在Instagram上之華人名人發文熱門程度預測

Abstract

近年來，智慧型手機的普及以及通訊技術的進步使得社群媒體蓬勃發展，人們可以藉由社群媒體互相溝通、分享訊息，甚至自我行銷。對於生活在這個世代的我們而言，沒有人能完全不接觸社群媒體。企業、廣告商或是網路紅人能夠藉由社群媒體提升在人群中的知名度，進而獲得收益。因此，對這部分的群體來說，如何維持甚至使名氣更進一步是一個值得深究的議題。 在本篇研究中，我們希望能透過發布貼文的內容，在文章發佈以前預測其在一天之內能夠獲得多少關注。測試的資料來自Instagram，它是一個人們可以透過照片以及些許附註文字來分享跟記錄生活的平台，它強調視覺性內容的分享方式，在青少年中深受喜愛。而目標群體是現實中已經具有一定知名度的人，對他們來說社群媒體上的行銷是一個維持曝光率甚至提升知名度重要的媒介。 我們藉由SnowNLP從貼文附註文字中萃取貼文的情緒，並使用DeepSentiBank作為圖片情緒以及圖片內容分類的工具，再加上使用者本身附帶的資訊，建構出能夠預測貼文熱門程度的預測器。我們的預測結果能在測試資料上獲得90.45%的量級準確率，同時與前人的結果相比，我們預測的結果更接近實際的關注數。此外，我們將熱門程度預測視為一個二元分類問題，即熱門以及不熱門，並在此分類問題上獲得88.71%的準確率。

In recent years, social media become much more popular according to the increasing usage of smartphone and the advances in communication technology. People can communicate, share information with others and even self-promotion via social media. To people who live in the era, no one can totally stay away from social media. Enterprises, advertisers and internet celebrities can gain revenue by increasing their popularity among users of social media. Therefore, how to keep the exposure rate or even increase the popularity on social media has become an issue to them. We aim to predict popularity at 24th hour in advance by post content in this study. Our data is from Instagram which is a platform people can record and share their life by photos. It is a popular social media among teenagers due to its image-emphasizing sharing method. And our target users are people who are well-known in the real world. Promotion in social media can be a vital way to keep or increase their popularity for these people. By using SnowNLP, we can retrieve text sentiment from caption of a post. Then we use DeepSentiBank model as an extractor which can get image emotion and image content category. Combining these features with information of user account, we build models to predict popularity. Predicted popularity has 90.45% log scale accuracy on test set. Comparing our result with previous work, we have closer numbers to the actual popularity. Also, we regard popularity prediction as binary classification task which is popular versus unpopular and we obtain 88.71% accuracy on this classification task.