1 | /*************************************************************************/ |
---|
2 | /* */ |
---|
3 | /* Evaluation of the subsetting of a discrete attribute */ |
---|
4 | /* ---------------------------------------------------- */ |
---|
5 | /* */ |
---|
6 | /*************************************************************************/ |
---|
7 | |
---|
8 | #include "buildex.i" |
---|
9 | |
---|
10 | ItemCount *Slice1, /* Slice1[c] = saved values of Freq[x][c] in subset.c */ |
---|
11 | *Slice2; /* Slice2[c] = saved values of Freq[y][c] */ |
---|
12 | |
---|
13 | Set **Subset; /* Subset[a][s] = subset s for att a */ |
---|
14 | |
---|
15 | short *Subsets; /* Subsets[a] = no. subsets for att a */ |
---|
16 | |
---|
17 | /*************************************************************************/ |
---|
18 | /* */ |
---|
19 | /* Evaluate subsetting a discrete attribute and form the chosen */ |
---|
20 | /* subsets Subset[Att][], setting Subsets[Att] to the number of */ |
---|
21 | /* subsets, and the Info[] and Gain[] of a test on the attribute */ |
---|
22 | /* */ |
---|
23 | /*************************************************************************/ |
---|
24 | |
---|
25 | EvalSubset(Att, Fp, Lp, Items) |
---|
26 | /* ---------- */ |
---|
27 | Attribute Att;ItemNo Fp, Lp;ItemCount Items; { |
---|
28 | DiscrValue V1, V2, BestV1, BestV2, Barred; |
---|
29 | ItemCount KnownItems; |
---|
30 | ClassNo c; |
---|
31 | float BaseInfo, MinGain, ThisGain, ThisInfo, Val, BestVal, BestGain, |
---|
32 | BestInfo, PrevVal, PrevGain, PrevInfo, DiscrKnownBaseInfo(), |
---|
33 | Worth(), ComputeGain(), TotalInfo(); |
---|
34 | short Blocks = 0, MissingValues = 0, ReasonableSubsets, Bytes, b; |
---|
35 | Boolean MergedSubsets = false; |
---|
36 | int SaveMINOBJS; |
---|
37 | |
---|
38 | SaveMINOBJS = MINOBJS; |
---|
39 | MINOBJS = 1; |
---|
40 | |
---|
41 | /* First compute Freq[][], ValFreq[], base info, and the gain |
---|
42 | and total info of a split on discrete attribute Att */ |
---|
43 | |
---|
44 | ComputeFrequencies(Att, Fp, Lp); |
---|
45 | |
---|
46 | KnownItems = Items - ValFreq[0]; |
---|
47 | if (KnownItems < Epsilon) { |
---|
48 | Verbosity(2) |
---|
49 | printf("\tAtt %s: no known values\n", AttName[Att]); |
---|
50 | |
---|
51 | Gain[Att] = -Epsilon; |
---|
52 | Info[Att] = 0; |
---|
53 | return; |
---|
54 | } |
---|
55 | |
---|
56 | BaseInfo = DiscrKnownBaseInfo(KnownItems, MaxAttVal[Att]); |
---|
57 | |
---|
58 | PrevGain = ComputeGain(BaseInfo, UnknownRate[Att], MaxAttVal[Att], |
---|
59 | KnownItems); |
---|
60 | PrevInfo = TotalInfo(ValFreq, 0, MaxAttVal[Att]) / Items; |
---|
61 | PrevVal = Worth(PrevInfo, PrevGain, Epsilon); |
---|
62 | |
---|
63 | Verbosity(2) { |
---|
64 | printf("\tAtt %s", AttName[Att]); |
---|
65 | |
---|
66 | Verbosity(3) |
---|
67 | PrintDistribution(Att, MaxAttVal[Att], true); |
---|
68 | |
---|
69 | printf("\tinf %.3f, gain %.3f, val=%.3f\n", PrevInfo, PrevGain, PrevVal); |
---|
70 | } |
---|
71 | |
---|
72 | /* Eliminate unrepresented attribute values from Freq[] and ValFreq[] |
---|
73 | and form a separate subset for each represented attribute value */ |
---|
74 | |
---|
75 | Bytes = (MaxAttVal[Att] >> 3) + 1; |
---|
76 | ClearBits(Bytes, Subset[Att][0]); |
---|
77 | |
---|
78 | ForEach(V1, 1, MaxAttVal[Att]) { |
---|
79 | if (ValFreq[V1] > 0.5) { |
---|
80 | if (++Blocks < V1) { |
---|
81 | ValFreq[Blocks] = ValFreq[V1]; |
---|
82 | ForEach(c, 0, MaxClass) { |
---|
83 | Freq[Blocks][c] = Freq[V1][c]; |
---|
84 | } |
---|
85 | } |
---|
86 | ClearBits(Bytes, Subset[Att][Blocks]); |
---|
87 | SetBit(V1, Subset[Att][Blocks]); |
---|
88 | } else { |
---|
89 | SetBit(V1, Subset[Att][0]); |
---|
90 | MissingValues++; |
---|
91 | } |
---|
92 | } |
---|
93 | |
---|
94 | /* Merge any single-class subsets with others of the same class */ |
---|
95 | /* Note: have ValFreq[V] > 0 for all V */ |
---|
96 | |
---|
97 | ForEach(V1, 1, Blocks-1) { |
---|
98 | for (c = 0; Freq[V1][c] < 0.1; c++) |
---|
99 | ; |
---|
100 | |
---|
101 | if (Freq[V1][c] < ValFreq[V1] - 0.1) |
---|
102 | continue; |
---|
103 | |
---|
104 | /* Now have a single class -- look for others */ |
---|
105 | |
---|
106 | for (V2 = V1 + 1; V2 <= Blocks;) { |
---|
107 | if (Freq[V2][c] < ValFreq[V2] - 0.1) { |
---|
108 | V2++; |
---|
109 | } else { |
---|
110 | /* Merge these subsets */ |
---|
111 | |
---|
112 | Combine(V1, V2, Blocks); |
---|
113 | |
---|
114 | ForEach(b, 0, Bytes-1) { |
---|
115 | Subset[Att][V1][b] |= Subset[Att][V2][b]; |
---|
116 | Subset[Att][V2][b] = Subset[Att][Blocks][b]; |
---|
117 | } |
---|
118 | |
---|
119 | Blocks--; |
---|
120 | MergedSubsets = true; |
---|
121 | } |
---|
122 | } |
---|
123 | } |
---|
124 | |
---|
125 | if (MergedSubsets) { |
---|
126 | PrevGain = ComputeGain(BaseInfo, UnknownRate[Att], Blocks, KnownItems); |
---|
127 | PrevInfo = TotalInfo(ValFreq, 0, Blocks) / Items; |
---|
128 | PrevVal = Worth(PrevInfo, PrevGain, Epsilon); |
---|
129 | |
---|
130 | Verbosity(2) { |
---|
131 | printf("\tAfter merging single-class subsets:"); |
---|
132 | |
---|
133 | Verbosity(3) |
---|
134 | PrintDistribution(Att, Blocks, false); |
---|
135 | |
---|
136 | printf("\tinf %.3f, gain %.3f, val=%.3f\n", PrevInfo, PrevGain, |
---|
137 | PrevVal); |
---|
138 | } |
---|
139 | } |
---|
140 | |
---|
141 | /* Examine possible pair mergers and hill-climb */ |
---|
142 | |
---|
143 | MinGain = PrevGain / 2; |
---|
144 | |
---|
145 | while (Blocks > 2) { |
---|
146 | BestVal = BestV1 = 0; |
---|
147 | BestGain = -Epsilon; |
---|
148 | |
---|
149 | /* Check reasonable subsets; if less than 3, bar mergers |
---|
150 | involving the largest block */ |
---|
151 | |
---|
152 | ReasonableSubsets = 0; |
---|
153 | Barred = 1; |
---|
154 | |
---|
155 | ForEach(V1, 1, Blocks) { |
---|
156 | if (ValFreq[V1] >= SaveMINOBJS) |
---|
157 | ReasonableSubsets++; |
---|
158 | |
---|
159 | if (ValFreq[V1] > ValFreq[Barred]) |
---|
160 | Barred = V1; |
---|
161 | } |
---|
162 | |
---|
163 | if (ReasonableSubsets >= 3) |
---|
164 | Barred = 0; |
---|
165 | |
---|
166 | /* For each possible pair of values, calculate the gain and |
---|
167 | total info of a split in which they are treated as one. |
---|
168 | Keep track of the pair with the best gain. */ |
---|
169 | |
---|
170 | ForEach(V1, 1, Blocks-1) { |
---|
171 | ForEach(V2, V1+1, Blocks) { |
---|
172 | if (V1 == Barred || V2 == Barred) |
---|
173 | continue; |
---|
174 | |
---|
175 | Combine(V1, V2, Blocks); |
---|
176 | |
---|
177 | ThisGain = ComputeGain(BaseInfo, UnknownRate[Att], Blocks - 1, |
---|
178 | KnownItems); |
---|
179 | ThisInfo = TotalInfo(ValFreq, 0, Blocks - 1) / Items; |
---|
180 | Val = Worth(ThisInfo, ThisGain, Epsilon); |
---|
181 | |
---|
182 | Verbosity(4) { |
---|
183 | printf("\tcombine %d %d info %.3f gain %.3f val %.3f", V1, |
---|
184 | V2, ThisInfo, ThisGain, Val); |
---|
185 | PrintDistribution(Att, Blocks - 1, false); |
---|
186 | } |
---|
187 | |
---|
188 | /* Force a split if |
---|
189 | less than two reasonable subsets, or |
---|
190 | using GAIN criterion |
---|
191 | Prefer this split to the previous one if |
---|
192 | gain >= MinGain (and previous < MinGain), or |
---|
193 | val >= previous best val */ |
---|
194 | |
---|
195 | if (ThisGain >= MinGain && BestGain < MinGain || Val >= BestVal |
---|
196 | || !BestV1 && (!GAINRATIO || ReasonableSubsets < 2)) { |
---|
197 | BestVal = Val; |
---|
198 | BestGain = ThisGain; |
---|
199 | BestInfo = ThisInfo; |
---|
200 | BestV1 = V1; |
---|
201 | BestV2 = V2; |
---|
202 | } |
---|
203 | |
---|
204 | Uncombine(V1, V2); |
---|
205 | } |
---|
206 | } |
---|
207 | |
---|
208 | if (GAINRATIO && ReasonableSubsets >= 2 && (!BestV1 || BestVal |
---|
209 | < PrevVal + 1E-5 || BestVal == PrevVal && BestGain < PrevGain)) |
---|
210 | break; |
---|
211 | |
---|
212 | PrevGain = BestGain; |
---|
213 | PrevInfo = BestInfo; |
---|
214 | PrevVal = BestVal; |
---|
215 | |
---|
216 | Combine(BestV1, BestV2, Blocks); |
---|
217 | |
---|
218 | ForEach(b, 0, Bytes-1) { |
---|
219 | Subset[Att][BestV1][b] |= Subset[Att][BestV2][b]; |
---|
220 | Subset[Att][BestV2][b] = Subset[Att][Blocks][b]; |
---|
221 | } |
---|
222 | |
---|
223 | Blocks--; |
---|
224 | |
---|
225 | Verbosity(2) { |
---|
226 | printf("\t\tform subset "); |
---|
227 | PrintSubset(Att, Subset[Att][BestV1]); |
---|
228 | printf(": %d subsets, inf %.3f, gain %.3f, val %.3f\n", Blocks, |
---|
229 | BestInfo, BestGain, BestVal); |
---|
230 | Verbosity(3) { |
---|
231 | printf("\t\tcombine %d, %d", BestV1, BestV2); |
---|
232 | PrintDistribution(Att, Blocks, false); |
---|
233 | } |
---|
234 | } |
---|
235 | } |
---|
236 | |
---|
237 | MINOBJS = SaveMINOBJS; |
---|
238 | |
---|
239 | if (PrevVal <= 0) { |
---|
240 | Gain[Att] = -Epsilon; |
---|
241 | Info[Att] = 0; |
---|
242 | } else { |
---|
243 | Gain[Att] = ComputeGain(BaseInfo, UnknownRate[Att], Blocks, KnownItems); |
---|
244 | Info[Att] = PrevInfo; |
---|
245 | |
---|
246 | if (MissingValues) { |
---|
247 | Blocks++; |
---|
248 | CopyBits(Bytes, Subset[Att][0], Subset[Att][Blocks]); |
---|
249 | } |
---|
250 | |
---|
251 | Subsets[Att] = Blocks; |
---|
252 | |
---|
253 | Verbosity(2) |
---|
254 | printf("\tFinal subsets:"); |
---|
255 | Verbosity(3) |
---|
256 | PrintDistribution(Att, Blocks, false); |
---|
257 | Verbosity(2) |
---|
258 | printf("\tinf %.3f gain %.3f val %.3f\n", Info[Att], Gain[Att], |
---|
259 | Worth(Info[Att], Gain[Att], Epsilon)); |
---|
260 | } |
---|
261 | } |
---|
262 | |
---|
263 | /*************************************************************************/ |
---|
264 | /* */ |
---|
265 | /* Combine the distribution figures of discrete attribute values */ |
---|
266 | /* x and y, putting the combined figures in Freq[x][] and */ |
---|
267 | /* ValFreq[x][], and saving old values in Slice1 and Slice2 */ |
---|
268 | /* */ |
---|
269 | /*************************************************************************/ |
---|
270 | |
---|
271 | Combine(x, y, Last) |
---|
272 | /* ------- */ |
---|
273 | DiscrValue x, y, Last; { |
---|
274 | ClassNo c; |
---|
275 | |
---|
276 | ForEach(c, 0, MaxClass) { |
---|
277 | Slice1[c] = Freq[x][c]; |
---|
278 | Slice2[c] = Freq[y][c]; |
---|
279 | |
---|
280 | Freq[x][c] += Freq[y][c]; |
---|
281 | Freq[y][c] = Freq[Last][c]; |
---|
282 | } |
---|
283 | |
---|
284 | Slice1[MaxClass + 1] = ValFreq[x]; |
---|
285 | Slice2[MaxClass + 1] = ValFreq[y]; |
---|
286 | |
---|
287 | ValFreq[x] += ValFreq[y]; |
---|
288 | ValFreq[y] = ValFreq[Last]; |
---|
289 | } |
---|
290 | |
---|
291 | /*************************************************************************/ |
---|
292 | /* */ |
---|
293 | /* Restore old class distribution figures of discrete attribute */ |
---|
294 | /* values x and y from Slice1 and Slice2 */ |
---|
295 | /* */ |
---|
296 | /*************************************************************************/ |
---|
297 | |
---|
298 | Uncombine(x, y) |
---|
299 | /* --------- */ |
---|
300 | DiscrValue x, y; { |
---|
301 | ClassNo c; |
---|
302 | |
---|
303 | ForEach(c, 0, MaxClass) { |
---|
304 | Freq[x][c] = Slice1[c]; |
---|
305 | Freq[y][c] = Slice2[c]; |
---|
306 | } |
---|
307 | |
---|
308 | ValFreq[x] = Slice1[MaxClass + 1]; |
---|
309 | ValFreq[y] = Slice2[MaxClass + 1]; |
---|
310 | } |
---|
311 | |
---|
312 | /*************************************************************************/ |
---|
313 | /* */ |
---|
314 | /* Print the values of attribute Att which are in the subset Ss */ |
---|
315 | /* */ |
---|
316 | /*************************************************************************/ |
---|
317 | |
---|
318 | PrintSubset(Att, Ss) |
---|
319 | /* ----------- */ |
---|
320 | Attribute Att;Set Ss; { |
---|
321 | DiscrValue V1; |
---|
322 | Boolean First = true; |
---|
323 | |
---|
324 | ForEach(V1, 1, MaxAttVal[Att]) { |
---|
325 | if (In(V1, Ss)) { |
---|
326 | if (First) { |
---|
327 | First = false; |
---|
328 | } else { |
---|
329 | printf(", "); |
---|
330 | } |
---|
331 | |
---|
332 | printf("%s", AttValName[Att][V1]); |
---|
333 | } |
---|
334 | } |
---|
335 | } |
---|
336 | |
---|
337 | /*************************************************************************/ |
---|
338 | /* */ |
---|
339 | /* Construct and return a node for a test on a subset of values */ |
---|
340 | /* */ |
---|
341 | /*************************************************************************/ |
---|
342 | |
---|
343 | SubsetTest(Node, Att) |
---|
344 | /* ----------- */ |
---|
345 | Tree Node;Attribute Att; { |
---|
346 | ItemCount CountItems(); |
---|
347 | short S, Bytes; |
---|
348 | |
---|
349 | Sprout(Node, Subsets[Att]); |
---|
350 | |
---|
351 | Node->NodeType = BrSubset; |
---|
352 | Node->Tested = Att; |
---|
353 | Node->Errors = 0; |
---|
354 | |
---|
355 | Bytes = (MaxAttVal[Att] >> 3) + 1; |
---|
356 | Node->Subset = (Set *) calloc(Subsets[Att] + 1, sizeof(Set)); |
---|
357 | ForEach(S, 1, Node->Forks) { |
---|
358 | Node->Subset[S] = (Set) malloc(Bytes); |
---|
359 | CopyBits(Bytes, Subset[Att][S], Node->Subset[S]); |
---|
360 | } |
---|
361 | } |
---|