A cold start is a typical situation when a sufficient amount of data has not yet been accumulated for the correct operation of the recommender system (for example, when a product is new or is just rarely bought). If the ratings of only three users estimate the average rating, such an assessment is not reliable, and users understand this. In such situations, ratings are often artificially adjusted.

The first way to do this is to show not the average value, but the smoothed average (damped mean). With a small number of ratings, the displayed rating leans more to a specific safe average indicator, and as soon as a sufficient number of new ratings are typed, the averaging adjustment stops operating.

Another approach is to calculate confidence intervals for each rating. Mathematically, the more estimates we have, the smaller the variation of the average will be and, therefore, the more confidence we have in its accuracy.

We can display, for example, the lower limit of the interval (low confidence interval (CI) bound) as a rating. At the same time, it is clear that such a system is quite conservative, with a tendency to underestimate ratings for new items.

Since the estimates are limited to a specific scale (for example, from 0 to 1), the usual methods for calculating the confidence interval are poorly applicable here, due to the distribution tails that go to infinity, and the symmetry of the interval itself. There is a more accurate way to calculate it—the Wilson CI.

The cold start problem is relevant for non-personalized recommendations too. The general approach here is to replace what currently cannot be calculated by different heuristics—for example, replace it with an average rating, use a simpler algorithm, or not use the product at all until the data is collected.

Another issue that should be considered when we develop a recommender system is the relevance of recommendations, which considers factors other than the user's interests—for example, it can be the freshness of a publication or a user's rating.