A.3 Creating additional attributes

As previously discussed, a sometimes effective way to decrease error rates is to develop attributes from the raw data that can be better discriminants than the raw data by itself. This is especially true of learners like C4.5, since if we select our attributes carefully so that the concept boundaries are parallel to the attributes we select, they will be able to segment the space more accurately. In this case, several attributes which were thought to be useful were added. These were:

• The difference between the bends of the MCPs. It was felt that in signing, the bend of a finger relative to the previous one might be significant, and provide more information than simply knowing the absolute bend of each finger.

• A measure of the closeness of the thumb to the little finger. This also was believed to be a useful piece of information, since some handshapes require the thumb and little finger to touch. The equation derived empirically and approximately for this was little-finger-rotation*(little-finger-mcp+little-finger-pip)+4*thumb-palm-rotation*(thumb-mcp+thumb-pip).

• Total finger bend. By summing the MCP and PIP we can get a measure of how far the finger is bent overall. It may be that this proves to be more invariant, particularly between signers than the raw values.

• The differences between total finger bend.

All up there were an additional 19 attributes suggested. They were tested together and their effect on the accuracy measured.

  
Table A.2: Results of learning algorithms on Vamplew's handshape data with additional attributes.

• A.3.1 Discussion