The MLP neural network and the random forest model did not perform as well as the naive model, but as seen by Figures \ref{576363} and \ref{508370}, did demonstrate what appears to be more geographic sensitivity. These models  picked up variations in Staten Island, Lower Manhattan, and Rockaway Peninsula, for example, that did not emerge in the linear regression approach, which is apparent at the census tract level. The tendency of these more complex models to over-fit might inhibit their predictive power but do indicate higher sensitivity, a characteristic that further analysis centered on interpolating possible correlations rather than prediction might exploit.
Additionally, several characteristics were found which were correlated with the likelihood of a gas leak at both the census tract and zip code levels.  At the zip code level, these indicators were complaints being referred to the New York City Housing Authority (NYCHA), the percentage of the population that is Black or African American in that zip code, the number of vacant, open or unguarded buildings, alterations of an occupied building without valid permits and the household mean income.   At the census tract level, these indicators were  the amount of open space in the census tract, permits for construction equipment, failure to paint sprinkler piping, failure to notify DOB prior to the cancellation of earthwork, and the percentage of the population that is Black or African American in that census tract.

Notes on performance

While total error was orders of magnitude higher for census tract than zip code, these were much more granular geographical areas and that added precision may have value if applied to prioritizing inspections. Initial analysis revealed that when the tract-level model predictions were aggregated back to the zip code level and compared, they fared better and indicate an area of further research -- particularly to address whether increasing the granularity of observations in this matter can counter-act the over-fitting that occurs with a small number of observations (such as 195 zips) and large dimensionality (735 features). Additionally, this evaluation metric of total RMSE penalizes extreme values, which influenced models that accurately predicted zips and tracts with little or no gas leaks (a large majority) while drastically under-estimating the top zips or tracts. If the primary goal is to prioritize top tracts, it may be better pursue more precise objectives, like optimizing models based only on specific zips or tracts with a high likelihood of incidents.
Figure 5 demonstrates the tendency of the Multi-Layer Perceptron neural network (in green below) to over-fit on the historical data (which is also what comprises the naive prediction, in red below) and how both approaches under-estimate the top zip codes, although the MLP neural network to a lesser extent. More data and better prediction objectives may reveal promising opportunities for neural networks.