Author: Roman Zykov, database expert, independent AI consultant at LENNUF.ru
When we talk about personalized recommendations, we often forget that there are different types of recommendations. In this article, I will look at the main differences between long term and short term recommendations.
Personal recommendations mean that we recommend something to the user, taking into account his personal interests, taking into account his signals. The signals themselves are of two types - explicit, when the user says what he likes (like, rating for a song, review for a product); and implicit, when the user performs some actions (looked at the product, listened to a song, looked for something), by which we cannot unambiguously evaluate whether the user likes / dislikes something. In the first versions of recommender systems, it was customary to use only explicit data, this can be seen both in the literature and in scientific articles of that time. There is usually much more implicit data - remember how many products you looked at (implicit signals) before buying,and then for how many products did you leave reviews (clear signals)? Therefore, in the last decade, there has been a shift towards implicit recommendations. Even Netflix representatives announced the importance of such data at the RecSys conference 5 years ago. How can we see this? For example, recommendations on youtube are constantly being adjusted according to the videos that you watched and those videos that you liked. The same is done by recommendation systems in streaming music services, social networks and online stores.The same is done by recommendation systems in streaming music services, social networks and online stores.The same is done by recommendation systems in streaming music services, social networks and online stores.
Let's go back to the issue of long-term / short-term recommendations. By long-term interests, I mean user interests that are relevant over a long period of time. These can be product categories, brands, genres of music, bands, etc. Under the short-term interests of the user - those that need to be instantly satisfied for his current needs. For example, a girl chooses boots - color, heel length, sole thickness, price, brand - are quite suitable for short-term interests. Both types of recommendations should satisfy these interests. For music - rhythm, presence / absence of vocals.
Long term recommendations
, (Recency, RFM!) (Frequency) , , … 80% . “” , -, . , , . , , .
, - , , , (batch) . , , - Recency Frequency. .
. , . - , . , , - "Application of Kullback-Leibler divergence for short-term user interest detection”.
, KL ( ) (, , ..). - . “” “”, . ? , , , .
- real-time , . . KL- - embeddings .
/ (recall- diversity tradeoff). , . , - , . “” , . explore-exploit reinforcement learning, . , (diversity) / .