Extended Data Fig. 4: Declining combinatorial novelty.

This figure shows changing patterns in the combinatorial novelty/conventionality of papers (a, n = 24,659,076) and patents (b, n = 3,912,353), using a previously proposed measure of “atypical combinations”¹⁴. The measure quantifies the degree to which the prior work cited by a paper or patent would be expected by chance. For papers, we follow prior work¹⁴ and consider combinations of cited journals. If a paper made three citations to prior work, and that work was published in three different journals—Nature, Cell, and Science—then there are three combinations—Nature × Cell, Nature × Science, and Science × Cell. To determine the degree to which each combination would be expected by chance, the frequency of observed pairings is compared to those in 10 “rewired” copies of the overall citation network, using a z-score. For patents, there is no natural analogue to journals, and therefore we consider pairings of primary United States Patent Classification (USPC) system codes. We present the results of this analysis following the approach of prior work¹⁴, which plots the cumulative distribution function of the measure. In general, there is a rightward shift in the cumulative distributions over time, suggesting that for both papers and patents, combinations are more conventional than would be expected by chance, consistent with what we would anticipate based on our results using the CD index. For patents, there is also a smaller shift in the opposite direction on the left side of the distribution, suggesting that novel patents in recent decades are somewhat more novel than novel patents in earlier decades. Overall, however, the bulk of the distribution is moving rightward, indicating greater conventionality.