func (data_set *DataSet) calculateCorrelations() (features_correlation [][]float64, answer_correlation []float64) {
// prepare data
features_correlation = make([][]float64, data_set.FeaturesNum)
for i := range features_correlation {
features_correlation[i] = make([]float64, data_set.FeaturesNum)
}
answer_correlation = make([]float64, data_set.FeaturesNum)
answers := make([]float64, len(data_set.Classes))
for i := range data_set.Classes {
answers[i] = float64(data_set.Classes[i])
}
// calculate correlations
for i := range data_set.SamplesByFeature {
for j := 0; j < i; j++ {
corr := math.Abs(stat.Correlation(data_set.SamplesByFeature[i], data_set.SamplesByFeature[j], nil))
features_correlation[i][j] = corr
features_correlation[j][i] = corr
}
}
for i := range data_set.SamplesByFeature {
answer_correlation[i] = math.Abs(stat.Correlation(data_set.SamplesByFeature[i], answers, nil))
}
return
}
func cost2(params score.Parameters, trainset []TrainItem) float64 {
var totalScore float64
enerchan := make(chan float64, len(trainset))
// rankchan := make(chan float64, len(trainset))
exp := make([]float64, len(trainset))
obs := make([]float64, len(trainset))
enerScore := 0.0
// rankScore := 0.0
for i := 0; i < len(trainset); i++ {
go func(i int) {
// protein := protein.LoadMol2("./traindata/" + traindata[i].Receptor)
protein := trainset[i].Receptor
// pos := ligand.LoadMol2("./traindata/" + traindata[i].Positive)
pos := trainset[i].Positive
total := params.Score(&protein, &pos)
// rk := 0.0
// for j := 0; j < len(trainset[i].Negatives); j++ {
// // neg := ligand.LoadMol2("./traindata/" + traindata[i].Negatives[j])
// neg := trainset[i].Negatives[j]
// negTotal := params.Score(&protein, &neg)
// if negTotal <= total {
// rk += 1.0
// }
// }
exp[i] = trainset[i].Energy
obs[i] = total
enerchan <- ((trainset[i].Energy - total) * (trainset[i].Energy - total))
// rankchan <- rk
}(i)
}
for i := 0; i < len(trainset); i++ {
enerScore += <-enerchan
// rankScore += <-rankchan
}
corr := stat.Correlation(exp, obs, nil)
// totalScore = enerScore/(corr*corr) + (enerScore / (corr * corr) * rankScore)
// totalScore = enerScore*(2000.0-1999.0*corr) + (math.Sqrt(enerScore) * (2000.0 - 1999.0*corr) * rankScore * 1000000.0)
totalScore = enerScore * (2000.0 - 1999.0*corr)
// fmt.Printf("PKD %f - Rank %f - Corr %f - TOTAL %f\n", enerScore, rankScore, corr, totalScore)
fmt.Printf("Energy %f - Corr %f - TOTAL %f\n", math.Sqrt(enerScore/float64(len(trainset))), corr, totalScore)
return totalScore
}
func cost3(params score.Parameters, trainset []TrainItem) float64 {
var totalScore float64
// enerchan := make(chan float64, len(trainset))
rankchan := make(chan float64, len(trainset))
exp := make([]float64, len(trainset))
obs := make([]float64, len(trainset))
// nNegatives := 0.0
// enerScore := 0.0
rankScore := 0.0
for i := 0; i < len(trainset); i++ {
go func(i int) {
protein := trainset[i].Receptor
pos := trainset[i].Positive
total := params.Score(&protein, &pos)
rk := 0
for j := 0; j < len(trainset[i].Negatives); j++ {
neg := trainset[i].Negatives[j]
negTotal := params.Score(&protein, &neg)
if negTotal <= total {
rk += 1
if rk == 5 {
break
}
if rk > 5 {
fmt.Println("F**K")
}
}
}
exp[i] = trainset[i].Energy
obs[i] = total
// enerchan <- ((trainset[i].Energy - total) * (trainset[i].Energy - total))
rankchan <- float64(rk) / 5.0
// nNegatives += float64(len(trainset[i].Negatives))
}(i)
}
// var ranktmp float64
for i := 0; i < len(trainset); i++ {
rankScore += <-rankchan
}
// ranktmp = <-rankchan
// if ranktmp > rankScore {
// rankScore = ranktmp
// }
// // enerScore += <-enerchan
// //
// }
rankScore = rankScore / float64(len(trainset))
corr := stat.Correlation(exp, obs, nil)
corrSquared := 0.0
if corr > 0.0 {
corrSquared = corr * corr
}
// totalScore = enerScore/(corr*corr) + (enerScore / (corr * corr) * rankScore)
// totalScore = enerScore*(2000.0-1999.0*corr) + (math.Sqrt(enerScore) * (2000.0 - 1999.0*corr) * rankScore * 1000000.0)
totalScore = (math.Pow(1-corrSquared, 2) + math.Pow(rankScore, 2)) * 1000
// fmt.Printf("PKD %f - Rank %f - Corr %f - TOTAL %f\n", enerScore, rankScore, corr, totalScore)
fmt.Printf("Corr %f - Corr^2 %f - Rank %f - TOTAL %f\n", corr, corrSquared, rankScore, totalScore)
return totalScore
}
