Dynamic Incremental Network for Estimating Chinese Musical Vogue Patterns of a Given Artist Over Time

Abstract

Estimating musical future trends is critical for the Chinese music industry, as it assists artists, producers, record labels, and investors in discerning public musical preferences. This knowledge facilitates market strategy optimization, recommendation system improvement, inventory management, and provides informed support for advertising and sponsorships. The benefits extend beyond enhancing user experience and contributing to cultural studies, as they are also invaluable for the discovery of emerging artists and the sustainable growth of the entire Chinese music ecosystem. Nonetheless, existing models for prolonged musical future trend estimation have been challenged by the severe decay of past information. Inspired by the concept of incremental learning, we propose the Dynamic Incremental Network (DINet). DINet capitalizes on high-dimensional sequential information embedded in the memory states of trained sequence models. It dynamically integrates information from both previous and current time steps when constructing the model to estimate the next time step's input sequence, enabling the incremental transmission of past information and addressing the decay problem in prolonged musical future trend estimation. Leveraging an open-source dataset from a musical pattern estimation contest, ablation experiments and comparative analyses have shown that DINet not only surpasses traditional and baseline models but also outperforms existing musical future trend estimation models, demonstrating its superior performance. In the future, we intend to employ the proposed Dynamic Incremental Method (DIM)an Occam's razor (simple yet effective) mechanism-to further enhance the performance of state-of-the-art musical and other sequential estimation tasks.