fdecomite – “Gems VI”
Businesses commonly have to make recommendations to their customers. These recommendations can be for products, pages, actions, partners, or many other aspects of their ecosystems. While it’s easy to dip your toes into making recommendations, you quickly hit a wall with regards to their effectiveness, nuance, and sparsity.
Part of the reason for this is due to the accessibility of the features of data that companies keep. For example, a t-shirt selling company may want to recommend t-shirts to customers, however, they can only access limited information about their t-shirts:
And similarly, only simple information about their customers:
Information about a customer’s amenability to future recommendations is highly coupled to information that can be extracted from other products they have looked at in the past. This situation leads to great recommendations for very popular items that have been tagged accurately with metadata, but there are almost certainly “hidden gems” in the alternatives that lie there, inscrutable.
While we bemoan the meagre information we have about the unseen items in our catalog, a lack of data isn’t strictly true…
The problem isn’t that there is too little data. The problems is that there is actually too-much uninterpretable data. What we would love is a function that would allow us to crunch our opaque data into a more meaningful form:
If we had such a function, not only could we automatically tag all of our catalog with useful metadata, but we could simultaneously build much more useful models of our customers – inferring a great many more actionable recommendations from their profile and history based on the additional nuances of the products that they have purchased or browsed.
We want to take high-dimensional, intractable data:
… and transform it into low-dimensional, actionable data:
Deep Learning is one approach that can be taken to mining hidden gems. The beauty of using deep learning is that the solution is so general. While enough domain-knowledge is required to determine if outcomes are effective, far-less is needed with regards to designing an algorithm to perform the task. The application of deep learning to this task can be done by experts in machine learning, who don’t necessarily have an understanding of your business.
If you’re still skeptical about the utility of mining hidden gems, then take a look at the Netflix Prize. Netflix was awarding one million dollars for a better recommendation algorithm – and they actually had great metadata, and were already doing a reasonable job! Imagine how much more of an impact hidden gems could make in a business where metadata is scarce and recommendations constricted.
Of course there are pitfalls to be avoided when taking this route. It is important that, while the neural network tasked with mining hidden gems needn’t know about your business, its mistakes and victories must be able to be analysed in the context of business needs and rules. It is also very useful to be able to use deep learning as an engine for generating fresh insight into your business. The black-box of deep learning must be cracked open, or at least made translucent to achieve these goals.
Sadly though, this post has reached its conclusion, and the interoperability of deep networks is a topic for another day.