PMI2 is a k=2 variation of pointwise mutual information for scoring semantic association between query terms. It compares how often two keywords are searched by the same users with how often each appears individually; higher scores indicate stronger relatedness and reduce sensitivity to frequency differences. section 6.3.1
Compute PMI2 as log(users_cooc² / (users_occ1 × users_occ2)). Build keyword–user pairs, count distinct users for each keyword and pair, then filter weak pairs such as co-occurrences at five or fewer users. PMI2 can still amplify sparse noise, so combining it with co-occurrence ranking provides a practical compromise. section 6.3.1
Query-log associations
Pairs such as “kenwood” and “alpine” or “green lantern” and “captain america” receive useful association scores, while “wireless mouse” and “godfather” illustrates noise from sparse co-occurrence.
