Predicting Temporal Patterns in Keyword Searches with Recurrent Neural Networks — Phenotyping Human Behaviour from Search Engine Usage

Web users worldwide rely on search engines daily, querying diverse terms to locate pertinent information. Due to the omnipresence of search engine in contemporary lives, we hypothesise that finely grained analyses of these search terms volume can offer valuable insights into societal trends, potentially reflecting economic conditions and overall quality of life. We examined Google Trends data, sampled hourly, for 61 specific search terms, revealing three primary patterns in how these keywords are used across daily search activities. We employ Dynamic Time Warping to compare the search volumes of these keywords, then apply hierarchical clustering for categorisation. Additionally, a Recurrent Neural Network (RNN) is used to learn the 24-hour time series patterns of these searches. We evaluate the RNN's effectiveness through two experiments, assessing its capacity to generalise across diverse keyword types and various dates. Incorporated into a broader framework, this RNN could potentially help monitor social welfare in near real time and guide policymaking addressing fundamental societal challenges.