Product embeddings

Question

Hei, I have a list of purchase baskets from customers and would like to build embeddings for the products.

For example:
BASKET1 = ['PRODUCT234', 'PRODUCT214', 'PRODUCT768']
BASKET2 = ['PRODUCT2', 'PRODUCT43', 'PRODUCT7684', 'PRODUCT65']

I was thinking of using something like Word2Vec by using the productIDs composing a basket as words and the baskets themselves would be sentences. The question that I am having at the moment is how to introduce sequentiality to my product basket as, unlike words, I do not have the sequence at which they were added to the basket. Think, for example, about a supermarket basket where items end up being scanned in a random order.

One way I was thinking of introducing some artificial sequence was by ordering productIDs but I do not have any rationale to justify such an approach.

Would you have any comments regarding this approach? Any suggestion as to an alternative model that would produce the desired embeddings? I would like to use embeddings to recommend similar products or products that occur in the same basket.

Answer 1

You can try simpler methods before word2vec type shallow neural networks.

1. for example: Create User product matrix, transpose it to get product-user matrix. Weight them with TF-IDF. Then by applying dimesionality reduction you can take products to lower dimensional space and calculate similarities. User-product matrix can be used to find similar users to later apply collaborative filtering on.
2. Another way is to get product-product matrix and UP`.UP. Then calculate Pointwise mutual information as weighting on that matrix to calculate similarity.

Here"s a sample notebook you can appropriate to your own purpose.