v-TCM: Vertical-aware Transformer Click Model for Web Search

Understanding and predicting user click behavior on a web search engine results page is critical for online advertising and recommendation engines. The click prediction results can be further used to estimate the relevance of the search engine results. Sequence-to-sequence models are effective for predicting user click behavior since they consider contextual information in the input sequence, and transformers are the recent state of the art in sequence learning. A novel transformer model is proposed in this paper for click prediction and relevance estimation that learns additionally from the vertical information, apart from the query and search engine results that are the inputs for the traditional click models. Vertical information is the representation style of a document in the search engine results page. The proposed vertical-aware transformer click model (v-TCM) takes into account both the position bias and the vertical bias of the search engine results. From our experiments, it is proved that v-TCM outperforms the existing click models for the click prediction and relevance estimation tasks.