ML-driven large-scale heterogeneous music-traffic data analysis

Abstract

A personalized music recommendation system can construct a user’s music preference model based on the user’s historical music behaviors and recommend music that the user is interested in. Typically, users’ music preferences are divided into long and short-term preferences. The ongoing development of new computing technologies, combined with machine learning, makes it possible to better mine user preferences and boost the effectiveness of recommendation systems. Music-traffic data is large-scale heterogeneous data, and information overload is becoming increasingly severe. Although recommendation algorithms can alleviate problems such as data sparsity, effectively integrating large-scale heterogeneous music-traffic data to maximize the data value is the biggest challenge facing current recommendation systems. As the scale of music-traffic data gradually increases, the energy consumption problem becomes increasingly prominent. How to achieve low-power processing of these algorithms has become an urgent problem to solve in realizing energy-efficient data analysis. To this end, first, a music recommendation model based on long short-term preference modeling is proposed. Then, considering the imbalance and sparsity problems in large-scale heterogeneous music-traffic data analysis, a two-stage decomposition method is proposed in the non-negative matrix decomposition. Finally, an energy-efficient scheduling strategy for heterogeneous music-traffic data is designed with the energy consumption problem faced by the increase of data scale. The ablation trial further validates the model’s effectiveness, and the experimental results demonstrate that the proposed music recommendation model performs better than benchmarks. Simultaneously, the experiment verifies the effectiveness of the proposed algorithm on energy-efficient.