Content recommendation method based on rank algorithm

Abstract

The invention discloses a content recommendation method based on a rank algorithm. According to the method, model training and sequence optimization are respectively carried out by using Pointer Network and reinforcement learning, and the method sequentially comprises the following processing flows of original data, preprocessing, input layer, coding layer, decoding layer, optimizing, evaluating and loop iteration, wherein the loop iteration process is that data is optimized and then returns to the input layer to be coded, decoded, optimized and evaluated. The method has the beneficial effectsthat a rank algorithm is provided to solve the potential influence and association between items and between user sequence behaviors, so that the overall long-term click / conversion rate is optimal; on one hand, modeling can be carried out on the relation between the items, and it is avoided that related clicks / conversion rates are independently calculated for the items in recommendation; and on the other hand, the long-term sequence behavior feedback of the user can be captured, so that the long-term overall target is optimal.

Description

technical field [0001] The invention relates to a content recommendation method based on a rerank algorithm, and belongs to the technical field of content recommendation. Background technique [0002] In the recommendation scenario, the content recommended to users will be restricted by various factors, such as content quality, user fatigue, richness of content, etc. From the perspective of content display, appropriate richness, category breakup, copywriting, color combination and other factors will greatly enhance the user experience, and thus increase the click / conversion rate. Therefore, the recalled massive content (item) is generally sorted twice, that is, rerank, in order to achieve the desired effect of recommendation. [0003] The traditional recommendation sorting algorithm based on learning to rank is difficult to consider the impact of different permutations and combinations of items on the click-through rate, such as pointwise and pairwise, generally optimize th...

AI Technical Summary

Problems solved by technology

[0004] Although the Listwise method can take the entire candidate item as the learning target and directly optimize the document sorting results, it increases the training complexity by calculating the loss based on the permutation.

Moreover, it cannot simulate the connection and influence between items, and it is also difficult to capture the sequential behavior of users. Many studies have shown that user behavior is not completely independent, but has a certain relationship.

Images