Model Building in B2B SaaS

For B2B Space, data from each customer is likely to be small and data between each company will have notable similaries, yet each customer will have its own flavour.

Baseline for B2B ML feature would be Non parameter model say, K-Nearest-Neighbours. For the nlp specifically, taking top k results from elastic search, cosine similarity on universal sentence encoder or sentence transformer. There is no model building effort in K nearest neighbour excepts that finding optimal k for the given problem. For same cases, it is the best choice

Level 1 is to always build a different model for each customer. That would definitely work. But we are missing out the learning from different customer.

Some problems are agnostic of the customer. Example. SA is the best example. Fine-turing the Ulmfit or Bert model for the all customer will result in better improvement than single source of data. Before that, it is important to get approval from the customer to do that.

For problems like CRS or FC is not possible as each company will have a set of classes. In that case, the Language Model training building from data from all companies will be helpful and later train classifier for each model.

This applies to the all the problems such as text, image, audio that Bert/Ulmfit Architecture can solve.

Tabular data like CCR can be not solved using Bert. Some ways to tackle is.

• Add a categorical variable (Customer type) and build the model from all the customer data.
• Create embedding for the categorical variables from all customer data and use the embedding features and numerical feature to train for the single customer
• FT-Transformer is the recent transformer for tabular data. Paper claims that FT-Transformer could not beat Xgboost in data set. More research is yet to happen :).

Next article will talk about improving the model in production. Please click here for next article

Please click here to go to the first article