Injecting GPT-4's reasoning into recommendation algorithms

Summary - injecting reasoning into recommendations

Redefining recommendation algorithms as a reasoning problem rather than just a pattern recognition task could give us a completely new way to make recommendations, aligning with what users might actually need, such as forgotten or complementary items to buy. We should explore using GPT-4 to inject reasoning capabilities into making recommendations to our users.

How recommendation algorithms work

Our current recommendation algorithms are based on recognising patterns and making inferences about future behaviour, such as:

• Past behaviour: predict what people will buy based on previous purchases
• User similarity: people of your profile or behaviour bought something similar
• Item similarity: if you bought an item, you might buy a similar one

These ideas kind of work - we do buy similar things to other people and we do tend to buy things we bought before, but we have all experienced how badly recommendation algorithms fail. For example, if I have purchased an umbrella, I'm probably not buying another umbrella for another 5 years so existing algorithms struggle to make any sensible recommendations.

Recommendation algorithms have a reasoning problem

Patterns are important, but it is hard to recommend something if you don't take into account what the customer is trying to accomplish with this purchase. Our current algorithms have been missing a key ingredient - reasoning. Using more abstract ways of thinking we can now ask questions like:

• What is this person trying to do?
• What could help them accomplish this task?
• What products could we recommend given the above?

So for example, if a person is buying crisps and beer, perhaps we shouldn't be recommending them cider or gin, but instead reason that they are probably organising a party and might need paper cups or a party game.

Using GPT-4's reasoning for recommendations

GPT-4's reasoning capabilities have been barely explored and I think there is a lot of scope to apply them to even tasks that seem to have been fairly well solved by traditional data science techniques. Below is Open AI's Playgroup with a system prompt that instructs GPT-4 to imagine what the user might be doing and what would implement their experience. As you can see, GPT-4 is giving pretty unusual but quite interesting set of recommendations that do not rely on previous purchases or what other customers have bought.

You can try it out for yourself, I have included the system prompt in 'Alt Text' of the image below. The prompt can be tuned a lot depending on the use case, e.g. target value or typical purchase frequency desirable recommendations can be additional requirements, but the core idea of trying to understand the user can be very powerful.