Pointwise mutual information (PMI) measures whether two events co-occur more often than chance predicts. In query analysis, higher scores indicate stronger semantic relatedness independent of individual popularity. section 6.3.1
A PMI2 variant compares user-level co-occurrence with each keyword’s occurrence count. Build keyword–user pairs, count distinct users for each keyword and pair, calculate PMI2, then filter sparse pairs with a minimum co-occurrence threshold.
Related-query ranking
A query-log pipeline finds pairs such as “iphone 4s” and “iphone,” or “bose” and “beats.” Combining co-occurrence rank with PMI2 rank balances popularity against association strength, while thresholds reduce noise.
