A Tensor Based Approach for Click Fraud Detection on Online Advertising Using BiLSTM and Attention based CNN

Online advertising is an effective way for businesses to reach a larger audience, generate leads and increase brand awareness, as the majority of people spend a significant amount of time online. The use of online advertising platforms employing pay-per-click (PPC) model allows effective dissemination of the products to the public. However, it is imperative for businesses to ensure the timely detection of fraudulent clicks, as such clicks can cause significant financial losses to the company. Therefore, early identification of fraudulent clicks are of paramount importance. One approach for identifying fraudulent clicks is through the use of machine learning algorithms that analyze data dynamically and adaptively. However, such algorithms typically treat multi dimensional data as a singular feature matrix or vector, which can make it difficult to uncover intrinsic relationships among a sequence of data. Additionally, the large number of daily fraudulent clicks across different types of devices and platforms can make it challenging for models to focus on specific patterns, leading to relatively low accuracy in predicting fraud. The LSH based tensor recovery mechanism is a solution for this problem, but providing this tensor as input to a simple neural network will not result in high accuracy. To solve this issue, this work introduces a tensor based approach for click fraud detection using BiLSTM and attention based CNN. A publicly available dataset named Talking Data is utilized to conduct comprehensive experiments and to validate the effectiveness of the proposed model. The new model outperforms state-of-the-art machine learning algorithms with regards to accuracy, F1-score, precision and recall rate, as demonstrated through rigorous evaluation and comparison.