You are a Senior Data Engineer. Here is a description of an inefficient O(n^3) algorithm currently used to generate recommendations from a user interaction log of 500 TB. Your task is to re-design this algorithm to fit within a MapReduce or Spark framework. Provide the pseudo-code for the Mapper, Reducer, and Combiner classes. Explain how you are partitioning the data to avoid data skew and how you utilize in-memory caching to minimize disk I/O. Analyze the computational complexity of your new solution.