Periodically I get emails from Amazon.com recommending things I might want to purchase, based on my previous purchase history.
The recommendations are always logical on a very low level, but completely nonsensical at a higher semantic level. The basic problem is that Amazon has no clue why I purchased something, which means they have no context for assessing what that purchase might predict about my future buying habits.
For example, today I received several recommendations, including one for a 75mm diameter curved watch glass. As it happens, I purchased a 50mm diameter watch glass through Amazon some months ago, because I needed a curved piece of glass for a scientific experiment. I didn’t actually need it for a watch.
It occurs to me now that even if I had needed it for a watch, it would have been for a specific 50mm diameter watch, not a hypothetical 75mm watch that I do not own.
Now my curved glass experiment is long finished, yet I continue to receive watch glass suggestions from Amazon, like pleas from a child who asks for a pony one more time, hoping to wear down a distracted grownup through sheer repetition.
I am left wondering — does this whole email recommendation thing ever actually work?
I heard two stories regarding the recommendation system.
1. case Amazon japan
There was a news while ago about Japanese amazon recommendation. They banned a specific combination of recommendation.
A while ago, some people posted about how to suicide, and what tools and stuffs you need to do that. Some people bought them through the amazon and actually dead. Then the originated site was banned, but amazon recommendation remembered the combination. So some people has a partial knowledge of this method, however, it is very odd combination for usual daily life and the recommendation gives you the information. That was a discussion, but amazon voluntarily banned the combination. This case is not for future purchases, but togather one.
2. case a local drag store
I think this from CASM (IEEE Spectrum) a while ago. A man came to a local drag store and complained to the store that sent a maternity goods to them. But the to his 16 years old daughter instead of his wife. The drag store’s recommendation system using Facebook analysis of the customers, it seems some messed up recommendation system. The store owner apologized him and the data was corrected. The system took the Facebook data and many of the purchases pattern connected to the account. Two months later, the man found his daughter is pregnant, she was following quite common pattern of the pregnant woman’s purchases pattern. This case seems the system is working.
Extra: case mine.
Mines doesn’t work. The problem is I buy birthday presents for others through the amazon time to time. I got recommendation which is good for my friend’s wife or one of my friends for sure, but I am totally not interested in it. This is the same that Amazon lacks why I bought it…