Collaborative filtering is one of the widely adopted approaches in recommender systems used for e-commerce applications, stating that users having similar tastes will have similar preferences in the future. The literature presents a number of similarity metrics such as the extended Jaccard coefficient to quantify these preference similarities. This paper aims to improve prediction accuracy by optimizing the similarity values computed using these metrics by adopting two biologically inspired approaches, namely artificial bee colony and genetic algorithms, with a bottom-up approach, suggesting that any improvement on a single-user basis will reflect on the overall prediction accuracy. Detailed statistical analysis was performed using the t-test, analysis of variance, and McNemar's test to see whether there were performance differences. The results show that statistically significant differences exist with high confidence levels.