Devlog IV: Bioblitz V1.1
Overview:
It may not look like much, but I added a markov chain based page-rank type system that was quite complex to recommend contests.
1. Page-rank recommendation system
-
From
playedGameIds, we derive a preference vector over:- topic distribution
- difficulty preference
- official vs user-created bias
- exploration pressure
-
Construct teleport distribution:
v_j ∝ w_t * topic(j)
+ w_d * difficulty(j)
+ w_q * quality(j)
+ w_f * freshness(j)
+ w_n * novelty(j)Normalize so that sum(v_j) = 1
-
Build transition matrix using top-k neighbors:
P_ij = w_ij / sum(w_ij over TopK neighbors)
-
PageRank iteration:
π_(t+1) = α * (P^T * π_t) + (1 - α) * v
-
Iterate ~24 steps until convergence, then rank by π
2. Rating graphs fixed
- Rating graphs did not handle multiple contests taken on the same day, but now they are treated as distinct nodes
- The color of the graph now matches the user’s rating tier, using hardcoded hex values
Log in to leave a comment