Ad Campaigns Generation Using Deep Learning Approach

Abstract

An advertising campaign is a set of advertising messages that related to each others. It identifies the message of promotion of a brand, a product, or a service. Campaigns usually use multiple media to promote the product and services such as text and Image/Video. Text campaigns are used in search engines such as Google and Bing, whereas the Image/Video campaigns are used in social media platforms and YouTube. Generating a text ad manually is a big challenge. It needs a human expert to create a text campaign to target a specific customers. This requires a lot of time, money, and efforts because they have to create and test new text campaigns that match customer interests and needs. In addition, some companies have a limited understanding of their customers because their customer data are scattered in multiple places such as emails and purchase orders. Thus, companies need a comprehensive understanding to take the actions and decisions in time. In this research, we develop a technique to generate text ads campaigns for search engines such as Google and Bing based on customer’s hot search keywords. Our technique consists of four main phases. The 1st phase is to collect good quality data. The 2nd phase is to preprocess collected data. The 3rd phase is to develop ads campaigns generation model using the recurrent neural networks (RNN) and deep learning approach. In this work, we investigate different NN architectures such as LSTM and GRU. The last phase in this work is the evaluation phase. We conduct several experiments to determine the best model by check the least training loss error value and the perplexity (PPL) of the generated text. In addition, results are evaluated using two metrics; include readability and relevancy of the generated text. The readability metric is evaluated by human annotators and by a Textstat tool, whereas the relevancy is only evaluated by human annotators. The evaluation shows that the generated texts by RNN are mostly easy to read and relevant to the provided keywords.