Machine Learning for Software Engineers — Part 2: GoLearn
This is part two of a two-part comprehensive, high-level exploration of Machine Learning for software engineers. If you have not read Part 1 click here.
Now that we have a decent understanding of how Machine Learning works and what it is used for, it is a good time to jump into some implementation. Below are some examples that any software engineer can use to improve their project without a complex deep understanding of the inner-workings of a deep neural network.
These examples are written in Go using the GoLearn library. It is very similar to the sci-kit learn library in Python, and all examples can be translated across languages if you desire.
Linear Regression
In high school we were all at some point given a graph with points on it and told to draw a line that goes through the middle of the data. This line was called the “line of best fit.” What they didn’t tell you is that you were really doing a statistical analysis of the data.
Linear regression employs supervised learning. A set of many data points is provided to a model. These data points may have many features (x1, x2, x3), but usually only two labels (y1, y2).
The model used these features and labels to calculate a line (y=wx+b). w is the weight that is applied to each feature and b is the bias or any initial condition for the line.
While training a model, the model’s Fit() function picks a random bias and weights that are used to calculate the MSE. The bias and weights are then changed to get the smallest MSE possible.
There are many other variations on linear regression. There is multi-linear regression that uses many lines to classify into more than two groups. There are also quadratic regression, cubic regression, and logistic regression, where other line definitions are used. In logistic regression the line is defined as y= 1/(1 + e^(-wx)) + b. Again
weights and bias are used to optimize the line while the Fit()function calculates an error.
Here is an example of a linear regression program written in Go using the GoLearn package.
The first step is to bring in the data. The data object here is a matrix implementation where each column corresponds to each feature in the CSV.
data, err := base.ParseCSVToInstances("../datasets/Advertising.csv", true)
if err != nil {
panic(err)
}Next, the data is split into training and test sets that are used to fit the model and to test the resulting model. The two parameters provided are the matrix that we want to split and the percentage that is to be in the training set. In this example, 70% is used for training and 30% is used for testing. It is important to note that this function does randomize the data as it splits between training and test sets. This is to prevent bias that may be included in the ordering of the data in the csv.
trainData, testData := base.InstancesTrainTestSplit(data, 0.70)Now to fit and test the model.
r := regression.NewLinearRegression()
err = r.Fit(trainData)
if err != nil {
panic(err)
}
predictions, err := r.Predict(testData)
if err != nil {
panic(err)
}Lastly, view the results of the fit.
fmt.Println("Linear Regression (information gain)")
cf, err := evaluation.GetConfusionMatrix(testData, predictions)
if err != nil {
panic(fmt.Errorf("Unable to get confusion matrix: %s", err.Error())) }
fmt.Println(evaluation.GetSummary(cf))
}Here a Confusion Matrix is used to display True Positive, False Positive, True Negatives, False Negatives, Precision, Recall, and the F1 score.
Decision Tree
Another commonly used model is the decision tree. The decision tree uses a series of nodes and branches to make selections across each of the data’s features. These trees can have hundreds of nodes and handle complex models with hundreds of features.
The following example will be an end to end implementation of a name classification model. It used a data set of tweets about the 2018 FIFA World Cup downloaded from Kaggle.
First thing is to get the data from the CSV. The data is stored into a slice and struct that will be used to get the appropriate features for training and testing the model.
// Tweet object for holding raw form of csv data.
type Tweet struct {
ID int
Lang string
Date int64
Source string
Length int
OrgTweet string
Tweets string
Likes int
Retweets int
Hashtag string
Usermention string
UserMentionID string
Name string
Place string
Followers int
Friends int
}func getData(file string) ([]Tweet, error) {
f, err := os.Open(file)
defer f.Close()
if err != nil {
return nil, err
}
lines, err := csv.NewReader(f).ReadAll()
if err != nil {
return nil, err
}
var tweets []Tweet
for i, line := range lines {
if i == 0 {
continue
}
id, _ := strconv.Atoi(line[0])
layout := "2006-01-02"
d, _ := time.Parse(layout, line[2])
l, _ := strconv.Atoi(line[4])
likes, _ := strconv.Atoi(line[7])
r, _ := strconv.Atoi(line[8])
f, _ := strconv.Atoi(line[14])
friends, _ := strconv.Atoi(line[14])
t := Tweet{
ID: id,
Lang: line[1],
Date: d.Unix(),
Source: line[3],
Length: l,
OrgTweet: line[5],
Tweets: line[6],
Likes: likes,
Retweets: r,
Hashtag: line[9],
Usermention: line[10],
UserMentionID: line[11],
Name: line[12],
Place: line[13],
Followers: f,
Friends: friends, }
tweets = append(tweets, t) } return tweets, err}
Feature engineering is the fancy name for getting the viable data for your model and cleaning it up. To classify names for other words, we need to pull out and label text features.
func plotWords(df dataframe.DataFrame) error {
var data []WordData
mapCounts := make(map[string]int, 100000)
// make map of names that will be labeled as true
// there are two columns that contain name data. WE need to take names from
// both columns
colNames := []string{"Name", "Usermention"}
for _, colName := range colNames { vals := df.Col(colName).Records()
for _, val := range vals {
// split multi word names into single words
words := strings.Split(val, " ")
for _, w := range words {
// remove spaces
w = strings.Trim(w, " ")
if _, ok := mapCounts[w];
!ok {
mapCounts[w] = 1
continue
}
mapCounts[w]++ } }
// add data to struct that contains important data information. isName := true
for name, count := range mapCounts {
d := WordData{
Word: name,
Occurances: count,
IsName: isName,
}
data = append(data, d)
} }
// parse each word in the tweets and lable it as name or not
records := df.Col("OrgTweet").Records()
tweetWordsCount := make(map[string]int, 100000)
for _, r := range records {
words := strings.Split(r, " ")
for _, w := range words {
w = strings.Trim(w, " ")
// skip names that we already have
if mapCounts[w] > 0 {
continue
}
// we assume we have all the names. This is going
// cause some un realiability in the data.
if _, ok := tweetWordsCount[w];
!ok {
tweetWordsCount[w] = 1
continue
}
tweetWordsCount[w]++
} }
// add stata to struct
for name, count := range tweetWordsCount {
d := WordData{
Word: name,
Occurances: count,
IsName: false,
}
data = append(data, d)
}
fmt.Printf("From the Fifa World Cup Data you have parsed out %d unique words. \n", len(data))
// create a new data frame
newDF := dataframe.LoadStructs(data)
fmt.Println(newDF)
// get summary of word occurances fmt.Println(newDF.Select([]string{"Occurances"}).Describe())
// cache data frame in csv
myFile, err := os.Create("data/words.csv")
if err != nil {
return err
}
err = newDF.WriteCSV(myFile)
if err != nil {
return err
}
return nil
}Just like before, we have to import data, train test split, define the model, Fit, Predict, and print the results.
// DecisionTree Write and train DecisionTree model.func DecisionTree(file string) error {
// var wg sync.WaitGroup
rand.Seed(44111342)
// Load in the words dataset
words, err := base.ParseCSVToInstances(file, true)
if err != nil {
return err
}
// Create a 70-30 training-test split trainData, testData := base.InstancesTrainTestSplit(words, 0.70) fmt.Println("ID3Decision Tree train") tree := trees.NewID3DecisionTree(0.6)
// Train the ID3 tree
err = tree.Fit(trainData)
if err != nil {
return err
}
// Generate predictions
predictions, err := tree.Predict(testData)
if err != nil {
return err
}
// Evaluate
fmt.Println("ID3 Performance (information gain)")
cf, err := evaluation.GetConfusionMatrix(testData, predictions)
if err != nil {
return fmt.Errorf("Unable to get confusion matrix: %s", err.Error())
}
fmt.Println(evaluation.GetSummary(cf))
tree.Save("models/DecisonTree.h")
return err
}
If you are brave you can go ahead, visit the GoLearn, and start implementing you own model.
For more information visit these sites:
