Multi-interest semantic changes over time in short-text microblogs

Abstract

Consumption of content in short-text microblogs is necessitated to a large extent by individual users and their friendship network interests. Based on the dynamism in the data throughput on such platforms, e.g., Twitter, prevailing conditions are bound to determine the nature of consumed or disseminated content. Therefore, semantic interests differ over time even for individual users. Detecting this semantic change over time is integral in mapping user profiles over a time period, especially in microblogs where only the extrinsic user profile identifiers provide metadata that seldom evolve. This is vital in serving relevant third-party content as well as in the computation of topical interest variations over time. In essence, current, and relevant topics of interest to a user on such a platform may not be representative of the same users’ interests a few months later. In our quest to identify the most user-representative interests at any given time, each topical term was modelled as the inner product between word embeddings and a time-based embedding representation of assigned topics at varied time periods. The model was fitted onto tweets as time-series documents. To validate the model, changes in the extracted user-representative interests over time were semantically weighed against a mirrored, time-variant dataset. Interest weights across the time-variant datasets were computed and validated in five sub-topics for a period spanning two and a half years. Linearity in the relationships between the test and validation sets could be identified, more so in emerging topics. A Pearson correlation coefficient as high as 0.871 was achieved in interest change verification over the tested period.