Large-scale gene co-expression networks are an effective methodology to analyze sets of co-expressed genes and discover new gene functions or associations. Distances between genes are estimated according to their expression profiles and are visualized in networks that may be further partitioned to reveal communities of co-expressed genes. Creating expression profiles is now eased by the large amounts of publicly available expression data (microarrays and RNA-seq). Although many distance calculation methods have been intensively compared and reviewed in the past, it is unclear how to proceed when many samples reflecting a wide range of different conditions are available. Should as many samples as possible be integrated into network construction or be partitioned into smaller sets of more related samples? Previous studies have indicated a saturation in network performances to capture known associations once a certain number of samples is included in distance calculations. Here, we examined the influence of sample size on co-expression network construction using microarray and RNA-seq expression data from three plant species. We tested different down-sampling methods and compared network performances in recovering known gene associations to networks obtained from full datasets. We further examined how aggregating networks may help increase this performance by testing six aggregation methods.