Towards Autonomous Video Delivery through Artificial Intelligence
Video service providers are dealing with hyper-competition in the media industry. To remain competitive, they must deliver not only excellent service quality, but also be agile and fast in terms of service development. Service failures need to be eliminated quickly and prevented before they happen. There are two challenges here, the first one being that raw data is siloed and needs to be transformed into insights readily available for different stakeholders. The second problem is that video ecosystems are becoming increasingly complex; how to manage that complexity with limited resources? The answers lie in a shift in focus to a holistic, silo meshing, data strategy and the power of automation, machine learning and artificial intelligence. In this paper, we share the steps that Divitel is taking towards the goal of building fully automated video services and explain how we plan to use Artificial Intelligence in the execution.
Video Service Proivider haben es mit einem Hyper-Wettbewerb in der Medienindustrie zu tun. Um wettbewerbsfähig zu bleiben, müssen sie nicht nur hervorragende Servicequalität liefern, sondern auch agil und schnell in der Entwicklung von Dienstleistungen sein. Ausfälle von Diensten müssen schnell beseitigt und verhindert werden, bevor sie auftreten. Hier gibt es zwei Herausforderungen. Die erste besteht darin, dass die Raw-Daten in leicht zugängliche Informationen umgewandelt werden müssen. Das zweite Problem besteht darin, dass Video-Ökosysteme immer komplexer werden. Wie ist diese Komplexität mit begrenzten Ressourcen zu bewältigen? Die Antworten bieten eine Verlagerung des Schwerpunkts auf eine ganzheitliche, vernetzte Datenstrategie, eine leistungsfähige Automatisierung, maschinelles Lernen und künstliche Intelligenz. In diesem Beitrag werden die Schritte vorgestellt, die Divitel auf dem Weg zum Ziel der Entwicklung vollautomatischer Videodienste unternimmt, und wie dabei künstliche Intelligenz eingesetzt werden soll.
A TV Platform is a complex ecosystem built out of many building blocks that are highly dependent on each other. For example, to play a recorded program, several platform components need to work in harmony to serve the customer the requested content and deliver an outstanding viewing quality. These components on their own are responsible for certain types and levels of quality. To offer an excellent holistic video delivery quality, these levels of quality cannot be managed in a siloed state.
Divitel has developed a holistic, data driven approach to model the interdependencies of these different components and their complex relationships along with a systematic framework to identify and solve confounding failures. The resulting delivery infrastructure achieves significant improvements in holistic viewer quality, measured in reduced churn rate, increased customer satisfaction and increased Net Promoter Score (NPS).
More importantly, it better enables to employ machine learning and artificial intelligence and paves the way for autonomous operations of video delivery ecosystems.
Three types of quality are not enough to satisfy all stakeholders
The quality of a video delivery service is most often evaluated by three dimensions. The first one being the Quality of Customer Care. This is where a lot of perceptual value can be gained or lost. Information is generated from the CRM, OSS and BSS, online payments and ticketing components, that in theory, should enable the 1st line and 2nd line support workers to do their job. The problem here is that in most cases, the insights needed to solve customer complaints are technically not only generated from data monitored in this specific customer care silo. The data could be generated elsewhere in the system and needs to be transformed into insights that are suitable for 1st line customer care support workers.
The second quality dimension concerns the controlplane, the engine of the TV ecosystem. Here the backend and the frontend work together (through many different components like CMS, DRM, CDN, recommendation, registration engines, etc.) This collaboration is central to what we in our industry call the Quality of Experience.
Most common metrics consist of different parameters reflecting viewer engagement like join time, buffering ratio, rate of buffering, viewing time and number of visits. Just as an example, research suggests that viewers begin to aband on a video after a 2-second delay in start time, and the abandonment rate increases by 6 % per second after that.1) The difficulty of getting this Quality of Experience right is that here too, problems with a negative impact on quality can arise anywhere in the video delivery chain, within or even outside of the control plane. As a result, video service providers must perform an impossible balancing act to deliver a service at optimal quality.
The third video quality dimension has to do with video transmission quality and service availability. Videocontent processing and delivery through connected components to generate what in our industry is known as the Quality of Service with low level, traffic-related metrics like packet loss, delay and jitter. Here the video content is processed and delivered, consisting of processes like transcoding, ad insertion, on the fly packaging, etc.).
Again, here too, it is not enough to only monitor the mentioned metrics to guarantee excellent overall video delivery quality.
The data from these different planes are interdependent and need to be transformed into the insights needed to meet the needs of different audiences and stakeholders. First line customer care support needs insights. Secondline support also needs it to be able to find and fix problems. Lastly, the software vendors within the planes need it in order to pinpoint the specific component causing issues so the right one can fix it.
Holistic, data driven approach
The first step towards the creation of a holistic view of quality and the overall ecosystem performance is to unify all data from previously mentioned planes and silos into one – the Data & Intelligence plane (figure 1).
From here, it becomes possible to explore and visualize the performance of processes throughout the entire ecosystem across silos, planes and components. Secondly, it becomes possible to follow trends of key metrics and correlate events. In this way, when problems occur, it will become easy to access log files and perform queries to pinpoint failure root causes from one point. Another significant benefit is that the data becomes the proof to identify which specific technology component is causing the problem, making orchestration of vendor fixes much more efficient.
Divitel is not just selling the results of the data analysis, but converts the results into insights and actions, and puts them into perspective. These insights inform our customers what was happening, what caused it and how to fix it. Was it a capacity issue, or did a server breakdown, or had the clients simply no route to the service?
Figure 2 shows an empirical case from a Divitel field deployment where we have applied mentioned holistic data insights approach. As the graph demonstrates, we have realized a clear decline in incoming customer reported incidents and an increase in machine-generated reported incidents. Other empirical results show that MTTR went from months to days even though the customer base of the video service provider has been steadily growing.
This holistic approach will enable us to deliver first line support workers the real-time insights they need to handle customer complaints in an extremely efficient way so they can truly help viewers when they call.
Artificial Intelligence and Automation
Now that we work with holistic data and insights, Divitel can develop algorithms that improve over time (the more data, the better the predictive abilities of our algorithms). By categorizing data, we will be able to train the model by analysing ‘good quality patterns,’ identified and fed to by human workers with video delivery domain knowledge. In this way, we will program the system to identify ‘the normal’ patterns and therefore detect anomalies.
Machine learning will allow further acceleration of root cause analysis beyond what we were already able to realize today. Here the possibilities are endless. One example of machine learning is in the field of capacity planning, where we could predict the required capacity and plan accordingly. Another example is that it becomes possible to predict connectivity issues a head of time and automatically adjust the streaming resolution to create enough playback buffer for a smooth userexperience.
Within our holistic approach to video delivery quality, many benefits can be gained by creating a digital brain that can make decisions and act autonomously. Artificial Intelligence can be applied to learn from previous outages and look for patterns in the current, data which might suggest a beginning of an incidentor correlate it through different sets of the ecosystem which we humans did not notice before.
First results from the field
The first step in the journey towards autonomous video delivery lies in mentioned holistic approach to managing data and creating stakeholder insights. We demonstrate that significant improvements can be achieved:
- 80 % faster root cause analysis,
- 65 % fewer tickets,
- 50 % faster Mean Time to Recover,
- 40 % increase in first time right,
This way of working is the key to managing video delivery ecosystems in a more efficient manner, not only reducing the number of incidents, but also by being able to solve the issues that occur, not by adding more resources, but by working with insights gained from all interdependent components, which in turn has significant benefits, not only to system performance, but also to viewer satisfaction.
Further warping of these results can be achieved by applying automation and machine learning. The final destination of this journey is the application of artificial intelligence and with it the creation of autonomous, excellent video delivery.