Time-Constrained Playlist Generation Using Text Embedding of Content Synopsis

This study proposes a playlist generation scheme that meets the requirement of time constraint. Playlist generation involves creating a personalized collection of audio/video contents and it plays a significant role in music and video streaming services. A personalized collection of contents automatically generated sequentially can be useful in on-demand streaming services during spare time; however, users may not be able to consume their desired suitable contents under limited time circumstances. To address this issue, we introduce a combinatorial optimization problem to meet a time constraint, namely the requirement that the total duration of the playlist is within the specified time length. The time-constraint optimization is formulated as a knapsack problem, which inevitably leads to an excessive computational burden for large-scale optimization. To reduce the high computational costs, candidate contents for the knapsack problem are narrowed down through a k-nearest neighbor search. An additional tuning parameter is incorporated to adjust the aggressiveness of selection depending on video duration, that is, prioritizing a specific range of video duration according to use cases. Experiments demonstrate that the proposed scheme realizes the playlist generation to meet the time constraint of total duration and policies for video duration priority.