Generating ad creatives using deep learning for search advertising

Abstract

We propose an automatic ad creative generation using a recurrent neural net- work. A landing page being advertised includes relevant text data, which can be used for generating ads. We formulated the ad creative generation task as a text summariza- tion problem and built a sequence to sequence model with a Long Short Term Memory network. In this network, a source sequence is encoded and represented as a fixed-size vector, which represents the context of the sequence. The decoder uses the source encoding to predict the target sequence. In order to assess the validity of the proposed approach, we experimented on four different datasets. The empirical evaluations have shown that the proposed model can generate relevant ad creatives on a template-based dataset Dtemp with moderate hyper-parameters. Training the model with more content increased the performance of the model due to better hyper-parameter tune-up. Since the rich-in-context dataset Drich includes 20x more unique words than Dtemp , we expect that increasing the size of Drich even further enables the network to generate more relevant ads. We observed that when the source and target shared common sequences during the training, the model produced the best ad creatives. Also, GloVe word embedding for input representation improved the performance of the network.