1 | /*************************************************************************/ |
---|
2 | /* */ |
---|
3 | /* Soften thresholds for continuous attributes */ |
---|
4 | /* ------------------------------------------- */ |
---|
5 | /* */ |
---|
6 | /*************************************************************************/ |
---|
7 | |
---|
8 | |
---|
9 | #include "defns.i" |
---|
10 | #include "types.i" |
---|
11 | #include "extern.i" |
---|
12 | |
---|
13 | |
---|
14 | Boolean *LHSErr, /* Does a misclassification occur with this value of an att */ |
---|
15 | *RHSErr; /* if the below or above threshold branches are taken */ |
---|
16 | |
---|
17 | ItemNo *ThreshErrs; /* ThreshErrs[i] is the no. of misclassifications if thresh is i */ |
---|
18 | |
---|
19 | float *CVals; /* All values of a continuous attribute */ |
---|
20 | |
---|
21 | |
---|
22 | #define Below(v,t) (v <= t + 1E-6) |
---|
23 | |
---|
24 | |
---|
25 | /*************************************************************************/ |
---|
26 | /* */ |
---|
27 | /* Soften all thresholds for continuous attributes in tree T */ |
---|
28 | /* */ |
---|
29 | /*************************************************************************/ |
---|
30 | |
---|
31 | |
---|
32 | SoftenThresh(T) |
---|
33 | /* ------------ */ |
---|
34 | Tree T; |
---|
35 | { |
---|
36 | CVals = (float *) calloc(MaxItem+1, sizeof(float)); |
---|
37 | LHSErr = (Boolean *) calloc(MaxItem+1, sizeof(Boolean)); |
---|
38 | RHSErr = (Boolean *) calloc(MaxItem+1, sizeof(Boolean)); |
---|
39 | ThreshErrs = (ItemNo *) calloc(MaxItem+1, sizeof(ItemNo)); |
---|
40 | |
---|
41 | InitialiseWeights(); |
---|
42 | |
---|
43 | ScanTree(T, 0, MaxItem); |
---|
44 | |
---|
45 | cfree(ThreshErrs); |
---|
46 | cfree(RHSErr); |
---|
47 | cfree(LHSErr); |
---|
48 | cfree(CVals); |
---|
49 | } |
---|
50 | |
---|
51 | |
---|
52 | |
---|
53 | /*************************************************************************/ |
---|
54 | /* */ |
---|
55 | /* Calculate upper and lower bounds for each test on a continuous */ |
---|
56 | /* attribute in tree T, using data items from Fp to Lp */ |
---|
57 | /* */ |
---|
58 | /*************************************************************************/ |
---|
59 | |
---|
60 | |
---|
61 | ScanTree(T, Fp, Lp) |
---|
62 | /* -------- */ |
---|
63 | Tree T; |
---|
64 | ItemNo Fp, Lp; |
---|
65 | { |
---|
66 | short v; |
---|
67 | float Val, Se, Limit, Lower, Upper, GreatestValueBelow(); |
---|
68 | ItemNo i, Kp, Ep, LastI, Errors, BaseErrors; |
---|
69 | ClassNo CaseClass, Class1, Class2, Category(); |
---|
70 | Boolean LeftThresh=false; |
---|
71 | Description CaseDesc; |
---|
72 | Attribute Att; |
---|
73 | void Swap(); |
---|
74 | |
---|
75 | /* Stop when get to a leaf */ |
---|
76 | |
---|
77 | if ( ! T->NodeType ) return; |
---|
78 | |
---|
79 | /* Group the unknowns together */ |
---|
80 | |
---|
81 | Kp = Group(0, Fp, Lp, T); |
---|
82 | |
---|
83 | /* Soften a threshold for a continuous attribute */ |
---|
84 | |
---|
85 | Att = T->Tested; |
---|
86 | |
---|
87 | if ( T->NodeType == ThreshContin ) |
---|
88 | { |
---|
89 | printf("\nTest %s <> %g\n", AttName[Att], T->Cut); |
---|
90 | |
---|
91 | Quicksort(Kp+1, Lp, Att, Swap); |
---|
92 | |
---|
93 | ForEach(i, Kp+1, Lp) |
---|
94 | { |
---|
95 | /* See how this item would be classified if its |
---|
96 | value were on each side of the threshold */ |
---|
97 | |
---|
98 | CaseDesc = Item[i]; |
---|
99 | CaseClass = Class(CaseDesc); |
---|
100 | Val = CVal(CaseDesc, Att); |
---|
101 | |
---|
102 | Class1 = Category(CaseDesc, T->Branch[1]); |
---|
103 | Class2 = Category(CaseDesc, T->Branch[2]); |
---|
104 | |
---|
105 | CVals[i] = Val; |
---|
106 | LHSErr[i] = (Class1 != CaseClass ? 1 : 0); |
---|
107 | RHSErr[i] = (Class2 != CaseClass ? 1 : 0); |
---|
108 | } |
---|
109 | |
---|
110 | /* Set Errors to total errors if take above thresh branch, |
---|
111 | and BaseErrors to errors if threshold has original value */ |
---|
112 | |
---|
113 | Errors = BaseErrors = 0; |
---|
114 | ForEach(i, Kp+1, Lp) |
---|
115 | { |
---|
116 | Errors += RHSErr[i]; |
---|
117 | |
---|
118 | if ( Below(CVals[i], T->Cut) ) |
---|
119 | { |
---|
120 | BaseErrors += LHSErr[i]; |
---|
121 | } |
---|
122 | else |
---|
123 | { |
---|
124 | BaseErrors += RHSErr[i]; |
---|
125 | } |
---|
126 | } |
---|
127 | |
---|
128 | /* Calculate standard deviation of the number of errors */ |
---|
129 | |
---|
130 | Se = sqrt( (BaseErrors+0.5) * (Lp-Kp-BaseErrors+0.5) / (Lp-Kp+1) ); |
---|
131 | Limit = BaseErrors + Se; |
---|
132 | |
---|
133 | Verbosity(1) |
---|
134 | { |
---|
135 | printf("\t\t\tBase errors %d, items %d, se=%.1f\n", |
---|
136 | BaseErrors, Lp-Kp, Se); |
---|
137 | printf("\n\tVal <= Errors\t\t+Errors\n"); |
---|
138 | printf("\t %6d\n", Errors); |
---|
139 | } |
---|
140 | |
---|
141 | /* Set ThreshErrs[i] to the no. of errors if the threshold were i */ |
---|
142 | |
---|
143 | ForEach(i, Kp+1, Lp) |
---|
144 | { |
---|
145 | ThreshErrs[i] = Errors = Errors + LHSErr[i] - RHSErr[i]; |
---|
146 | |
---|
147 | if ( i == Lp || CVals[i] != CVals[i+1] ) |
---|
148 | { |
---|
149 | Verbosity(1) |
---|
150 | printf("\t%6g %6d\t\t%7d\n", |
---|
151 | CVals[i], Errors, Errors - BaseErrors); |
---|
152 | } |
---|
153 | } |
---|
154 | |
---|
155 | /* Choose Lower and Upper so that if threshold were set to |
---|
156 | either, the number of items misclassified would be one |
---|
157 | standard deviation above BaseErrors */ |
---|
158 | |
---|
159 | LastI = Kp+1; |
---|
160 | Lower = Min(T->Cut, CVals[LastI]); |
---|
161 | Upper = Max(T->Cut, CVals[Lp]); |
---|
162 | while ( CVals[LastI+1] == CVals[LastI] ) LastI++; |
---|
163 | |
---|
164 | while ( LastI < Lp ) |
---|
165 | { |
---|
166 | i = LastI + 1; |
---|
167 | while ( i < Lp && CVals[i+1] == CVals[i] ) i++; |
---|
168 | |
---|
169 | if ( ! LeftThresh && |
---|
170 | ThreshErrs[LastI] > Limit && |
---|
171 | ThreshErrs[i] <= Limit && |
---|
172 | Below(CVals[i], T->Cut) ) |
---|
173 | { |
---|
174 | Lower = CVals[i] - |
---|
175 | (CVals[i] - CVals[LastI]) * (Limit - ThreshErrs[i]) / |
---|
176 | (ThreshErrs[LastI] - ThreshErrs[i]); |
---|
177 | LeftThresh = true; |
---|
178 | } |
---|
179 | else |
---|
180 | if ( ThreshErrs[LastI] <= Limit && |
---|
181 | ThreshErrs[i] > Limit && |
---|
182 | ! Below(CVals[i], T->Cut) ) |
---|
183 | { |
---|
184 | Upper = CVals[LastI] + |
---|
185 | (CVals[i] - CVals[LastI]) * (Limit - ThreshErrs[LastI]) / |
---|
186 | (ThreshErrs[i] - ThreshErrs[LastI]); |
---|
187 | if ( Upper < T->Cut ) Upper = T->Cut; |
---|
188 | } |
---|
189 | |
---|
190 | LastI = i; |
---|
191 | } |
---|
192 | |
---|
193 | T->Lower = Lower; |
---|
194 | T->Upper = Upper; |
---|
195 | |
---|
196 | Verbosity(1) printf("\n"); |
---|
197 | |
---|
198 | printf("\tLower = %g, Upper = %g\n", T->Lower, T->Upper); |
---|
199 | } |
---|
200 | |
---|
201 | /* Recursively scan each branch */ |
---|
202 | |
---|
203 | ForEach(v, 1, T->Forks) |
---|
204 | { |
---|
205 | Ep = Group(v, Kp+1, Lp, T); |
---|
206 | |
---|
207 | if ( Kp < Ep ) |
---|
208 | { |
---|
209 | ScanTree(T->Branch[v], Kp+1, Ep); |
---|
210 | Kp = Ep; |
---|
211 | } |
---|
212 | } |
---|
213 | } |
---|