Business, Data, And The Reality Of Recommendation

Welcome to a world run on recommendations…

For businesses, there is little of more importance than connecting with customers. But how do they do it? They come up with a decent product, make sure it fulfils an unmet need, and get their marketing right. Each of these stages can be broken down and subdivided into a long list of processes that could turn consumers off at any point.

However, there are various ways that companies can use data to work out what users might want to consume next. Those insights then become targeted recommendations. This can include simple information retrieval as well as more complicated matrix factorisation techniques. But what considerations should businesses make before using a recommendation engine?

The rush for recommendations

Generally speaking, the aim of recommendation is to connect with customers. Matrix factorisation is just one of the many methods that companies can use to generate these suggestions. It attempts to explore underlying features in interactions between users and content. Netflix, for example, uses this method to fill in the gaps in user ratings and work out if an individual would enjoy a certain programme or film based on matrices such as cast members, setting, and genre. Spotify uses a form of matrix factorisation called Logistic Matrix Factorisation to suggest related artists. So, if you listened to AC/DC, you might be a fan of Led Zepplin too. But, for Spotify at least, recommendations have another advantage in reducing costs. By recommending lesser known artists, the company can cut down on licensing fees and the advantage to the consumer is novel discovery.

Undoubtedly the biggest benefit of recommendations, though, is pushing the consumption of content, products and services with a focus on personalisation.  Accurately targeted suggestions are already having a positive impact in consumer facing industries like hospitality, healthcare, retail, and of course entertainment. B2B relationships could profit from this model too, making the customer journey easier for clients and encouraging engagement.

That said, there are some instances in which recommendations could be seen as unhelpful. This largely depends on the consumer – some people, for example, may not want visual evidence that their viewing habits are being analysed by an algorithm. They may also want to expose themselves to new things, and enjoy a level of randomness. Unfortunately for them, most content distributors now rely on recommendations. If someone is surrounded by recommendations based on what they would ‘usually’ or ‘normally’ consume, then spontaneity is removed. Restricting the scope and variety of what someone is exposed to forces them, albeit unwittingly, into a niche. Ultimately, this could reinforce bias.

Key considerations…

For the most part, generating accurate recommendations is a great way to engage with audiences and improve both the reach and popularity of products and services. Through recommendation engines, businesses can get closer to working out exactly what consumers want. But before they can do this, they need to have a firm grasp of their main purpose – for example, is the goal to improve retention rates, or increase sales? Another major consideration should be the variety of user profiles. What method will work best for specific demographics? Then there are the ever present challenges of operational costs and technical challenges. What resources are needed? Does it mean hiring or retraining staff? Answering these questions is the first step – the next stage is to pick the most appropriate method, and to test it as extensively as possible.

Giant content distributors (think Amazon, Netflix, Spotify, YouTube…) have done well from the recommendation business model. Recommendations based on consumer habits seem to appeal to individuals but can quickly snowball into a lack of variety. Sometimes, the element of surprise can be just as effective.

For DISRUPTIONHUB’s weekly insights straight to your inbox, sign up for our newsletter here.