Predictive Video Analysis and Viewer Engagement
The Influencer Marketing (or Creator Marketing) Industry is forecast to increase from approximately £6bn in 2019 to almost £12bn by 2022, and is set to be the fastest growing online-acquisition channel.
On top of that, video marketing is in growing demand; users want to see more video and brands are seeing an increase in leads.
Recognising the need to capitalise on both themes, a digital advertising agency working with industry leading content creators and brand managers approached T-DAB to predict the level of engagement achieved by online influencer videos on social media platforms.
Guided by their industry expertise, we proposed to develop a machine learning engine and supporting architecture that YouTube content creators and brand content managers could use to gain predictive insight into the performance of their content on YouTube and provide guidance for improving the quality of the videos.
To predict viewer engagement with a video, a multi-step machine learning architecture was devised. The first step was feature extraction using out-of-the-box deep learning tools from Microsoft Cognitive Services, videos were analyzed, and extracted information was stored.
In the second step, bespoke unsupervised machine learning models performed clustering of videos on a range of attributes in order to quantify attributes of videos (such as edginess, uniqueness, etc.) in relation to others, as well as perform trend analysis of topics extracted to provide information on relevancy, trend, and remaining growth.
In a third step, extracted features from the first two steps together with video meta-data were used to train a supervised ensemble model to predict audience engagement. Said ensemble model was a heterogeneous mix of weak learners, each trained on a unique part of the overall input data. Model type for each weak learner was chosen automatically from several options, depending on the test accuracy.
Additional to the prediction of engagement, an interactive dashboard serving video-specific insights was created and integrated to the database. The dashboard consisted of elements relating to video’s performance on YouTube, performance relative to other creators, forecasted trendiness of topics discussed in the video and some others.
Video meta data and extracted features were stored on the SQL Server in Microsoft Azure. The overall data architecture allowed the collection and loading processes to be automated, CI/CD pipeline allowed for the database to be quickly changed and updated.
The benefits of this system accrued for two end users; influencers and agencies.
For influencers, the tool allows them to upload and test videos, predict their engagement, and provide insights and benchmarks to help them identify areas for improvement.
For agencies, it provides a tool for testing videos across a whole campaign, identify which influencers are generating impactful content, and identify micro-influencers who are predicted to have disproportionate engagement.
The system can predict with up to 75% accuracy (more than three times as good as random) the level of engagement achieved by a video.