We propose a modification of the Pattern Informatics (PI) method that has been developed for forecasting the locations of future large earthquakes. This forecast is based on analyzing the space-time patterns of past earthquakes to find possible locations where future large earthquakes are expected to occur. A characteristic of our modification is that the effect of errors in the locations of past earthquakes on the output forecast is reduced. We apply the modified and original methods to seismicity in the central part of Japan and compared the forecast performances. We also invoke the Relative Intensity (RI) of seismic activity and randomized catalogs to constitute null hypotheses. We do statistical tests using the Molchan and Relative Operating Characteristic (ROC) diagrams and the log-likelihoods and show that the forecast for using the modified PI method is generally better than the competing original-PI forecast and the forecasts from the null hypotheses. Using the bootstrap technique with Monte-Carlo simulations, we further confirm that earthquake sequences simulated based on the modified-PI forecast can be statistically the same as the real earthquake sequence so that the forecast is acceptable. The main and innovative science in this paper is the modification of the PI method and the demonstration of its applicability, showing a considerable promise as an intermediate-term earthquake forecasting tool.