Is Amazon A9 dead? A deep analysis of the COSMO-LM model

Recently, many Amazon sellers said that the A9 algorithm is dead, and asked me to take a look at Amazon’s latest AI algorithm.
I read the COSMO-LM model research report by the Amazon team.

I have brought you a Chinese COSMO-LM model research report for better understanding. If you are interested in this, you can contact me privately to get the translated version.

Image source: COSMO: A Large-Scale E-commerce Common Sense Knowledge Generation and Serving System at Amazon

Let me first state the conclusion:

1. A9 is not dead, or the core is still there;
   

2. Copywriting may need to be adjusted to cater to the new algorithm.

Simply put, both the traditional A9 and COSMO-LM models focus on correlation.

Looking further, they work on solving two problems: first, search relevance, and second, purchase relevance.

The relevance of the search is determined by the text content of the title, list items, description, etc.

The relevance of purchases is affected by data indicators such as click-through rate, conversion rate, and sales volume.

Traditional A9 focuses more on the search term itself and recommends related products based on the comprehensive score. In short, what you search for, I recommend. And the products that are bought more will be recommended more frequently in the front.

In contrast, the COSMO-LM model uses user portraits or big data search transaction habits to deduce precise search terms behind potentially inaccurate search terms. For example, if you want to buy a warm coat, users in the north and south will definitely have different requirements for thickness and fabric. The COSMO-LM model may deduce a coat style that better suits your needs based on your historical purchase records or location.

Of course, people who have been using large language models for a long time have had similar experiences: if the inductive model does not have enough knowledge input, it will sometimes give outrageous results. This also explains why some unknown traffic entered the advertising before.

So what should we do when it comes to tweaking copy?

Although the report only gives a general introduction to the instruction fine-tuning process of the COSMO-LM model without detailed parameters, we can give some optimization understanding based on the known logical parameters. The core idea is: we give AI what it wants to see.

As for data indicators, everyone can continue to promote according to the previous promotion plan, there is not much need for changes.

In terms of copywriting, it is mainly necessary to supplement it based on the relationship types derived by AI to make it more specific and more in line with user needs.