Rec. Sys For Product Page — High Level Design

How to up-sell and cross-sell your products

Introduction

There is no one size fits all solution to every type of business, hence one
needs to be considerate enough to understand limitations that can help to
build the ultimate product. ML models in development are challenging, but
deploying them to production creates a whole new set of challenges. Model
performance can degrade over time, which can cause losses for the business.
So, throughout development and deployment, it might turn out that the best
models aren’t even the right choice for your project.

When building AI systems, it’s always good to take a divide-and-conquer
approach with your goal. This means breaking down the problem statement
into solvable components, and studying how machine learning could help
alleviate certain problems.

Upselling

Upselling is aimed at increasing a sale’s value by selling a more expensive
option. The product is essentially the same but costs more. The
recommendations in the product page should be more product centric rather
than the user centric. Comparing the primary product with the
recommendations can help users decide based on the values added. A general
recommendation system consists of three stages:

Candidate generator
Filter out most similar products to the primary product. This
helps reduce the number of products to consider which helps in
scalability and removing noise. Similarity based on vectors
derived from:

• TF-IDF using products title, description, category etc.
• Item-Item collaborative filtering: Users’ interaction with the
  product.
• Item2vec: Using word2vec model and taking product
  purchase sequence.

Re-ranker
The filtered out candidates will be assigned a score based on their
relevance to the primary product and the user. Multiple ways the
candidates can be scored:

• Sales in the last 1,2,4..n weeks sales etc.
• Price multiplier i.e. Primary product price/ Candidate price
• Supervised deep learning algorithm that takes the above factors into account and the user purchase history, candidate embeddings, user demographics etc. This approach takes into account the user features that might help in purchase decisions.

Optimisation
In this stage, freshness and uniqueness of the candidates can be
taken into consideration using the click and impression data of
the user within the session or historical or both. Also, filtering out or pushing down the candidates that are priced lesser.

Cross-selling

Cross selling is intended to sell additional products, not change what the
customer already intends to purchase. Ideally any product that compliments
the primary product can help make a cross sell. Following ways can used to
model out cross selling:

1. Market basket analysis

• Use historical transactional data to find out which products are
  added to cart together frequently. If the data is sparse, click or
  browsing history of products within a session can be used.
• To decide the best complimentary product support, confidence
  and lift should be considered.

2. Collaborative filtering

• Helps in finding out which other products similar users had purchased or liked based popularly known as wisdom of crowds.

3. Next best product prediction

• Based on the historical sequentences of purchases/ clicks made
  by users, RNN can be trained to predict the next product.

AB test or MAB can be used to figure out the best strategy through analytics.
Though offline validation is always required before taking the strategy
online.