The Secret Sauce for TV Recommendations (do not share!)
So I have been at The Filter a little over 6 months now and this has given me a great insight into what is really powering the TV personalisation services we offer: More Like This, Recommended For You, Because You Watched and even Search. Prior to The Filter, I was on the other side of the fence working for OTT and digital TV operators. So, it is fascinating now to be inside the data science rather than just seeing the outputs and I just want to bring to life what has surprised me the most – what the secret sauce is...
- It is NOT the recommendation models per se – yes, we have a team of super smart data scientists that talk about the finer mathematical subtleties of recommender systems. Yes, the models are really complex and train themselves every day. However, the models we have built and refined are not at the cutting edge of AI; we are not talking self-driving cars here. They are proven and essential, but they are not the secret sauce.
- The secret sauce is enriching the data – the content data we get from clients is generally aimed at EPGs. It is not designed to power personalisation but to read well. Data science models work best when there is just the right degree of overlap between keywords. Too much overlap or too little and the models just don’t show you what a customer would expect.
In order to get meaningful results that move core KPIs like churn and value perception, you have to enrich content data. This means both significant investment in automated data science models that do this for you and also, for the most popular content, we sometimes have to roll-up our sleeves and manually add tags and franchises. While the data science purists amongst you will shudder at this, we have the results from millions of actual views to show just how important this is. The Filter has been providing TV recommendations for over 10 years now, so we have built up a substantial bank of these proprietary tags.
Above is an example of raw content keyword data from a client and shows the lack of consistency and overlap. This data is not wrong; it is just not helpful for powering recommendations. The Filter enriches these titles to make sure they are grouped appropriately.
- Automated data enriching processes are vital – this is still the bulk of the work we do when we first get a new client’s content data. From the example above, it is relatively easy to develop models that can combine ‘police officer’ with ‘detective‘ with ‘cop’. It is almost like a TV focused thesaurus. We also pull out the more important keyword from, say, an item’s description.
- The Filter Tags are the icing on the cake – this is where we can add our real secret sauce that no other company out there can match. We have tags from ‘car chase’ to ‘war’. Moods are in there like ‘creepy’ and ‘romantic’. We also tag documentaries with words such as ‘sharks’ or ‘oceans’. One interesting thing I’ve learnt is that we only have 700 of these tags. This is due to the need for recommendation’s data to drive the right balance of overlap between titles. Not too much or too little. We often get asked is tagging scalable and it always is. In practice you only manually tag the top 10%–20% of content by popularity and once live, this is only for new–to–platform content. We also have a long history that has given us an established bank of tags.
Example of the before and after for enriching data for The Joker. Links to the Batman franchise are not obvious from the existing metadata but work when our franchise tags are added.
- Tags and franchises need to relate to TV – the output from everything we do here at The Filter is for viewers of TV content. So, it is clearly of paramount importance that when you are enhancing data you do it in a way that makes sense for TV viewers when they are in front of that TV platform with a TV mindset. Franchise is the most obvious manifestation of this. Customers expect More Like This for Fast & Furious 6 to include the other classics of this franchise. Not exclusively as actors, themes, moods, recency all contribute but if the other parts of the franchise are not there, people start to lose trust in the recommendations. Type ‘BOND’ into most TV searches and see what you get.
- Model tuning is an art form – it is not simply a case of refining our models and putting them live. There is a significant pre-live stage required to work out which of the many factors work best and with what weightings: description, time of day, recency, genre, box office, Filter Tags, etc. This will always have the viewer as the primary stakeholder, but can include business and brand objectives from our clients. Perhaps they want to slightly push original content. Once a model is live, then the power of a/b testing is unleashed to fine-tune these weightings, but it is vital to hit the ground running with a model that works well from the start; a/b testing takes time and you can only change so much in any given test. Again, our years of working with many clients has given The Filter a significant head start here.
So I have learnt lots over the last 6 months and suspect that curve will continue. The art of getting really good TV personalisation is as much about enhancing the data as it is in having super smart recommendation models. It is a fascinating subject and one where focused expertise from TV lovers with years of experience is needed. Long may all our learning continue.
If you would like to know more, please email Damien.firstname.lastname@example.org
About The Filter
The Filter offers personalisation of key customer touch points based on proven data science. We put the right content in front of a customer at the right time. More views means more loyalty, lower churn and more revenue.
The Filter prides itself on working closely with clients both during set-up and afterwards to make sure that the client’s customers are always getting the optimal recommendations. We are an agency not a SaaS provider. We have a genuinely deep TV understanding which means our many data science solutions work exceptionally well for your viewers. We work with you to tailor our AI to your customer experience, your audiences and your content. We also integrate effortlessly into your data ecosystem allowing you to benefit from the power of personalisation easily and quickly.