Tom Slee has a fun dishpan model (term) of the effect of recommendation systems on product diversity. The short form is that introducing a recommendation system drives the market toward reduced diversity. He launches off from a paper that shows much the same result. He builds a little simulation.
His model is straight forward. Buyers and products are situated in a feature space. Buyers are most satisfied when they acquire a product close to them in this space. His model runs in two modes: with and without a recommendation system. The products and buyers are drawn randomly from a normal distribution. I guess we could label the buyers on the tails of that distribution eccentric, and the products on the tail as niche.
Absent the recommendation system buyers find products with some tendency to find ones nearby in the feature space. Those products satisfy them to some extent and that knowledge can then be rolled up into a recommendation system. Naturally, once the recommendation system is turned buyer choice is effected, to some degree, by this pooled knowledge.
There are lots of affordances to play with on a model like this. For example if we assume that the buyer’s position in the feature space is known to the recommendation system it can vary it’s advice. Alternatively we could imagine changing buyer’s minds by moving their preferences in the feature space or changing how they weigh one or anther dimension of the space.
It is conventional in discussions of markets to rant about the evils of the intermediary, i.e. the agent. In Tom’s model the agent is the recommendation system. Tom’s conclusion is that the recommendation system leads to a sharp decrease in the diversity of products cleared through the market. (I am reminded of how the real estate agent has a strong preference for bathrooms in white.) I suspect that if he did the sums it would also show a decrease in the overall buyer satisfaction – but that’s less clear.
Dishpan models, particularly coupled with a simulation, are extermely powerful tools for thinking about problems like this. Valuing their outputs can be trouble, but ignoring them is worse.