Evolution of approaches to personalized user experience in retail

• Recommender systems and their development
• Basic approaches to creating recommendations
• User behavior and interests
• Inside the ecosystem
• Existing recommender systems for e-commerce
• Problems with Classical Recommender Systems
• Virtual assistant
• Key findings
• Questions and Answers

In today's world of e-commerce, personalized customer experience is becoming a key success factor. Recommender systems, which have undergone a long evolution since the 1990s, play a key role in this process. However, classic approaches face a number of problems, such as the echo chamber effect and excessive personalization, which leads to a limited choice of products for the user.

In response to these challenges, ecom.tech has developed an innovative approach - a virtual assistant capable of responding to the actual needs of users in real time.

Technical implementation

The system has evolved from simple filtering methods to the use of advanced technologies. The development went from basic collaborative filtering through contextual analysis and hybrid models to matrix recommendations and deep learning, reaching the level of modern dynamic recommendations and the use of large language models (LLM).

The virtual assistant is implemented as a test Telegram bot with plans for integration into the company's main application. It is capable of conducting a dialogue with the user, analyzing their current needs and generating personalized offers in the context of the client's daily life.

Recommender systems and their development

Since the 1990s, recommendation systems in e-commerce have gone through several stages in their evolution. In the process, approaches to filtering users have changed.

Among the important stages are:

• Collaborative filtering. Basic and simplified ranking system without using a personalized approach. Similar products are offered to the same users;
• Contextual analysis. Filtering is based on content. If a person buys something, then similar products are given to him in the recommendation;
• Hybrid model. Combines collaborative and contextual filtering, reducing the disadvantages of each;
• Matrix recommendation. The result of adding the basic, contextual approaches and the environment analysis. Focused on target groups.
• Deep learning. Recommendations are based on studying a significant amount of information of different directions, using neural networks;
• Dynamic recommendations. Used since the early 2020s. Work in real time with the study of all current changes in the user's status - mood, location, preferences;
• LLM. Large language models process text, images, video, audio, analyze context and dependencies.

Each new generation of models has improved the accuracy, relevance, and personalization of recommendations.

Basic approaches to creating recommendations

Systems are built taking into account various factors, which allows us to highlight their features. The most common ones are related to the study of user behavior and interests, including actions within the created ecosystems.

User behavior and interests

Studying behavioral factors and buyers’ interests creates a significant database for recommendation systems. The approach increases consumer satisfaction and improves conversion rates into purchases. According to the international consulting company McKinsey&Company, recommendations generate about 35% of Amazon’s sales. Netflix claims that 80% of content views are related to personalized offers that are visible to users.