Progetto svolto in collaborazione con Alessandro Vavalà e Giuseppe Milazzo

Data Mining: Advanced Topics and Applications

Tasks of the project:

• Explore and prepare the dataset. You are allowed to take inspiration from the associated GitHub repository and figure out your personal research perspective (from choosing a subset of variables to the class to predict…). You are welcome in creating new variables and performing all the pre-processing steps the dataset needs.
• Define one or more (simple) classification tasks and solve it with Decision Tree and KNN. You decide the target variable.
• Identify the top 1% outliers: adopt at least three different methods from different families (e.g., density-based, angle-based… ) and compare the results. Deal with the outliers by removing them from the dataset or by treating the anomalous variables as missing values and employing replacement techniques. In this second case, you should check that the outliers are not outliers anymore. Justify your choices in every step.
• Analyze the value distribution of the class to predict with respect to point 2; if it is unbalanced leave it as it is, otherwise turn the dataset into an imbalanced version (e.g., 96% - 4%, for binary classification). Then solve the classification task using the Decision Tree or the KNN by adopting various techniques of imbalanced learning.
• Draw your conclusions about the techniques adopted in this analysis.

• Solve the classification task defined in Module 1 (or define new ones) with the other classification methods analyzed during the course: Naive Bayes Classifier, Logistic Regression, Rule-based Classifiers, Support Vector Machines, Neural Networks, Ensemble Methods and evaluate each classifier with the techniques presented in Module 1 (accuracy, precision, recall, F1-score, ROC curve). Perform hyper-parameter tuning phases and justify your choices.
• Besides the numerical evaluation draw your conclusions about the various classifiers, e.g. for Neural Networks: what are the parameter sets or the convergence criteria which avoid overfitting? For Ensemble classifiers how the number of base models impacts the classification performance? For any classifier which is the minimum amount of data required to guarantee an acceptable level of performance? Is this level the same for any classifier? What is revealing the feature importance of Random Forests?
• Select two continuous attributes, define a regression problem and try to solve it using different techniques reporting various evaluation measures. Plot the two-dimensional dataset. Then generalize to multiple linear regression and observe how the performance varies.

• Select the feature(s) you prefer and use it (them) as a time series. You can use the temporal information provided by the authors’ datasets, but you are also welcome in exploring the .mp3 files to build your own dataset of time series according to your purposes. You should prepare a dataset on which you can run time series clustering; motif/anomaly discovery and classification.
• On the dataset created, compute clustering based on Euclidean/Manhattan and DTW distances and compare the results. To perform the clustering you can choose among different distance functions and clustering algorithms. Remember that you can reduce the dimensionality through approximation. Analyze the clusters and highlight similarities and differences.
• Analyze the dataset for finding motifs and/or anomalies. Visualize and discuss them and their relationship with other features.
• Solve the classification task on the time series dataset(s) and evaluate each result. In particular, you should use shapelet-based classifiers. Analyze the shapelets retrieved and discuss if there are any similarities/differences with motifs and/or shapelets.

• Sequential Pattern Mining: Convert the time series into a discrete format (e.g., by using SAX) and extract the most frequent sequential patterns (of at least length 3/4) using different values of support, then discuss the most interesting sequences.
• Advanced Clustering: On a dataset already prepared for one of the previous tasks in Module 1 or Module 2, run at least one clustering algorithm presented in the advanced clustering lectures (e.g. X-Means, Bisecting K-Means, OPTICS). Discuss the results that you find analyzing the clusters and reporting external validation measures (e.g SSE, silhouette).
• Transactional Clustering: By using categorical features, or by turning a dataset with continuous variables into a dataset with categorical variables (e.g. by using binning), run at least one clustering algorithm presented in the transactional clustering lectures (e.g. K-Modes, ROCK). Discuss the results that you find analyzing the clusters and reporting external validation measures (e.g SSE, silhouette).

• Try to use one or more explanation methods (e.g., LIME, LORE, SHAP, etc.) to illustrate the reasons for the classification in one of the steps of the previous tasks.