For an online retailer with an e-commerce website, we were able to follow the sequence of actions of users and analyze what is the product they are most likely to buy or to predict if they are going to cancel the subscription. The input to the algorithm was the sequence of actions with respect to time:
- browsing of certain products and categories,
- history of completed orders,
- history of interaction with promotional emails and reminders,
- history of links visited on the website, etc.
The algorithm was able to identify which actions lead to which outcomes, and thus allowed customer service to attempt to retain some customers by appropriate and personalized promotions.
The recommender system utilized a hybrid approach combining content-based recommendations and collaborative filtering. The latter captured implicit feedback based on the following actions:
- whether the customer added something to their wish list or shopping cart;
- what was bought;
- whether the customer returned an item;
- whether he previewed something and never bought it, etc.
This was used for recommending more relevant items, that ultimately increased sales (evident through transaction history) and customer satisfaction (registered by surveys).
By applying techniques for affinity analysis and association rule learning we built another module that identifies combinations of products that frequently co-occur in transactions. The system was successfully applied by the retailer for:
- Targeted and personalized emails to past customers about promotions of products that frequently were bought together with some of their recently purchased items by other customers.
- Suggesting bundles of products that can be sold for a discounted price compared to the sum of prices of the individual products.
- Augmented recommender systems so potentially products are recommended when browsing a particular product.
We built a system for weekly forecasts of consumption using trend analysis based on historical data about sales, stock on a particular point in time for products, product categories and subcategories. It featured warnings about low stocks of items that would be in demand in the following weeks. As a result, with the implemented system the frequency of how often there were items out of stock but in demand was reduced to almost zero, without piling up high quantities of items on stock.