Scholarlink

<p>Social media platforms like X offer a valuable avenue for analyzing user behavior, interests, and demographic attributes. Accurately predicting a user’s age is particularly beneficial for applications such as targeted advertising, personalized recommendations, and broader user profiling. However, progress in this area is hindered by the scarcity of high-quality publicly available datasets, many of which are biased toward younger users. To overcome this limitation, we developed a large-scale publicly accessible dataset comprising over 7,000 X (formerly Twitter) profiles, covering users aged 13 to 75. Candidate users were initially discovered using tweets that contain explicit birthday and age phrases (e.g., “Happy 21st birthday”) to obtain reliable age labels; to prevent label leakage, all such age reference tweets were then excluded from the training and testing corpora, and models were trained only on each user’s recent non-label tweets and metadata. This dataset includes recent tweets and user metadata, collected through a carefully designed scraping and cleaning pipeline to ensure reliable age annotation. Users were categorized into life stage-based age groups, and we employed deep learning models that leverage both language patterns and metadata from X to predict user age. Our experimental setup utilized a stacking ensemble approach, incorporating base models such as Gradient Boosting, Decision Trees, and Support Vector Machines. Beyond simple accuracy, we evaluated model performance using macro average and weighted F1 scores, reaching up to 0.87 and 0.90, respectively. Additionally, our best performing models achieved precision scores as high as 0.92 and recall values up to 0.93. Despite challenges such as class imbalance and overlapping age categories, our methodology yielded significant improvements in predictive accuracy. This work has practical implications for marketing and content recommendation, where precise age prediction can optimize targeting and reduce operational costs. The publicly released dataset serves as a strong benchmark for future research in demographic prediction on social media, enabling continued progress in AI-driven user modeling.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A large-scale dataset and one-vs-all ensemble with confidence-based decision fusion approach for age prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月28日

查看全文
  • Muhammad Bilal Awais,
  • Maryam Bashir

摘要

Social media platforms like X offer a valuable avenue for analyzing user behavior, interests, and demographic attributes. Accurately predicting a user’s age is particularly beneficial for applications such as targeted advertising, personalized recommendations, and broader user profiling. However, progress in this area is hindered by the scarcity of high-quality publicly available datasets, many of which are biased toward younger users. To overcome this limitation, we developed a large-scale publicly accessible dataset comprising over 7,000 X (formerly Twitter) profiles, covering users aged 13 to 75. Candidate users were initially discovered using tweets that contain explicit birthday and age phrases (e.g., “Happy 21st birthday”) to obtain reliable age labels; to prevent label leakage, all such age reference tweets were then excluded from the training and testing corpora, and models were trained only on each user’s recent non-label tweets and metadata. This dataset includes recent tweets and user metadata, collected through a carefully designed scraping and cleaning pipeline to ensure reliable age annotation. Users were categorized into life stage-based age groups, and we employed deep learning models that leverage both language patterns and metadata from X to predict user age. Our experimental setup utilized a stacking ensemble approach, incorporating base models such as Gradient Boosting, Decision Trees, and Support Vector Machines. Beyond simple accuracy, we evaluated model performance using macro average and weighted F1 scores, reaching up to 0.87 and 0.90, respectively. Additionally, our best performing models achieved precision scores as high as 0.92 and recall values up to 0.93. Despite challenges such as class imbalance and overlapping age categories, our methodology yielded significant improvements in predictive accuracy. This work has practical implications for marketing and content recommendation, where precise age prediction can optimize targeting and reduce operational costs. The publicly released dataset serves as a strong benchmark for future research in demographic prediction on social media, enabling continued progress in AI-driven user modeling.