In radio interferometry, observed visibilities are intrinsically sampled at some interval in time and frequency. Modern interferometers are capable of producing data at very high time and frequency resolution; practical limits on storage and computation costs require that some form of data compression be imposed. The traditional form of compression is a simple averaging of the visibilities over coarser time and frequency bins. This has an undesired side effect: the resulting averaged visibilities "decorrelate", and do so differently depending on the baseline length and averaging interval. This translates into a non-trivial signature in the image domain known as "smearing", which manifests itself as an attenuation in amplitude towards off-centre sources. With the increasing fields of view and/or longer baselines employed in modern and future instruments, the trade-off between data rate and smearing becomes increasingly unfavourable. In this work we investigate alternative approaches to low-loss data compression. We show that averaging of the visibility data can be treated as a form of convolution by a boxcar-like window function, and that by employing alternative baseline-dependent window functions a more optimal interferometer smearing response may be induced. In particular, we show improved amplitude response over a chosen field of interest, and better attenuation of sources outside the field of interest. The main cost of this technique is a reduction in nominal sensitivity; we investigate the smearing vs. sensitivity trade-off, and show that in certain regimes a favourable compromise can be achieved. We show the application of this technique to simulated data from the Karl G. Jansky Very Large Array (VLA) and the European Very-long-baseline interferometry Network (EVN).