dérive

A recommender that takes you on a walk through rare and strange concepts. Every pick is a choice between going deeper into what you like and exploring somewhere new. Many come from a list I have kept while reading for years.

Cognitive science

Apophenia

Perceiving meaningful connections between unrelated things. It also covers gamblers who find patterns in random results.

Wikipedia (opens in a new tab)

How it works

Each concept is a point in a space of meaning, computed from its term and description. You choose the next one among three:

Closest: the unread concept most similar to the one you are reading.
Other field: the unread concept in another field most similar to your taste.
Far: one of the concepts least similar to those you read recently.

Your taste follows Rocchio relevance feedback: what you picked and saved, minus half of what you disliked.

The math

Each concept $c$ is an embedding $v_c$ of its term and description. $k$ is the concept you are reading, $f(c)$ the field of $c$ and $U$ the concepts you have not read. Among your last 50 concepts, $P$ holds those you picked or saved and $D$ those you disliked. Your taste is the Rocchio vector

t = \frac{1}{|P|} \sum_{c \in P} v_c - \frac{1}{2} \cdot \frac{1}{|D|} \sum_{c \in D} v_c

or $v_k$ before any pick, save or dislike. With $F$ the fields of $k$ and of Closest,

\text{Closest} = \operatorname{arg\,max}_{c \in U} \cos(v_c, v_k)

\text{Other field} = \operatorname{arg\,max}_{c \in U,\, f(c) \notin F} \cos(v_c, t)

Far is drawn at random from the 30 concepts of $U$ with the lowest $\max_{r \in R} \cos(v_c, v_r)$ , where $R$ is your last 10 concepts.