Asset Embedding

What it Means and How to Use it

What is Embedding?

Embedding is a relatively new concept in Artificial Intelligence. It first emerged from the research done in Natural Language Processing (NLP) where researchers wanted their AI algorithms to really understand a given text as opposed to treating it as bits and bytes. The benchmark for understanding here is the human intellect. When we hear the word apple, depending on the context, we think of it as either a fruit or a company. If it is a fruit, then our brain understands it as something which is close to (or conceptually related to) entities like an apple tree as well as functions like eating and smelling. If on the other hand the context implies that it is a company, then our brain associates the word with such entities as a Mac Book, iPhone etc. In other words our brain identifies the meaning of the word using the context and more importantly associates it with related concepts.

To achieve a near-human understanding of the written text, the NLP researchers designed neural networks and fed it with huge amounts of text. The job of this neural networks was to find a representation for each word which encapsulated the context-aware meaning of the word in it. This representation is nothing but a vector of real valued numbers. The size of this vector could be anything from 50 to 500. But as the magic of mathematics has it, the resulting vectors really did capture the meaning (a.k.a semantics) of each word. For example, the vector representing the word Hockey would be more similar to the vectors representing the name of the countries where hockey is a major/national sport compared to other countries and indeed all other words. What's more is that if you sum the vectors of "Hockey" and "Canada", you would get a vector which is very similar to those representing famous Canadian hockey players. We call the process of generating these vectors Embedding. The word embedding ultimately enabled inventions like chatbots.

Embedding Extends Its Scope

The success of word embedding in NLP was so great that researchers soon scrambled to see if they can find other applications for it, and this they found. For instance, we now have patient embedding where all the health related information of a patient is considered a person's health context. Given this context and, again, using deep neural networks the patient vectors were achieved. These vectors too have beautiful mathematical properties. You can find which chronic diseases are more similar to the patient vector and by doing that estimate the likelihood of this patient getting that disease. Other applications are also possible.
At this point you might ask, ok, for word embedding, we had textual data, what type of data do we have for patients which enables embedding? The answer to this question is: Knowledge Graphs. What happens in case of patient embedding is that the data is actually a knowledge graphs where patients, doctors, diseases/illnesses and drugs and medications are all linked to one another according to the health records. A deep neural network uses this knowledge graph to embed each entity's meaning and generate its vector.


Asset Embedding

The same idea, generally speaking, applies to financial, social and political entities. Given a knowledge graph which connects companies, countries, governments, individuals, events and much more, we use a deep neural network to perform embedding. When the entity is a financial asset we can rightly call it Asset Embedding.
Though the general concept of embedding is similar across application areas, each of these fields have their own challenges. However, financial and media entities have a lot of unique features which when considered properly in the mathematics of AI, it will give rise to a large number of innovative concepts. These concepts, complex in nature and unachievable using traditional machine learning models are extremely exciting. Please read our article about them here to see the many things that we can do using asset embedding. Due to the limitless opportunities which are now possible using our asset embedding, we have decided to provide the Asset Vectors in their raw form as an additional data product to our clients. It would be great to know what novel concepts you can invent with these vectors. If you do, please contact us to have what will surely be an exciting conversation with us.

Generic placeholder image

Founder & CEO

March 10, 2021