Product recommendations for personalized mailings and website

General Information

Product recommendations are a tool for creating effective customer communications via email campaigns, as well as widgets on the website and in the Mobile Application.

Key objectives of product recommendation tools:

• Personalize interactions with customers;
• Encourage customers to perform specific actions aligned with the Company’s goals and interests;
• Achieve business objectives;
• Automate marketing activities and test hypotheses.

Product recommendations are generated using the Loymax AI module whose algorithms enable fast and accurate processing and analysis of large data volumes. Marketers and analysts can focus on developing business strategies while delegating technical big-data processing tasks to artificial intelligence for specific business purposes.

Business goals achieved through personalized communications using product recommendations:

• Improving customer interaction quality with the Company;
• Enhancing customer experience through recommendations;
• Suggesting alternative products;
• Expanding the shopping cart;
• Increasing purchase frequency;
• Increasing average order value;
• Promoting specific products, product lines, and brands;
• Clearing out inventory;
• Other business metrics marketers aim to influence through campaigns and customer engagement scenarios.

Product recommendations are generated using statistical analysis and machine learning algorithms. These algorithms leverage both historical data and real-time online actions by customers on the website. They analyze preferences, predict behavior, and deliver the most relevant product recommendations for each individual customer.

Interaction of Loymax Infrastructure Components for Generating Product Recommendations

Data sources for receipts, customers, and products from the product catalog used to generate product recommendations include the DWH BI data warehouse and/or the ClickHouse SmartCom database management system. This data is transferred to the Loymax AI module, where it is processed and product recommendations are calculated using ML algorithms. The resulting recommendations are then delivered via API to data marts. The API-proxy server validates incoming requests and formats responses for Smart Communications. Smart Communications retrieves responses from the API-proxy server, enriches them with data (images, product names, links, prices, etc.), inserts them into predefined templates, and sends mass messagings containing the generated product recommendations to customers.

Currently, 5 algorithms are available for generating product recommendations:

Recommendations produced by each algorithm differ because they serve distinct business goals and are based on different models, algorithms, hyperparameters, etc. When selecting a specific algorithm and usage scenario, it’s essential to consider multiple factors:

• The objective of the specific marketing campaign (information/communication);
• The desired customer action following the communication;
• Customer touchpoints and content delivery channels.

This means that combining multiple recommendation algorithms can lead to the most favorable outcome for the Company and effectively motivate customers to take specific actions.

Communication channels using product recommendations:

1. Email,
2. Website widget (a graphical application that displays content on a desktop, smartphone/tablet screen, or web page),
3. Mobile Application,
4. Personal Account.

“Popular Products” Algorithm

1. Description of the Popular Products Algorithm

The Popular Products algorithm calculates a ranking across the entire* product list. It recommends the top-N most popular products from each category based on each product’s position in the overall ranking.  
This approach recommends a broader range of products (i.e., several items from different categories), which is preferable for achieving business goals.  
Product rankings are calculated based on total units sold across all customer purchases—i.e., the most purchased items.  
All purchases include both online and offline transactions processed through Loymax.

2. Use Cases for the Popular Products Algorithm in Personalized Campaigns

Email Campaigns:

3. Examples of Using the Popular Products Algorithm for Website Personalization

4. General Concept of the Popular Products Algorithm

Example: Generating recommendations by selecting the top 1 most popular product from each category.

“Popular Products in Category” Algorithm

1. Description of the Popular Products in Category Algorithm

The Popular Products in Category algorithm calculates a popularity ranking among products within each specific category.

It recommends the top-N products from the category.

Product rankings are calculated based on total units sold across all customer purchases—i.e., the most purchased items.  
All purchases include both online and offline transactions processed through Loymax.

2. Use Cases for the Popular Products in Category Algorithm in Personalized Campaigns

Recommended products are selected based on the customer’s previously viewed/ordered/purchased items and the category (i.e., category ID) to which those items belong.

The category/classifier level assigned to a product can be either the lowest or a higher level in the hierarchy.

Email Campaigns:

3. Use Cases for the Popular Products in Category Algorithm in Website Personalization

Recommended products are selected based on the customer’s on-site behavior (event tracking) and the category (i.e., category ID) of the viewed product. The classifier level (i.e., category) can be configured—it may be the lowest level or a higher-level category.

4. General Concept of the Algorithm

Product rankings are calculated at each classifier level (Group ID, Subgroup ID, Category ID) across all products belonging to that hierarchical level—i.e., all items in the group/subgroup/category—regardless of deeper nesting.

For example, if a Subgroup ID (a higher-level classifier than category) is provided in the request, the ranking is calculated across all products in every category belonging to that subgroup (see example below for classifier level 2).

Example: A customer is viewing Product 6. The diagram below illustrates the 5 product recommendations they might receive when querying by Category ID, Subgroup ID, or Group ID.

“Similar Products” Algorithm

1. Description of the Similar Products Algorithm

The Similar Products algorithm recommends alternative products* that are the most similar to the viewed/ordered/purchased item.

It recommends N products that have similar attributes to the viewed/ordered/purchased product.

2. Use Cases for the Similar Products Algorithm in Personalized Campaigns

Email Campaigns:

3. Use Cases for the Similar Products Algorithm in Website Personalization

4. General Concept of the Algorithm

For each product, a neural network generates an N-dimensional vector representation (embedding) based on product descriptions, attributes, purchase/view statistics/content. The 3D space shown visualizes high-dimensional vectors (dimensionality > 50). Each point represents a product. Larger points indicate clusters of closely related items. Nearest-neighbor clusters form with minimal intra-cluster distances and significant inter-cluster separation. Similar products are selected based on minimal cosine distance between vectors—i.e., those sharing the most similar combination of product attributes.

“Bought Together” Algorithm

1. Description of the Bought Together Algorithm

The Bought Together algorithm recommends products that most frequently appear in the same transaction as the viewed/ordered/purchased items.

2. Use Cases for the Bought Together Algorithm in Personalized Campaigns

Email Campaigns:

3. Use Cases for the Bought Together Algorithm in Website Personalization

4. General Concept of the Bought Together Algorithm

This algorithm incorporates a series of models and techniques based on the Association Rule Learning methodology.

Association rules for product recommendations consist of logical rules that identify relationships between products and generate recommendations based on those relationships. These rules help predict which products may interest a customer and offer relevant suggestions based on prior purchases or current browsing behavior, thereby enhancing personalization and customer satisfaction.

At first glance, each customer’s in-store purchases appear unique. However, customers develop similar behavioral patterns driven by shared needs. People's needs often overlap, allowing us to identify typical purchasing behaviors under specific conditions.

Association rules can be complex and include multiple conditions.

Assume we have customer purchase history data indicating which products were bought together. Association rules identify product relationships and generate recommendations accordingly.

“Personalized Product Recommendations” Algorithm

1. Description of the Personalized Product Recommendations Algorithm

The Personalized Product Recommendations algorithm generates product suggestions based on an individual customer’s behavior (Customer ID) and the behavior of similar customers.  
In other words, it predicts unknown preferences for a specific user by leveraging preference data from a group of users.

To predict optimal customer preferences, the machine learning model uses data on customer behavior and purchase history, online activity, customer profiles and attributes, product attributes and statistics, and other parameters.

2. Use Cases for the Personalized Product Recommendations Algorithm in Personalized Campaigns

Email Campaigns:

3. Use Cases for the Personalized Product Recommendations Algorithm in Website Personalization

4. General Concept of the Personalized Product Recommendations Algorithm

The Personalized Product Recommendations algorithm is an ensemble of multiple models and algorithms. Its foundation is the Collaborative Filtering technique.
Using this method within an ensemble of models—alongside other algorithms—produces more relevant product recommendations for each customer.

Collaborative filtering is a method for generating predictions (recommendations) in recommendation systems by leveraging known preferences (ratings) from a group of customers to predict unknown preferences for another customer.

The core assumption of this method is: customers who have purchased similar products/categories in the past are likely to make similar future purchases of other products they haven’t bought yet—but that their nearest “neighbors” (i.e., customers with highly similar purchase histories) have purchased.