[SOLVED] Cs6035 machine learning 2025

Learning Goals of this Project:Important HighlightsImportant Reference Materials:Project Overview VideoThis is a 16 minute video by the project creator, it covers project concepts.There are other videos on the Setup page that cover installation and other subjects.BACKGROUNDMany of the Projects in CS6035 are focused on offensive security tasks. These are related to Red Team activities/tasks that many of us may associate with cybersecurity. This project will be focused on defensive security tasks, which are usually considered Blue Team activities that are done by many corporate teams.Historically, many defensive security professionals have investigated malicious activity, files, and code. They investigate these to create patterns (often called signatures) that can be used to detect (and prevent) malicious activity, files, and code when that pattern is used again. What this means is that these simple methods only were effective on known threats.This approach was relatively effective in preventing known malware from infecting systems, but it did nothing to protect against novel attacks. As attackers became more sophisticated, they learned to tweak or simply encode their malicious activity, files, or code to avoid detection from these simple pattern matching detections.With this background information, it would be nice if a more general solution could give a score to activity, files, and code that pass through corporate systems every day. This solution would inform the security team that while a certain pattern may not exactly fit a signature of known malicious activity, files, or code it appears to be very similar to examples that were seen in the past that were malicious.Luckily machine learning models can do exactly that if provided with proper training data! Thus, it is no surprise that one of the most powerful tools in the hands of defensive cybersecurity professionals is Machine Learning. Modern detection systems usually use a combination of machine learning models and pattern matching (regular expressions) to detect and prevent malicious activity on networks and devices.This project will focus on teaching the fundamentals of data analysis and building/testing your own machine learning models in python. You’ll be using the open source libraries Pandas and Scikit-Learn.Cybersecurity Machine Learning Careers and TrendsAdditional InformationTable of contents Task 1:Task 1For the first task, let’s get familiar with some pandas basics. pandas is a Python library that deals with Dataframes, which you can think of as a Python class that handles tabular data. In the real world, you would create graphics and other visuals to better understand the dataset you are working with. You would also use plotting tools like PowerBi, Tableau, Data Studio, and Matplotlib. This step is generally known as Exploratory Data Analysis. Since we are using an autograder for this class, we will skip the plotting for this project.For this task, we have released a local test suite. If you are struggling to understand the expected input and outputs for a function, please set up the test suite and use it to debug your function. Please note that the return lines for the provided skeleton functions are placeholders for the data types that the tests are expecting.It’s critical you pass all tests locally before you submit to Gradescope for credit. Do not use Gradescope for debugging.TheoryIn this Task, we’re not yet getting into theory. It’s more nuts and bolts – you will learn the basics of pandas. pandas dataframes are something of a glorified list of lists, mixed in with a dictionary. You get a table of values with rows and columns, and you can modify the column names and index values for the rows. There are numerous functions built into pandas to let you manipulate the data in the dataframe.To be clear, pandas is not part of Python, so when you look up docs, you’ll specifically want the official Pydata pandas docs. Note that we linked to the API docs here, this is the core of the docs you’ll be looking at.You can always get started trying to solve a problem by looking at Stack Overflow posts in Google search results. There you’ll find ideas about how to use the pandas library. In the end, however, you should find yourself in the habit of looking directly at the docs for whichever library you are using, pandas in this case.For those who might need a concrete example to get started, here’s how you would take a pandas dataframe column and return the average of its values:import pandas as pd# create a dataframe from a Python dictdf = pd.DataFrame({“color”:[“yellow”, “green”, “purple”, “red”], “weight”:[124,4.56,384,-2]})df # shows the dataframeNote that the column names are [“color”,”weight”] while the index is [0,1,2,3…] where […] the brackets denote a list.Now that we have created a dataframe, we can find the average weight by summing the values under ‘weight’ and dividing them by the sum:average = df[‘weight’].sum() / len(df[‘weight’])average # if you put a variable as the last line, the variable is printed127.63999999999999Note: In the example above, we’re not paying attention to rounding, you will need to round your answers to the precision asked for in each Task.Also note, we are using slightly older versions of the pandas, Python and other libraries so be sure to look at the docs for the appropriate library version. Often there’s a drop-down at the top of docs sites to select the older version.Refer to the Submissions page for details about submitting your work.Useful Links:Deliverables:Instructions:The Task1.py file has function skeletons that you will complete with Python code, mostly using the pandas library. The goal of each of these functions is to give you familiarity with the pandas library and some general Python concepts like classes, which you may not have seen before. See information about the function’s inputs, outputs, and skeletons below.Table of contentsfind_data_typeIn this function you will take a dataset and the name of a column in it. You will return the column’s data type.Useful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.dtypes.htmlINPUTSOUTPUTSnp.dtype – data type of the columnFunction Skeletondef find_data_type(dataset:pd.DataFrame,column_name:str) -> np.dtype:return np.dtype()set_index_colIn this function you will take a dataset and a series and set the index of the dataset to be the seriesUseful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.Index.htmlINPUTSOUTPUTSa pandas DataFrame indexed by the given index seriesFunction Skeletondef set_index_col(dataset:pd.DataFrame,index:pd.Series) -> pd.DataFrame:return pd.DataFrame()reset_index_colIn this function you will take a dataset with an index already set and reindex the dataset from 0 to n-1, dropping the old indexUseful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.reset_index.htmlINPUTSOUTPUTSa pandas DataFrame indexed from 0 to n-1Function Skeletondef reset_index_col(dataset:pd.DataFrame) -> pd.DataFrame:return pd.DataFrame()set_col_typeIn this function you will be given a DataFrame, column name and column type. You will edit the dataset to take the column name you are given and set it to be the type given in the input variableUseful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.astype.htmlINPUTSOUTPUTSa pandas DataFrame with the column in column_name changed to the type in new_col_typeFunction Skeleton# Set astype (string, int, datetime)def set_col_type(dataset:pd.DataFrame,column_name:str,new_col_type:type) -> pd.DataFrame:return pd.DataFrame()make_DF_from_2d_arrayIn this function you will take data in an array as well as column and row labels and use that information to create a pandas DataFrameUseful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.htmlINPUTSOUTPUTSa pandas DataFrame with columns set from column_name_list, row index set from index and data set from array_2dFunction Skeleton# Take Matrix of numbers and make it into a DataFrame with column name and index numberingdef make_DF_from_2d_array(array_2d:np.array,column_name_list:list[str],index:pd.Series) -> pd.DataFrame:return pd.DataFrame()sort_DF_by_columnIn this function, you are given a dataset and column name. You will return a sorted dataset (sorting rows by the value of the specified column) either in descending or ascending order, depending on the value in the descending variable.Useful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.sort_values.htmlINPUTSOUTPUTSa pandas DataFrame sorted by the given column name and in descending or ascending order depending on the value of the descending variableFunction Skeleton# Sort DataFrame by valuesdef sort_DF_by_column(dataset:pd.DataFrame,column_name:str,descending:bool) -> pd.DataFrame:return pd.DataFrame()drop_NA_colsIn this function you are given a DataFrame. You will return a DataFrame with any columns containing NA values droppedUseful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.dropna.htmlINPUTSOUTPUTSa pandas DataFrame with any columns that contain an NA value droppedFunction Skeleton# Drop NA values in DataFrame Columns def drop_NA_cols(dataset:pd.DataFrame) -> pd.DataFrame:return pd.DataFrame()drop_NA_rowsIn this function you are given a DataFrame you will return a DataFrame with any rows containing NA values droppedUseful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.dropna.htmlINPUTSOUTPUTSa pandas DataFrame with any rows that contain an NA value droppedFunction Skeletondef drop_NA_rows(dataset:pd.DataFrame) -> pd.DataFrame:return pd.DataFrame()make_new_columnIn this function you are given a dataset, a new column name and a string value to fill in the new column. Add the new column to the dataset and return the dataset.Useful Resourceshttps://pandas.pydata.org/pandas-docs/stable/getting_started/intro_tutorials/05_add_columns.htmlINPUTSOUTPUTSa pandas DataFrame with the new column created named new_column_name and filled with the value in new_column_valueFunction Skeletondef make_new_column(dataset:pd.DataFrame,new_column_name:str,new_column_value:list) -> pd.DataFrame:return pd.DataFrame()left_merge_DFs_by_columnIn this function you are given 2 datasets and the name of a column with which you will left join them on using the pandas merge method. The left dataset is dataset1 right dataset is dataset2, for example purposes.Useful Resourceshttps://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.merge.html https://stackoverflow.com/questions/53645882/pandas-merging-101INPUTSOUTPUTSa pandas DataFrame containing the two datasets left joined together on the given column nameFunction Skeletondef left_merge_DFs_by_column(left_dataset:pd.DataFrame,right_dataset:pd.DataFrame,join_col_name:str) -> pd.DataFrame:return pd.DataFrame()simpleClassThis project will require you to work with Python Classes. If you are not familiar with them we suggest learning a bit more about them.You will take the inputs into the class initialization and set them as instance variables (of the same name) in the Python class.Useful Resourceshttps://www.w3schools.com/python/python_classes.aspINPUTSOUTPUTSNone, just setup the init method in the class.Function Skeletonclass simpleClass():def __init__(self, length:int, width:int, height:int):passfind_dataset_statisticsNow that you have learned a bit about pandas DataFrames, we will use them to generate some simple summary statistics for a DataFrame. You will be given the dataset as an input variable, as well as a column name for a column in the dataset that serves as a label column. This label column contains binary values (0 and 1) that you also summarize, and also the variable to predict.In this context:This type of binary classification is common in machine learning tasks where we want to be able to predict the field. An example of where this could be useful would be if we were looking at network data, and the label column was IsVirus. We could then analyze the network data of Georgia Tech services and predict if incoming files look like a virus (and if we should alert the security team).Useful ResourcesINPUTSOUTPUTSHint: Consider using the int function to type cast decimalsFunction Skeletondef find_dataset_statistics(dataset:pd.DataFrame,label_col:str) -> tuple[int,int,int,int,int]:n_records = #TODOn_columns = #TODOn_negative = #TODOn_positive = #TODOperc_positive = #TODOreturn n_records,n_columns,n_negative,n_positive,perc_positive    Task 2:Now that you have a basic understanding of pandas and the dataset, it is time to dive into some more complex data processing tasks.TheoryIn machine learning a common goal is to train a model on one set of data. Then we validate the model on a similarly structured but different set of data. You could, for example, train the model on data you have collected historically. Then you would validate the model against real-time data as it comes in, seeing how well it predicts the new data coming in.If you’re looking at a past dataset as we are in these tasks, we need to treat different parts of the data differently to be able to develop and test models. We segregate the data into test and training portions. We train the model on the training data and test the developed model on the test data to see how well it predicts the results.You should never train your models on test data, only on training data.NotesAt a high level it is important to hold out a subset of your data when you train a model. You can see what the expected performance is on unseen sample. Thus, you can determine if the resulting model is overfit (performs much better on training data vs test data).Preprocessing data is essential because most models only take in numerical values. Therefore, categorical features need to be “encoded” to numerical values so that models can use them. A machine learning model may not be able to make sense of “green”, “blue” and “red.” In preprocessing, we’ll convert those to integer values 1, 2 and 3, for example. It’s an interesting question as to what happens when you have training data that has “green,” “red” and blue,” but your testing data says “yellow.”Numerical scaling can be more or less useful depending on the type of model used, but it is especially important in linear models. Numerical scaling is typically taking positive value and “compressing” them into a range between 0 and 1 (inclusive) that retains the relationships among the original data.These preprocessing techniques will provide you with options to augment your dataset and improve model performance.Useful Links:Deliverables:Instructions:The Task2.py File has function skeletons that you will complete with python code (mostly using the pandas and scikit-learn libraries). The Goal of each of these functions is to give you familiarity with the applied concepts of Splitting and Preprocessing Data. See information about the Function’s Inputs, Outputs and Skeletons belowTable of contentsttsIn this function, you will take:You will return features and labels for the training and test sets.At a high level, you can separate the task into two subtasks. The first is splitting your dataset into both features and labels (by columns), and the second is splitting your dataset into training and test sets (by rows). You should use the scikit-learn train_test_split function but will have to write wrapper code around it based on the input values we give you.Useful ResourcesINPUTSOUTPUTSFunction Skeletondef tts(  dataset: pd.DataFrame,label_col: str,test_size: float,should_stratify: bool,random_state: int) -> tuple[pd.DataFrame,pd.DataFrame,pd.Series,pd.Series]:# TODOreturn train_features,test_features,train_labels,test_labelsPreprocessDatasetThe PreprocessDataset Class contains a code skeleton with nine methods for you to implement. Most methods will be split into two parts: one that will be run on the training dataset and one that will be run on the test dataset. In Data Science/Machine Learning, this is done to avoid something called Data Leakage.For this assignment, we don’t expect you to understand the nuances of the concept, but we will have you follow principles that will minimize the chances of it occurring. You will accomplish this by splitting data into training and test datasets and processing those datasets in slightly different ways.Generally, for everything you do in this project, and if you do any ML or Data Science work in the future, you should train/fit on the training data first, then predict/transform on the training and test data. That holds up for basic preprocessing steps like task 2 and for complex models like you will see in tasks 3 and 4.For the purposes of this project, you should never train or fit on the test data (and more generally in any ML project) because your test data is expected to give you an understanding of how your model/predictions will perform on unseen data. If you fit even a preprocessing step to your test data, then you are either giving the model information about the test set it wouldn’t have about unseen data (if you combine train and test and fit to both), or you are providing a different preprocessing than the model is expecting (if you fit a different preprocessor to the test data), and your model would not be expected to perform well.Note: You should train/fit using the train dataset; then, once you have a fit encoder/scaler/pca/model instance, you can transform/predict on the training and test data.You will also notice that we are only preprocessing the Features and not the Labels. There are a few cases where preprocessing steps on labels may be helpful in modeling, but they are definitely more advanced and out of the scope of this introduction. Generally, you will not need to do any preprocessing to your labels beyond potentially encoding a string value (i.e., “Malware” or “Benign”) into an integer value (0 or 1), which is called Label Encoding.PreprocessDataset:__init__Similar to the Task1 simpleClass subtask you previously completed you will initialize the class by adding instance variables (add all the inputs to the class).Useful ResourcesINPUTSExample of feature_engineering_functions:def double_height(dataframe:pd.DataFrame):return dataframe[“height”] * 2 def half_height(dataframe:pd.DataFrame):return dataframe[“height”] / 2 feature_engineering_functions = {“double_height”:double_height,”half_height”:half_height}Don’t worry about copying it we also have examples in the local test cases this is just provided as an illustration of what to expect in your function.OUTPUTSNone, just assign all the input parameters to class variables.Also per the instructions below, you’ll return here and create another instance variable: a scikit-learn OneHotEncoder with any Parameters you may need later.Function Skeletondef __init__(self,one_hot_encode_cols:list[str],min_max_scale_cols:list[str],n_components:int,feature_engineering_functions:dict):# TODO: Add any instance variables you may need to make your functions workreturnPreprocessDataset:one_hot_encode_columns_train and one_hot_encode_columns_testOne Hot Encoding is the process of taking a column and returning a binary vector representing the various values within it. There is a separate function for the training and test datasets since they should be handled separately to avoid data leakage (see the 3rd link in Useful Resources for a little more info on how to handle them).Pseudocodeone_hot_encode_columns_train()one_hot_encode_columns_test()Example Walkthrough (from Local Testing suite):INPUTS:one_hot_encode_cols[“src_ip”,”protocol”]Train FeaturesTest FeaturesTrain DataFrames at each step:2.DataFrame with columns to encode:DataFrame with other columns:4.One Hot Encoded 2d array:5.One Hot Encoded DataFrame with Index and Column Names6.Final DataFrame with passthrough/other columns joined backTest DataFrames at each step:1.DataFrame with columns to encode:DataFrame with other columns:2.One Hot Encoded 2d array:3.One Hot Encoded DataFrame with Index and Column Names4.Final DataFrame with passthrough columns joined backNote: For the local tests and autograder use the column naming scheme of joining the previous column name and the column value with an underscore (similar to above where Type -> Type_Fruit and Type_Vegetable)Note 2: Since you should only be fitting your encoder on the training data, if there are values in your test set that are different than those in the training set, you will denote that with 0s. In the example above, let’s say we have a row in the test set with pizza, which is neither a fruit nor vegetable for the Type_Fruit and Type_Vegetable. It should result in a 0 for both columns. If you don’t handle these properly, you may get errors like Test Failed: Found unknown categories.Note 3: You may be tempted to use the pandas function get_dummies to solve this task, but its a trap. It seems easier, but you will have to do a lot more work to make it handle a train/test split. So, we suggest you use scikit-learn’s OneHotEncoder.Useful ResourcesINPUTSOUTPUTSa pandas DataFrame with the columns listed in one_hot_encode_cols one hot encoded and all other columns in the DataFrame unchangedFunction Skeletondef one_hot_encode_columns_train(self,train_features:pd.DataFrame) -> pd.DataFrame:one_hot_encoded_dataset = pd.DataFrame()return one_hot_encoded_datasetdef one_hot_encode_columns_test(self,test_features:pd.DataFrame) -> pd.DataFrame:one_hot_encoded_dataset = pd.DataFrame()return one_hot_encoded_datasetPreprocessDataset:min_max_scaled_columns_train and min_max_scaled_columns_testMin/Max Scaling is a process to transform numerical features to a specific range, typically [0, 1], to ensure that input values are comparable (similar to how you may have heard of “normalizing” data) and is a crucial preprocessing step for many machine learning algos. In particular this standardization is essential for algorithms like linear regression, logistic regression, k-means, and neural networks, which can be sensitive to the scale of input features, whereas some algos like decision trees are less impacted.By applying Min/Max Scaling, we prevent feature dominance, to ideally improve performance and accuracy of these algorithms and improve training convergence. It’s a recommended step to ensure your models are trained on consistent and standardized data.For the provided assignment you should use the scikit-learn MinMaxScaler function (linked in the resources below) rather than attempting to implement your own scaling function.The rough implementation of the scikit-learn function is provided below for educational purposes.X_std = (X – X.min(axis=0)) / (X.max(axis=0) – X.min(axis=0))X_scaled = X_std * (max – min) + minNote: There are separate functions for the training and test datasets to help avoid data leakage between the test/train datasets. Please refer to the 3rd link in Useful Resources for more information on how to handle this – namely that we should still scale the test data based on our “knowledge” of the train dataset.Example Dataframe:Example Min Max Scaled Dataframe (rounded to 4 decimal places):Note: For the Autograder use the same column name as the original column (ex: Price -> Price)Useful ResourcesINPUTSOUTPUTSa pandas DataFrame with the columns listed in min_max_scale_cols min/max scaled and all other columns in the DataFrame unchangedFunction Skeletondef min_max_scaled_columns_train(self,train_features:pd.DataFrame) -> pd.DataFrame:min_max_scaled_dataset = pd.DataFrame()return min_max_scaled_datasetdef min_max_scaled_columns_test(self,test_features:pd.DataFrame) -> pd.DataFrame:min_max_scaled_dataset = pd.DataFrame()return min_max_scaled_datasetPreprocessDataset:pca_train and pca_testPrincipal Component Analysis is a dimensionality reduction technique (column reduction). It aims to take the variance in your input columns and map the columns into N columns that contain as much of the variance as it can. This technique can be useful if you are trying to train a model faster and has some more advanced uses, especially when training models on data which has many columns but few rows. There is a separate function for the training and test datasets because they should be handled separately to avoid data leakage (see the 3rd link in Useful Resources for a little more info on how to handle them).Note 1: For the local tests and autograder, use the column naming scheme of column names: component_1, component_2 .. component_n for the n_components passed into the __init__ method.Note 2: For your PCA outputs to match the local tests and autograder, make sure you set the seed using a random state of 0 when you initialize the PCA function.Note 3: Since PCA does not work with NA values, make sure you drop any columns that have NA values before running PCA.Useful ResourcesINPUTSOUTPUTSa pandas DataFrame with the generated pca values and using column names: component_1, component_2 .. component_nFunction Skeletondef pca_train(self,train_features:pd.DataFrame) -> pd.DataFrame:# TODO: Read the function description in https://github.gatech.edu/pages/cs6035-tools/cs6035-tools.github.io/Projects/Machine_Learning/Task2.html and implement the function as describedpca_dataset = pd.DataFrame()return pca_dataset def pca_test(self,test_features:pd.DataFrame) -> pd.DataFrame:# TODO: Read the function description in https://github.gatech.edu/pages/cs6035-tools/cs6035-tools.github.io/Projects/Machine_Learning/Task2.html and implement the function as describedpca_dataset = pd.DataFrame()return pca_datasetPreprocessDataset:feature_engineering_train, feature_engineering_testFeature Engineering is a process of using domain knowledge (physics, geometry, sports statistics, business metrics, etc.) to create new features (columns) out of the existing data. This could mean creating an area feature when given the length and width of a triangle or extracting the major and minor version number from a software version or more complex logic depending on the scenario.In cybersecurity in particular, feature engineering is crucial for using domain expert’s (e.g. a security analyst) experience to identify anomalous behavior that might signify a security breach. This could involve creating features that represent deviations from established baselines, such as unusual file access patterns, unexpected network connections, or sudden spikes in CPU usage. These anomaly-based features can help distinguish malicious activity from normal system operations, but the system does not know what data patterns mean anomalous off-hand – that is where you as the domain expert can help by creating features.These methods utilize a dictionary, feature_engineering_functions, passed to the class constructor (__init__). This dictionary defines how to generate new features:Example of what could be passed as the feature_engineering_functions dictionary to __init__:import pandas as pddef double_height(dataframe: pd.DataFrame) -> pd.Series:return dataframe[“height”] * 2 def half_height(dataframe: pd.DataFrame) -> pd.Series:return dataframe[“height”] / 2 example_feature_engineering_functions = {“double_height”: double_height, # Note that functions in python can be passed around and used just like data!“half_height”: half_height} # and the class may be been created like this…# preprocessor = PreprocessDataset(…, feature_engineering_functions=example_feature_engineering_functions, …)In particular for this method, you will be taking in a dictionary with a column name and a function that takes in a DataFrame and returns a column. You’ll be using that to create a new column with the name in the dictionary key. Therefore if you were given the above functions, you would create two new columns named “double_height” and “half_height” in your Dataframe.Useful ResourcesINPUTSOUTPUTSa pandas dataframe with the features described in feature_engineering_train and feature_engineering_test added as new columns and all other columns in the dataframe unchangedFunction Skeletondef feature_engineering_train(self,train_features:pd.DataFrame) -> pd.DataFrame:feature_engineered_dataset = pd.DataFrame()return feature_engineered_datasetdef feature_engineering_test(self,test_features:pd.DataFrame) -> pd.DataFrame:feature_engineered_dataset = pd.DataFrame()return feature_engineered_datasetPreprocessDataset:preprocess_train, preprocess_testNow, we will put three of the above methods together into a preprocess function. This function will take in a dataset and perform encoding, scaling, and feature engineering using the above methods and their respective columns. You should not perform PCA for this function.Useful ResourcesSee resources for one hot encoding, min/max scaling and feature engineering aboveINPUTSOUTPUTSa pandas dataframe for both test and train features with the columns in one_hot_encode_cols encoded, the columns in min_max_scale_cols scaled and the columns described in feature_engineering_functions engineered. You do not need to use PCA here.Function Skeletondef preprocess_train(self,train_features:pd.DataFrame) -> pd.DataFrame:train_features = pd.DataFrame()return train_features def preprocess_test(self,test_features:pd.DataFrame) -> pd.DataFrame:test_features = pd.DataFrame()return test_features Task 3In Task 2 you learned how to split a dataset into training and testing components. Now it’s time to learn about using a K-means model. We will run a basic model on the data to cluster files (rows) with similar attributes together. We will use an unsupervised model.TheoryAn unsupervised model has no label column. By constrast, in supervised learning (which you’ll see in Task 4) the data has features and targets/labels. These labels are effectively an answer key to the data in the feature columns. You don’t have this answer key in unsupervised learning, instead you’re working on data without labels. You’ll need to choose algorithms that can learn from the data, exclusively, without the benefit of lablels.We start with K-means because it is simple to understand the algorithm. For the Mathematics people, you can look at the underlying data structure, a Voronoi diagram. Based on squared Euclidian distances, K-means creates clusters of similar datapoints. Each cluster has a centroid. The idea is that for each sample, it’s associated/clustered with the centroid that is the “closest.”Closest is an interesting concept in higher dimensions. You can think of each feature in a dataset as a dimension in the data. If it’s 2d or 3d, we can visualize it easily. Concepts of distance are clear in 2d and 3d, and they work similarly in 4+d.If you read the Wikipedia articles for K-means you’ll see a discussion of the use of “squared Euclidean distances” in K-means. This is compared with simple Euclidean distances in the Weber problem, and better approaches resulting from k-medians and k-mediods is discussed.Please use scikit-learn to create the model and Yellowbrick to determine the optimal value of k for the dataset.So far, we have functions to split the data and preprocess it. Now, we will run a basic model on the data to cluster files (rows) with similar attributes together. We will use an unsupervised model (model with no label column), K-means. Again, use scikit-learn to create the model and Yellowbrick to determine the optimal value of k for the dataset.Refer to the Submissions page for details about submitting your work.Useful Links:Deliverables:Local Test Dataset InformationFor this task the local test dataset we are using is the NATICUSdroid dataset, which contains 86 columns of data related to android permissions used by benign and malicious Android applications released between 2010 and 2019. For more information such as the introductory paper and the Citations/Acknowledgements you can view the dataset site in the UCI ML repository. In this specific case clustering can be a useful tool to group apps that request similar permissions together. The team that created this dataset hypothesized that malicious apps would exhibit distinct patterns in the types of permissions they request compared to benign apps. This difference in permission request patterns could potentially be used to distinguish between malicious and benign applications.Instructions:The Task3.py File has function skeletons that you will complete with Python code. You will mostly be using the pandas, Yellowbrick and scikit-learn libraries. The goal of each of these functions is to give you familiarity with the applied concepts of Unsupervised Learning. See information about the function’s Inputs, Outputs and Skeletons below.KmeansClusteringThe KmeansClustering Class contains a code skeleton with 4 methods for you to implement.Note: You should train/fit using the train dataset then once you have a Yellowbrick/K-means model instance you can transform/predict on the training and test data.KmeansClustering:__init__Similar to Task 1, you will initialize the class by adding instance variables as needed.Useful ResourcesINPUTSOUTPUTSNoneFunction Skeletondef __init__(self,random_state: int):# TODO: Add any state variables you may need to make your functions workpassKmeansClustering:kmeans_trainKmeans Clustering is a process of grouping together similar rows together and assigning them to a cluster. For this method you will use the training data to fit an optimal K-means cluster on the data.To help you get started we have provided a list of subtasks to complete for this task:Useful ResourcesINPUTSOUTPUTSa list of cluster ids that the K-means model has assigned for each row in the train datasetFunction Skeletondef kmeans_train(self,train_features:pd.DataFrame) -> list:cluster_ids = list()return cluster_idsKmeansClustering:kmeans_testK-means clustering is a process of grouping together similar rows together and assigning them to a cluster. For this method you will use the training data to fit an optimal K-means cluster on the test data.To help you get started, we have provided a list of subtasks to complete for this task:Useful ResourcesINPUTSOUTPUTSa list of cluster ids that the K-means model has assigned for each row in the test datasetFunction Skeletondef kmeans_test(self,test_features:pd.DataFrame) -> list:cluster_ids = list()return cluster_idsKmeansClustering:train_add_kmeans_cluster_id_feature, test_add_kmeans_cluster_id_featureUsing the two methods you completed above (kmeans_train and kmeans_test) you will add a new feature(column) to the training and test dataframes. This is similar to the feature engineering method in Task 2 where you appended new columns onto an existing dataframe.To do this, use the output of the methods (the list of cluster ids you return) from the corresponding train or test method and add it as a new column named kmeans_cluster_id in the input dataframe, then return the full dataframe.Useful ResourcesINPUTSUse the needed instance variables you set in the __init__ method and the kmeans_train and kmeans_test methods you wrote above to produce the needed output.OUTPUTSA pandas dataframe with the kmeans_cluster_id added as a feature and all other input columns unchanged, for each of the two methods train_add_kmeans_cluster_id_feature and test_add_kmeans_cluster_id_feature.Function Skeletondef train_add_kmeans_cluster_id_feature(self,train_features:pd.DataFrame) -> pd.DataFrame:output_df = pd.DataFrame()return output_df def test_add_kmeans_cluster_id_feature(self,test_features:pd.DataFrame) -> pd.DataFrame:output_df = pd.DataFrame()return output_df        Task 4 Now let’s try a few supervised classification models:We have chosen a few commonly used models for you to use here, but there are many options. In the real world, specific algorithms may fit a specific dataset better than other algorithms.You won’t be doing any hyperparameter tuning yet, so you can better focus on writing the basic code. You will:(Note on feature importance: You should use RFE for determining feature importance of your Logistic Regression model, but do NOT use RFE for your Random Forest or Gradient Boosting models to determine feature importance. Please use their built-in values for this.)Useful Links:Deliverables:Local Test Dataset InformationFor this task the local test dataset we are using is the NATICUSdroid dataset, which contains 86 columns of data related to android permissions used by benign and malicious Android applications released between 2010 and 2019. For more information such as the introductory paper and the Citations/Acknowledgements you can view the dataset site in the UCI ML repository. If you look at the online poster for the paper that the dataset creators wrote from their research, they trained a variety of different models including Random Forest, Logistic Regression and XGBoost and calculated a variety of metrics related to training and detection performance. In this task we will guide you through training ML models and calculating performance metrics to compare the predictive abilities of different models.Instructions:The Task4.py File has function skeletons that you will complete with Python code (mostly using the pandas and scikit-learn libraries).The goal of each of these functions is to give you familiarity with the applied concepts of training a model, using it to score records and calculating performance metrics for it. See information about the function inputs, outputs and skeletons below.Table of contentsModelMetricscalculate_naive_metricsA Naive model is a very simple model/prediction that can help to frame how well a more sophisticated model is doing. At best, such a model has random competence at predicting things. At worst, it’s wrong all the time.Since a naive model is incredibly basic (often a constant or randomly selected result), we can expect that any more sophisticated model that we train should outperform it. If the naive Model beats our trained model, it can mean that additional data (rows or columns) is needed in the dataset to improve our model. It can also mean that the dataset doesn’t have a strong enough signal for the target we want to predict.In this function, you’ll implement a simple model that always predicts a constant (function-provided) number, regardless of the input values. Specifically, you’ll use a given constant integer, provided as the parameter naive_assumption, as the model’s prediction. This means the model will always output this constant value, without considering the actual data. Afterward, you will calculate four metrics—accuracy, recall, precision, and F1-score—for both the training and test datasets.[1] Refer to the resources below.Useful ResourcesINPUTSOUTPUTSA completed ModelMetrics object with a training and test metrics dictionary with each one of the metrics rounded to 4 decimal placesFunction Skeletondef calculate_naive_metrics(train_features:pd.DataFrame, test_features:pd.DataFrame, train_targets:pd.Series, test_targets:pd.Series, naive_assumption:int) -> ModelMetrics:train_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0}test_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0}naive_metrics = ModelMetrics(“Naive”,train_metrics,test_metrics,None)return naive_metricscalculate_logistic_regression_metricsA logistic regression model is a simple and more explainable statistical model that can be used to estimate the probability of an event (log-odds). At a high level, a logistic regression model uses data in the training set to estimate a column’s weight in a linear approximation function. Conceptually this is similar to estimating m for each column in the line formula you probably know well from geometry: y = m*x + b. If you are interested in learning more, you can read up on the math behind how this works. For this project, we are more focused on showing you how to apply these models, so you can simply use a scikit-learn Logistic Regression model in your code.For this task use scikit-learn’s LogisticRegression class and complete the following subtasks:NOTE: Make sure you use the predicted probabilities for roc aucUseful ResourcesINPUTSThe first 4 are similar to the tts function you created in Task 2:OUTPUTSFunction Skeletondef calculate_logistic_regression_metrics(train_features:pd.DataFrame, test_features:pd.DataFrame, train_targets:pd.Series, test_targets:pd.Series, logreg_kwargs) -> tuple[ModelMetrics,LogisticRegression]:model = LogisticRegression()train_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0,“fpr” : 0,“fnr” : 0,“roc_auc” : 0}test_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0,“fpr” : 0,“fnr” : 0,“roc_auc” : 0} log_reg_importance = pd.DataFrame()log_reg_metrics = ModelMetrics(“Logistic Regression”,train_metrics,test_metrics,log_reg_importance) return log_reg_metrics,modelExample of Feature Importance DataFramecalculate_random_forest_metricsA Random Forest model is a more complex model than the naive and Logistic Regression Models you have trained so far. It can still be used to estimate the probability of an event, but achieves this using a different underlying structure: a tree-based model.
Conceptually, this looks a lot like many if/else statements chained together into a “tree”. A Random Forest expands on this and trains different trees with different subsets of the data and starting conditions. It does this to get a better estimate than a single tree would give. For this project, we are more focused on showing you how to apply these models, so you can simply use the scikit-learn Random Forest model in your code.For this task use scikit-learn’s Random Forest Classifier class and complete the following subtasks:NOTE: Make sure you use the predicted probabilities for roc aucUseful ResourcesINPUTSOUTPUTSFunction Skeletondef calculate_random_forest_metrics(train_features:pd.DataFrame, test_features:pd.DataFrame, train_targets:pd.Series, test_targets:pd.Series, rf_kwargs) -> tuple[ModelMetrics,RandomForestClassifier]: model = RandomForestClassifier() train_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0,“fpr” : 0,“fnr” : 0,“roc_auc” : 0} test_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0,“fpr” : 0,“fnr” : 0,“roc_auc” : 0} rf_importance = pd.DataFrame()rf_metrics = ModelMetrics(“Random Forest”,train_metrics,test_metrics,rf_importance) return rf_metrics,modelExample of Feature Importance DataFramecalculate_gradient_boosting_metricsA Gradient Boosted model is more complex than the Naive and Logistic Regression models and similar in structure to the Random Forest model you just trained. A Gradient Boosted model expands on the tree-based model by using its additional trees to predict the errors from the previous tree. For this project, we are more focused on showing you how to apply these models, so you can simply use the scikit-learn Gradient Boosted Model in your code.For this task use scikit-learn’s Gradient Boosting Classifier class and complete the following subtasks:NOTE: Make sure you use the predicted probabilities for roc aucRefer to the Submissions page for details about submitting your work.Useful ResourcesINPUTSOUTPUTSFunction Skeletondef calculate_gradient_boosting_metrics(train_features:pd.DataFrame, test_features:pd.DataFrame, train_targets:pd.Series, test_targets:pd.Series, gb_kwargs) -> tuple[ModelMetrics,GradientBoostingClassifier]:model = GradientBoostingClassifier()train_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0,“fpr” : 0,“fnr” : 0,“roc_auc” : 0}test_metrics = {“accuracy” : 0,“recall” : 0,“precision” : 0,“fscore” : 0,“fpr” : 0,“fnr” : 0,“roc_auc” : 0} gb_importance = pd.DataFrame()gb_metrics = ModelMetrics(“Gradient Boosting”,train_metrics,test_metrics,gb_importance) return gb_metrics,modelExample of Feature Importance DataFrame     Task 5: Model Training and Evaluation:Now that you have written functions for different steps of the model-building process, you will put it all together. You will write code that trains a model with hyperparameters you determine (you should do any tuning locally or in a notebook, i.e., don’t tune your model in gradescope since the autograder will likely timeout).Important: Conduct hyperparameter tuning locally or in a separate notebook. Avoid tuning within Gradescope to prevent autograder timeouts.Develop your own local tests to ensure your code functions correctly before submitting to Gradescope. Do not share these tests with other students.train_model_return_scores (ClaMP Dataset)Instructions (10 points):This function focuses on training a model using the ClaMP dataset and evaluating its performance on a test set.Sample Submission (ClaMP):Function Skeleton (ClaMP):import pandas as pd def train_model_return_scores(train_df, test_df) -> pd.DataFrame:“””Trains a model on the ClaMP training data and returns predicted probabilitiesfor the test data. Args:train_df (pd.DataFrame): ClaMP training data with ‘class’ column.test_df (pd.DataFrame): ClaMP test data without ‘class’ column. Returns:pd.DataFrame: DataFrame with ‘index’ and ‘malware_score’ columns.“””# TODO: Implement the model training and prediction logic as described above.test_scores = pd.DataFrame()  # Replace with your implementationreturn test_scorestrain_model_unsw_return_scores (UNSW-NB15 Dataset)Instructions (10 points):This function is similar to the previous one but uses the UNSW-NB15 dataset.Sample Submission (UNSW-NB15):Function Skeleton (UNSW-NB15):import pandas as pd def train_model_unsw_return_scores(train_df, test_df) -> pd.DataFrame:“””Trains a model on the UNSW-NB15 training data and returns predictedprobabilities for the test data. Args:train_df (pd.DataFrame): UNSW-NB15 training data with ‘class’ column.test_df (pd.DataFrame): UNSW-NB15 test data without ‘class’ column. Returns:pd.DataFrame: DataFrame with ‘index’ and ‘prob_class_1’ columns.“””# TODO: Implement the model training and prediction logic as described above.test_scores = pd.DataFrame()  # Replace with your implementationreturn test_scoresDeliverablesDataset InformationClaMP DatasetUNSW-NB15 Dataset