First, the application scenarios
When the recall items based on user interest, each user has interest tags, can sometimes be very much interested in labels, each label has calculated weights to sort from high in the end. Making recommendations, we were in the end what interest-based labels recommend it, select only topN do, or all? If you select only topN, that every recommendation results are relatively similar, and the weight lower interest tag seems not recommended; if recommended, and follow all label may be very large amount of calculation.
This time can be screened weighted sampling mode user interest tags, each recall not all interested in the label, but in accordance with the sampling weights recall part of the label, the benefits of doing so are obvious:
1) reduce the complexity of the recall time
2) can retain more tags
3) calculate the recall have been changed every time, meet the recommended diversity
4) Notwithstanding the changes, but it is still subject to major label small relative weight constraints
Top Charts of the show, the weighted sampling method can also be used, such as the popular show only 10 bits, it can be a time-weighted sampling 10 shows, there are subtle changes to refresh.
Therefore, the weighted sampling method should belong to the skill of the recommendation system, users can improve subtle recommendation result, it is still very important.