Many sellers will ask if Amazon has any rules. In fact, they can change their minds and understand the algorithm recommended by Amazon, and easily master the rules of Amazon. Some recommendation algorithms seen now are generally based on the comprehensive filtering recommendation of the similarity of the items themselves, user browsing habits, likes, purchase records and other data. The following will share with sellers the rules of Amazon’s recommended products.
1. Collaborative filtering based on object dimensions.
For example, users 1, 2, 3, 4, 5, 6 buy item A, and users 1, 3, 4, 5, 7 buy item B, then 123456 is used as the characteristic attribute data of item A, and 13457 is used as the characteristic data of item B, and then the similarity between items A and B is calculated.
Because the same person bought A and B, then A and B must have a certain relationship.
2. Collaborative filtering based on user dimensions.
Collect user purchase, browsing, collection and other product data, and list the products purchased by users as user attribute dimensions. For example, if user A buys item 1.2.3.4.5 and user B buys item 1.2.5.6, then you can simply use 12345 and 1256 as the attribute feature strings of A and B, calculate the similarity between A and B, and cluster the users into several categories (neighbors) through simple clustering.
3. Based on the similarity of the items themselves.
For example, clothing A and clothing B, used to calculate their similarity in style, price segment, classification, attributes, brand positioning, etc. If the similarity is high, users can recommend B when browsing A. Of course, it is not that simple.
These attributes of clothes do not depend on users. It does not depend on the user’s behavioral data, so it is relatively rigid and there is no personalized recommendation.
Many people know the idea of this algorithm, but the simpler the algorithm, the more difficult it is to achieve good results, especially the algorithm with a very low conversion rate.
There are dozens of attributes for commodities. For commodities of different categories, not all attributes need to be included in the similarity calculation. The attributes are included but the importance is different. Therefore, screening the necessary attributes of different categories and setting the weight values in the similarity calculation is a very large project. Amazon’s recommendation system is also the earliest industry in the world. I believe they must have their own fast and effective methods on this issue.
4. Apply strong association rules before application.
The focus is on the same purchase record (of course not necessarily, it depends on your choice).
First, the data collection needs to filter out orders for the purchase of one commodity. Then perform one-to-one extraction statistics for each record. The simple one is the two statistical times, which are the number of times two commodities are purchased at the same time, which is suitable for one-to-one recommendation.
The other is the FPTreee algorithm, which is not only one-to-one recommendation, but also one-to-two and two-to-one.
In this process, the association rule mining algorithm is very important, and the confidence and support also need to be constantly adjusted.
5. Algorithms learn from each other.
Data sharing between all recommendation systems. Regular automatic data update. Automatic learning
Generally speaking, most recommendation algorithms are simple, but they need to be used well, which cannot be achieved without long-term accumulation.
It is basically impossible to achieve good recommendation results by just hiring some algorithm engineers and using some algorithm frameworks. Only when the algorithm is combined with the business can a chemical reaction occur.
The above is the rule of Amazon product recommendation shared with sellers, and I hope it will be helpful for sellers to open stores. Lianlian Cross-border Payment always pays attention to every little thing of the sellers, and will also bring articles on related aspects in future articles to help sellers operate better.