eCommerce Optimization


Advanced Purchase Propensity Prediction Algorithm: Supervised training using dependent variables as customers, pets, store visits, sales, social media, news, products, margins, date, weather, etc.

Trained Model Use Examples:

  • Customer segment most likely to purchase

  • Best product and discount for clearance-products promotions

  • Best days-dates with the highest probability of selling to specific segments


  • Generating initial insights about new commercial actions with a high probability of generating high margins.

  • For this purpose, the following Machine Learning techniques can be used:

1. Train Data Embeddings Space (DES) - Data Vector:

      • Since on ecommerce data there are synonyms and the data comes in various formats, styles and vocabulary, the Continuous Bag of Words (CBOW) technique can be used in TensorFlow.

      • Specific Language pre-delivered word embeddings (300 dimensions) is used to bring them to 200 dimensions

      • With both word embedding spaces, a hybrid space is built from where the word retrieval is performed, which allows the quality of the training to be compared.

      • The training result is an algorithm that learns the data formats of the eCommerce site and leaves them in suitable formats for the platform (Store Vec (SV))

2.Train k-Means Clustering for Customer Segmentation:

      • Why K-Means Clustering: It is on of best clustering models when there are dependent variables (margin, sales, satisfaction) whose result depends on the independent variables that we choose (price, discount, product). For example, what is the margin probability if I launch the promotion?

      • Potential Value (RFM): Customers are segmented with the criterion of potential value for the client

      • The Potential Value Function is trained using the marketing concept RFM (last purchase, frequency, amount = RFM; Recency – How recently did the customer purchase?; Frequency – How often do they purchase?; Monetary Value – How much do they spend?

    • Find Optimal Number of Segments (Ks):

      • An explained variance loop is created in the K-Means Clustering analysis. That is, what percentage of the variance of the data is explained with each number of segments.

      • The optimal number of segments will be the one where, by adding another segment, the increase in the explained variance is infinitesimal.

3.Recommender System V1:

      • User-Item Collaborative Filtering: Used by eCommerce companies like AMAZON to recommend products, it is the training method to use that allow answering: "Customers who are similar to this customer A have a high probability of buying the products customer A buys"

      • Market Basket: This is a "market basket" recommender. That is, the system is trained to "learn" which items tend to be purchased with other items and recommends bundling offering based on those estimations.

      • Rating Based: Information on customer service and product satisfaction can be integrated and produce better trainings.

      • Margin Function Maximization: A margin function determined as the percentage that represents the gross margin of each product is integrated into the Training process.