Dataset for the Article "A Predictive Method to Improve the Effectiveness of Twitter Communication in a Cultural Heritage Scenario"

This is the <strong>dataset</strong> for the article "<strong>A Predictive Method to Improve the Effectiveness of Twitter Communication in a Cultural Heritage Scenario</strong>". <strong>Abstract: </strong> Museums are embracing social technologies in the attempt to broaden their audience and to engage people. Although social communication seems an easy task, media managers know how hard it is to reach millions of people with a simple message. Indeed, millions of posts are competing every day to get visibility in terms of likes and shares and very little research focused on museums communication to identify best practices. In this paper, we focus on Twitter and we propose a novel method that exploits interpretable machine learning techniques to: (a) predict whether a tweet will likely be appreciated by Twitter users or not; (b) present simple suggestions that will help enhancing the message and increasing the probability of its success. Using a real-world dataset of around 40,000 tweets written by 23 world famous museums, we show that our proposed method allows identifying tweet features that are more likely to influence the tweet success. <strong>Code to run a selection of experiments</strong> is available at https://github.com/rmartoglia/predict-twitter-ch <strong>Dataset structure</strong> The dataset contains the dataset used in the experiments of the above research paper. Only the extracted features for the museum tweet threads (and not the message full text) are provided and needed for the analyses. We selected 23 well known world spread art museums and grouped them into five groups: G1 (museums with at least three million of followers); G2 (museums with more than one million of followers); G3 (museums with more than 400,000 followers); G4 (museums with more that 200,000 followers); G5 (Italian museums). From these museums, we analyzed ca. 40,000 tweets, with a number varying from 5k ca. to 11k ca. for each museum group, depending on the number of museums in each group. Content features: these are the features that can be drawn form the content of the tweet itself. We further divide such features in the following two categories: – Countable: these features have a value ranging into different intervals. We take into consideration: the number of hashtags (i.e., words preceded by #) in the tweet, the number of URLs (i.e., links to external resources), the number of images (e.g., photos and graphical emoticons), the number of mentions (i.e., twitter accounts preceded by @), the length of the tweet; – On-Off : these features have binary values in {0, 1}. We observe whether the tweet has exclamation marks, question marks, person names, place names, organization names, other names. Moreover, we also take into consideration the tweet topic density: assuming that the involved topics correspond to the hashtags mentioned in the text, we define a tweet as dense of topics if the number of hashtags it contains is greater than a given threshold, set to 5. Finally, we observe the tweet sentiment that might be present (positive or negative) or not (neutral). Context features: these features are not drawn form the content of the tweet itself and might give a larger picture of the context in which the tweet was sent. Namely, we take into consideration the part of the day in which the tweet was sent (morning, afternoon, evening and night respectively from 5:00am to 11:59am, from 12:00pm to 5:59pm, from 6:00pm to 10:59pm and from 11pm to 4:59am), and a boolean feature indicating whether the tweet is a retweet or not. User features: these features are proper of the user that sent the tweet, and are the same for all the tweets of this user. Namely we consider the name of the museum and the number of followers of the user.本数据集（dataset）对应论文**《提升文化遗产场景下Twitter传播效果的预测方法》**（原标题：A Predictive Method to Improve the Effectiveness of Twitter Communication in a Cultural Heritage Scenario）。**摘要**：博物馆正积极引入社交技术，以期扩大受众群体并增强公众参与度。尽管社交传播看似易事，但社交媒体运营者深知，仅凭一则简单讯息触达百万受众绝非易事。如今每日有数以百万计的帖文竞相争夺点赞与转发的曝光量，而针对博物馆传播场景以识别最优实践的研究却寥寥无几。本文聚焦Twitter平台，提出一种可解释机器学习（interpretable machine learning）技术驱动的新型方法，可实现两大目标：(a) 预测某条推文（tweet）大概率会获得Twitter用户的认可与否；(b) 提供简明建议，助力优化讯息内容并提升传播成功概率。本文使用由23家全球知名博物馆发布的约4万条推文构成的真实数据集，验证了所提方法可有效识别对推文传播效果影响更为显著的特征。 **可运行部分实验的代码**公开地址为：https://github.com/rmartoglia/predict-twitter-ch ## **数据集结构** 本数据集仅包含博物馆推文线程（tweet threads）的提取特征，而非完整帖文文本，可直接用于后续分析。 我们选取了23家全球知名的艺术博物馆，并将其划分为5个组别： - G1：粉丝量不少于300万的博物馆； - G2：粉丝量超过100万的博物馆； - G3：粉丝量超过40万的博物馆； - G4：粉丝量超过20万的博物馆； - G5：意大利籍博物馆。 从上述博物馆中，我们共分析了约4万条推文，每个博物馆组别的推文数量约为5000条至11000条不等，具体数量取决于该组别内的博物馆数量。 ### **内容特征（Content features）** 此类特征可从推文自身内容中提取，我们进一步将其划分为两类： 1. 计数型特征（Countable features）：此类特征的取值分布于不同区间，包含：推文中的话题标签（hashtags，即前缀为#的词汇）数量、统一资源定位符（URL，即外部资源链接）数量、图片（含照片与图形表情符号）数量、提及对象（mentions，即前缀为@的Twitter账号）数量，以及推文的字符长度。 2. 二值型特征（On-Off features）：此类特征的取值为{0, 1}的二值变量。我们考察的内容包括：推文中是否包含感叹号、问号、人名、地名、机构名及其他专有名词；此外还考量了推文的话题密度：假设涉及的话题对应文本中提及的话题标签，我们将话题标签数量超过设定阈值（阈值设为5）的推文定义为高话题密度推文；最后，我们还关注推文中可能存在的情感倾向（正面、负面或中性）。 ### **上下文特征（Context features）** 此类特征并非从推文自身内容中提取，而是用于刻画推文发布时的整体语境。具体而言，我们考量了推文的发布时段（划分为四个时段：凌晨5:00至上午11:59为晨间，中午12:00至下午5:59为午后，下午6:00至晚上10:59为晚间，晚上11:00至次日凌晨4:59为深夜），以及一个布尔型特征，用于标识该推文是否为转发帖（retweet）。 ### **用户特征（User features）** 此类特征属于发布推文的用户（即博物馆账号），且该用户发布的所有推文均共享相同的用户特征。具体包括：博物馆名称，以及该账号的粉丝数量。