Some aspects of the use of recommendation systems

Recommender systems daily influence decisions made in the process of using Internet resources. While the development and implementation of significant attention is paid to issues such as the accuracy and privacy of recommendations, the long-term mutual feedback between recommendation systems and user recommendations has never been subjected to serious research.

Despite the widespread belief that recommendations help users find something new, long-term use of recommendation systems can contribute to the tremendous growth in popularity of some elements and, ultimately, to narrow the user's choice. These results are confirmed by some studies in real systems.

Even if we do not notice the effect, our online life depends on recommendations. Popular websites such as Netflix, YouTube, Amazon, and Ozon, in an effort to facilitate navigation, offer opportunities to provide relevant elements. Thus, an increase in user satisfaction is achieved and, importantly, profit. Today, various working recommendation algorithms are widespread, ranging from the simplest options “customers who chose X also purchased Y” to complex options such as singular decomposition .

Although many users still operate independently of any automated assistance, the use of recommendation systems is growing steadily every day. The main feature of any recommendation system is the ability to match customer requests with related products. This task is especially important and difficult for less popular elements for which user patterns cannot be easily identified. Proper harmonization of less popular elements is crucial for e-commerce. Studies have shown that between 20% and 40% of sales on Amazon are not among the most popular products. If you rank the goods and compare their sales, you can get a certain schedule with a long tail of distribution, which contains a large number of niche elements.

In this regard, recommender algorithms make it possible to identify hidden distribution resources with a long tail . Recommendations work by expanding the variety of recommended elements and more evenly spreading the user's attention. However, at the moment, the algorithms implemented on many popular resources do not fully fulfill their functions, which is why the tail of the distribution of popularity is getting shorter. At the same time, the most popular elements still make up a significant share of total sales. Such adverse effects can ultimately lead to a loss of balance in the system.

Algorithms built on the basis of precision-oriented metrics, by their nature, cannot explain this behavior. Although, if you look closely, you can find the inverse relationship between the choice of users and the recommendation system. This interaction is similar for many physical systems. Therefore, it is quite acceptable to use the physical approach for research.

Currently, recommendation systems are most often investigated to achieve short-term indicators such as accuracy and diversity, while it is clear that the use of conventional recommendation algorithms leads to the fact that the system reaches a stationary state in which users are focused on a small number of elements (products), but not distributed over a wide range. In other words, conventional recommendation algorithms ultimately narrow the user's choices and lower information horizons instead of expanding them.

In some cases, hysteresis is also observed., which implies a serious dependence of the system on the initial state. Which at the same time indicates the insufficiency of the current recommendation systems for online stores, applications, search engines, social networks and the media. And also about the necessary compromise between the short-term and long-term effects of the use of recommendation systems in the design and implementation of the next generation of recommendation systems. So at the moment, you need to engage in broader research in this area, since the gains from a more complete use of the entire distribution can be huge and be a kind of reserve for the entire online business.