Getting the most from recommendations

Recommendations are everywhere, you get them from friends and family, you see them when purchasing items from websites, and (surprise, surprise) you even get them on your TV. Implemented well they can be an extremely powerful asset, helping to reduce customer churn and increase engagement. However, we often see that recommendation algorithms are just added to OTT platforms without much thought as to what purpose they are serving, which can often lead to less than desirable results.

So how do you get the most from recommendations?

Align recommendation strategies against your objectives

Different types of recommendation strategies exist to help achieve specific objectives. Understanding how best to leverage these strategies and measure results accurately is fundamental to success and often neglected. In order to understand how recommendations can help, it can be useful to think of them in the following ways:

  • Personalised recommendations – these are generally the most sophisticated, using a mix of both content and user data to surface relevant recommendations. They are unique to each user and are a great way of showcasing highly relevant content to users which they may not be aware of. These recommendations are often labelled on OTT Platforms as ‘Recommended For You’, and ‘We think you will like’. These can be extremely useful for platforms which have large catalogues, covering a wide range of tastes. But beware, unless you get the right model combination you could end doing more harm than good, a poor recommendation can also leave a user with a negative sentiment if it does not match what they would expect to see.
  • Global recommendations – based on content popularity and trends, these recommendations can be extremely effective in showcasing content which has wide appeal and is proven to get the attention of viewers. You often see OTT platforms label these types of recommendations on their players as ‘Most Popular’, ‘Trending Now’, and ‘Top 10’. It is important however to make sure the parameters for these algorithms are tuned correctly, as often you can see recommendations which are very static and rarely change. This is typically due to either the date ranges the algorithm use being set too wide or the unit of measure used to calculate popularity being incorrect.
  • Contextual recommendations – these are typically tailored according to content attributes (e.g. genre, mood) and are a great way of recommending content which is similar to what the user has watched previously. They are often labelled on OTT platforms as ‘More Like This’ or ‘Because You Watched’ and are a highly effective way of showcasing relevant content, which may not be new or well-known. These algorithms can be a highly effective way of showcasing the types of content the user loves, however, when used in excess can create ‘filter bubbles’.

Each of these recommendation strategies have strengths and weaknesses and when implemented effectively can help meet your desired objectives. For example, if you identify that users who watch lesser-known titles within your catalogue are less likely to churn you may want to increase the prominence of personalised and/or contextual recommendations. This does not mean that global recommendations should not be used, just that in this case they are not the best way to meet this objective.

Test and learn

Recommendations are fundamental to the success of many OTT platforms (just look at how they have helped Netflix), which is why you will find that most successful services will be using all 3 types of recommendations in varying ways. However, it is important to understand that just putting recommendations on the platform will not gain you instant results, especially if the models have not been tuned and optimised.

Adopting a culture of experimentation is fundamental – rail positioning, model blending (e.g. mixing recommendation strategies into one set of results), and even the naming of rails can have a large impact on how users engage with your platform. Some changes will have positive impacts, whilst others will not. Running experiments will enable you to identify the good from the bad, allowing you to take a data-led approach to improving your customer’s experience.

At The Filter we are constantly working with our clients to help them shape their objectives and optimise their platforms. We always split test our models across user groups to measure effectiveness and demonstrate success or failure, using our analytics tools to measure the metrics that matter and help ensure you are getting the most from recommendations. Our ever-growing suite of personalisation solutions means our clients can quickly and easily utilise proven solutions which can be tailored to their needs.