以搜集本次研究中需要的品牌與行銷影片的相關資料。
並列摘要
The vigorous development of social media has changed the way people interact with brands. Therefore, it is not hard to imagine that launching videos on YouTube to promote brands and create trending topics worldwide is a common marketing strategy for businesses. Most of existing studies focus on predicting the future popularity of videos by analyzing the patterns of the early observed time-series popularity data with some complementary metadata. However, it is may not be a practical approach in most real-world scenarios since it is usually too late for branding companies to evaluate the performance of a video after all of the planning stages are finished. Therefore, in this research, we propose a diffusion learning model which takes YouTube video metadata as model input to predict the diffusion process of a video before it is released. Given the metadata of a YouTube video, we aim to predict the diffusion curve of the video and then use the predicted diffusion curve to derive daily accumulated view counts. Therefore, in order to estimate the parameters of the innovation diffusion model, we propose a deep-learning-based method with two phases. In Phase 1, we do curve-fitting on the original training instances for estimating the parameters of diffusion curves, and phase 2 is for diffusion model learning. We also develop a data collecting system utilizing YouTube Data API v3 and Google Cloud Platform.