In this episode, our hosts Simon McGrath and Alex Borland are joined by Veena Chandran, Senior Solutions Architect at Amazon Web Services (AWS), and Chris Birkinshaw, Technology Principal at Merapar, to discuss all things AWS Personalize.
With consumers expecting more from their OTT platforms, this episode explores how personalisation and curated recommendations can be key to delivering a successful customer experience. We take a deep dive into the role that AWS Personalize plays in helping organisations to achieve this.
Inside the episode
Simon McGrath and Alex Borland start from how far recommendations have come since the early video-on-demand days, when the job was to sell the next pay-per-view title. Veena Chandran describes working with nonprofit customers at AWS and on the Personalize deployment with Merapar for PBS. Chris Birkinshaw explains how she brought Personalize into NBCUniversal’s Hayu.
From there the conversation gets practical: what data Personalize needs, how it copes with a viewer it has never seen, why deploying it took more than a typical cloud pipeline, and the development kit Merapar built from that experience. It ends on metadata, which mattered less to accuracy than the team expected, and on how to tell whether recommendations are working.
Highlights
- 07:16 Chris Birkinshaw on Hayu: as she understood it, the first deployment of Personalize in media in Europe.
- 08:11 “The good thing with Personalize is that it’s very easy to do prototyping.” Chris Birkinshaw
- 09:26 “It was something that Amazon.com had been doing for a decade plus.” Veena Chandran on where Personalize came from.
- 17:46 “It’s garbage in, garbage out.” Veena Chandran
- 25:01 With interactions data alone, one client saw a precision at five of 20: show five recommendations and one gets watched.
- 26:04 “I don’t organise my apps into folders on my phone anymore.” Chris Birkinshaw, who relies on suggestions and search instead.
Key questions answered
What is Personalize, and where did it come from?
Veena Chandran explains that Amazon.com had been generating recommendations for more than a decade, and Amazon turned that into a managed service. It was deliberately not built for one domain, but retail and media customers took it up quickly. It now has a video on demand domain with ready-made rails most media services want, such as Trending Now, “because you watched X” and “more like X”. Compared with building your own engine, Veena says, it takes away much of the work of feature and algorithm selection, tuning and deployment.
What data does it need, and how much does metadata matter?
Three data sets: interactions between users and items, user metadata such as demographics, and item metadata such as genre and length. Interactions are the most important signal. A new user has no history, but within a few clicks Personalize adjusts to what they are looking for, without retraining the whole model. Chris Birkinshaw found that adding content metadata made less difference to accuracy than expected; it proved valuable for new items and users, and for filtering, as when one client wanted a 50-50 split between VOD and linear. That contrasts with older media-focused platforms that rely heavily on metadata. Alex Borland expects a packaged product to take as long to integrate, without the same targeting to an operator’s audience and content.
What makes deploying it harder than other managed services?
Chris names two things. First, preparing the data: merging sources, denormalising, and filling gaps such as shows with no genre. Second, importing gigabytes of data and training a model take tens of minutes, not the seconds it takes to deploy a cloud function. So the team used AWS Step Functions for the long-running jobs, and turned what they learned into CDK constructs: the Merapar Development Kit (MDK) for Personalize, which Merapar was preparing to launch at IBC 2023 when the episode was recorded. Veena adds that the Data Insights feature, then a few months old, gives a sanity check on input data.
How do you know the recommendations are working?
In the Personalize console, Chris says, you can track the share of recommendations viewed or interacted with, and how much of the catalogue is being recommended, and tune for more exploration so titles from the deep archive get surfaced. A business will also look at click-through on rails and conversion funnels from its analytics platforms. Veena lists the KPIs customers look for: engagement, acquisition, sometimes revenue (or donations, for a nonprofit) and discoverability of the catalogue.
Speakers
- Veena Chandran, Senior Solutions Architect in the nonprofit sector, AWS
- Chris Birkinshaw, Technology Principal, Merapar
- Simon McGrath, Merapar (host)
- Alex Borland, VP Strategy & Business Development, Merapar (host)
