Large Scale Open Source Video Recommender Tool Using Metadata Surrogates

Video and multi-media sharing is a significant activity on social media platforms. Learning patterns of activities using raw video data is computationally intensive and impractical, and manual inspection is not scalable and prohibitively expensive. An alternate strategy is to learn information about video content using far less compute intensive metadata surrogates. This paper describes a video recommender tool implemented in GovCloud using a novel approach of using lightweight video metadata to learn and classify video content. In contrast to popular video recommender systems that use consumption models for classification, the new approach used in our tool is based solely on the video metadata along with domain expertise used to truth a relatively small subset of relevant video content. The tool is very user-friendly and captures practical knowledge of the user resulting in good learning model. The architecture and implementation specifics of the tool is outlined in this paper. The classifier performance using metadata from tens of thousands of real postings exceeds 90% for both recall and ROC metrics. This tool has shown promise in providing a console for aggregating social media videos for analysts to train the system consistent with the context and task at hand.