Machine Learning as a service (MLaaS) is a burgeoning field in the digital TV space. Its goal is to create AI/ML based revenue generating products. In this study, a novel use case is presented along with machine learning based enhancements. TV viewers routinely encounter shows that they dislike, but they are unable to avoid seeing them. While the consumer opinions are highly subjective, the end-result is the same: flipping the channel, which leads to advertising revenue loss for the programmer. Although retaining viewership of the channel is highly desired, technical challenges have precluded a satisfactory solution thus far.
The selected use case is of interest because unappealing content and recommendations contrast each other (dissuade vs. persuade). This distinction also manifests in the solution structure. For example, Recommender Systems (RS) are based on user ratings of liked content. In contrast, ‘disliked content’ maybe so averse to a viewer thus it is not even rated. Not having user ratings is a barrier for applying the RS model, which uses similarity measures in the latent space to determine affinity. Hence, in this study a different metric based on implicit data is used for feature vector creation. The goal is to illustrate the challenges and opportunities in developing MLaaS products for carrier-grade video.
Presented is a distributed solution* applicable to vMVPD service. Enhancements to IP content delivery pipeline and Machine learning based automation are key for replacing disliked content. Additional scopes for MLaaS applications are also discussed.
*patent filing (16/167,766)