Well, what would one say is a Product Recommendation System?
Most of you have bought something online and there is a good chance that when you bought a product online, the online system would have recommended something else to you, which is similar to the item you just bought OR is very likely to be bought together. The technology which does this behind the scenes is called a Product Recommendation System.
The reasoning is simple, the recommendation systems uses 3 sets of information to suggest you something
- A buyer’s past buying history
- Buying history of buyer’s who share a demographic similarity to you
- Pattern of products which are always brought together and hence can be reasonably assumed to be complementary to each other
How does one benefit from a Product Recommendation System?
The system, when used correctly will have a dual benefit on both the buyer and the seller. Recommending the right product to the buyer takes out the cognitive stress on a buyer about whether they are making the right decision. And if they do end up buying the product recommended, it adds to the bottomline of the seller. Moreover, the entire experience leaves the buyer with the feeling of having a personal shopper and will make them keep coming back and also having a relatively positive experience with the whole thing.
How does one go about building a Machine Learning system?
As with all Machine Learning systems, the first step is to gather the data. This data can be split into 2 areas – buyer and product data.
Buyer data would be demographic (which is income, education) and behavioural (which is propensity to buy certain kinds of products)
Product data would be attributes of the product – viz. Is it a personal care product, an aspirational product and so on?
The normal machine learning processes of using an algorithm to derive patterns from the data takes over from this point onwards.
What are the different kinds of Product Machine Learning Recommendation Algorithms?
The 2 main buckets into which most of the recommender algorithms can be classified into are – Content based and collaborative filtering
Content-based algorithm: which focuses on a buyer by itself and ends up recommending products which the buyer may have bought in the past
Collaborative algorithm: would be based on 2 buyers sharing similar characteristics and have bought similar items in the past and then recommending the exact item from each buyer to the other buyer. Of course, it is assuming that the other buyer has not bought the exact same item.
What problems are faced while building or deploying Product Recommendation Systems
Let’s focus on 2 problems – Cold start and Data Scarcity – both of which are inter-related.
Cold Start
This phrasing could be inspired from starting a car in cold weather. So the problem here is when either a buyer or a product is new to market and there is no history behind it.
The solution for the new buyer would be to pitch the most popular products in each category.
The solution for the product is to use the attributes of the product and then pitch to buyers who buy similar products.
A proper solution for both would be to solve the Data Scarcity problem
Data Scarcity
There are multiple mathematical approaches to solve this problem.