In the overall link of real-time bidding advertising, the media side initiates a request and sends it to each DSP through ADX. The DSP returns the advertisement to participate in the bidding based on the traffic evaluation. After the ADX compares the prices, it returns the winning advertisement, and the media displays and reports to complete the entire advertising process.
Among them, after DSP receives an advertisement request, it needs to find out a matching advertisement and return it to ADX within 100-200ms. The number of candidates in the advertisement library is huge, generally between hundreds of thousands and millions, and it needs to be done in a very short time. To complete the phone number list scoring and evaluation of massive advertisements, the common practice in the industry is similar to the recommendation system, which is divided into two main links: recall + sorting.
The purpose of recall is to reduce the number of candidates (try to keep it within 1000), so as to facilitate accurate sorting using complex models in the subsequent sorting process; because a large number of candidates are evaluated in a short phone number list period of time, the key point of recall is a fast word, which is limited by this and sorting In contrast, the algorithm model of recall is relatively simple and uses fewer features.
The method commonly used in the industry is multi-way recall, that is, from multiple dimensions, the candidates with high correlation are found as much as possible in the massive library.
Multiple simultaneous recalls are due to various considerations:
Diversity, starting from different dimensions to find relevant candidates;
Robustness, even if there is a problem with the recall, other recall channels will operate normally and will not block the main process;
Interpretable and flexible, each channel can explain the logic of recall from a separate dimension, and if the effect is not ideal, it will be less complex and more flexible to adjust.
When a certain recall is selected, the direction is selected to determine the corresponding scoring function, and then the scoring, sorting, and truncated recall topN are performed. Each recall is independent of each other, and the winning candidates are not comparable.
Recall mainly starts from the general directions of user (U), context (C), search term (Q), and advertisement (A). The refinement dimension can be combined with actual business scenarios and can be based on contextual title/description/category/tag/picture , user basic statistical information/interest tags/historical behavior, search terms, item-based collaborative filtering, and the processed information is multimodal, including text, pictures, videos, etc.
There are two ideas for specific recall matching:
Text hard matching based on tags/keywords, such as region and gender in ad targeting, either matches or does not match, with poor scalability and flexibility;
Vector-based semantic soft matching. Select effective original features for embedding in vector space, learn to obtain the vector representation of users and advertisements through the twin tower model, and use dot product, cosine similarity or Euclidean distance to calculate vector correlation, similar to Youtube's recommended twin towers and Microsoft DSSM; it can not only adjust the number of recalls by changing the threshold but also meet the performance requirements, which is the mainstream form of current recall.
Unlike the sorting stage, recall does not directly affect business indicators. After rough sorting and fine sorting, the impact on the final result is small, and it is relatively difficult to measure quality. You can try to evaluate from two aspects:
Uniqueness, the irreplaceability of the recall result of a certain channel or the repeatability with other channels, the higher the repeatability, the smaller the recall value of the channel;
Conversion effect, follow-up performance of recall results, such as ranking after refined ranking or whether users click after exposure, the better the effect, the higher the value.
If recall is to provide possibility, ranking is to provide certainty: find the most suitable candidate and push it to the user.
Sorting refinement can be divided into coarse sorting, fine sorting, and rearranging.
The number of candidates returned in the recall phase is too large for the rough sorting, and the direct processing performance of the fine sorting cannot be satisfied. The rough sorting uses a simple model to filter again to reduce the number, which is an optional link;
Rearrangement is to process the refined results for business considerations: diversity, frequency control, category control, specific result escalation, etc.;
Refinement is the key to the entire sorting, the main battlefield of various models, and the protagonist of our discussion.
Compared with recall, the number of candidates faced by sorting is drastically reduced: only the candidates that win in the recall stage need to be processed. "If the magnitude of the entire creative library is around a thousand (such as open-screen ads/store native, etc.), you can all Recall is directly used for sorting"; this lays a good foundation for sorting: use more features and feature combinations, complex models to uniformly evaluate and score the results of multiple recalls, sort, and truncate the topN output.