An Efficient Hybrid Model-Based Classification of Online Personalized Ad and User Intent Detection Using CNN and Deep Q-Networks

Abstract

As a vast market, online advertising has promptly gained aggregate interest in a vast array of platforms, including mobile apps, search engines, third-party websites, and social media. In online marketing, the success of online campaigns is a challenge, which is often assessed by user response using several measures, such as product subscriptions, explicit customer feedback, clicks on ad creatives, or transactions obtained through online surveys. Auditing advertising images, or "creatives," prior to their appearance on publishers' websites is one of the most crucial quality control steps in online advertising. This ensures that advertisements only appear on websites that are appropriate for them. The user experience, the publisher's reputation, and maybe legal ramifications can all be negatively impacted if a sensitive creative is shown on the incorrect website. To detect and classify whether an advertisement has any sensitive content, we use a machine learning algorithm to process the creative image and combine this with the historical distribution of sensitive categories linked to the creative's landing page. To protect against this, the study presents an efficient hybrid model using CNN and Deep Q-networks for the classification of online personalized ads and user intent detection. Initially, the HCOAUICNN-DQN model uses data preprocessing to preprocess the input dataset from the online advertisement website. Next, extract those input features through the pretrained CNN-feature maps. Following, DQN is applied for the user intent detection classification of and online personalized ads. Finally, the hyperparameter tuning using Optuna optimization algorithm is applied to improve the real-time classification performance. A set of experiments was conducted to analyze model efficiency using different datasets of advertising images. The performance of the HCOAUICNN-DQN model will be assessed through accuracy, F1-score, recall, and precision metrics. The HCOAUICNN-DQN method obtains superior outcomes when compared to other existing approaches.