Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/majkeloess/fashionmnist

The project aims to use modern data science tools
https://github.com/majkeloess/fashionmnist

Last synced: about 1 month ago
JSON representation

The project aims to use modern data science tools

Awesome Lists containing this project

README 

        
# fashionMnist

The project provides a foundation for understanding the practical challenges and nuances associated with real-world data analysis tasks. The success of the classification model suggests the importance of supervised learning techniques, especially in scenarios where distinct class labels are available. 

## [Kaggle link](https://www.kaggle.com/code/majkeloess/fashion-mnist-dataset-analysis-ds-winter-project?scriptVersionId=159907141)



# Dataset Description

The `Fashion-MNIST` dataset serves as the cornerstone of our exploration into practical data analysis, offering a nuanced departure from the traditional MNIST dataset. Crafted as a response to the need for more challenging validation datasets, Fashion-MNIST introduces a diverse collection of grayscale images, each representing various clothing articles from Zalando's inventory. With ten distinct classes, including T-shirts, trousers, dresses, and more, the dataset expands the horizons of image classification. Notably, the images are formatted as **28x28** grayscale representations, presenting a unique challenge and opportunity for algorithms to discern meaningful features from relatively small images.



Examples and labels: 



![image](https://github.com/majkeloess/fashionMnist/assets/118011581/1416faf9-b53b-466e-926b-c4f674df0fd5) 



| Label | Description |

| --- | --- |

| 0 | T-shirt/top |

| 1 | Trouser |

| 2 | Pullover |

| 3 | Dress |

| 4 | Coat |

| 5 | Sandal |

| 6 | Shirt |

| 7 | Sneaker |

| 8 | Bag |

| 9 | Ankle boot |



# Dimensionality Reduction 



One prominent technique for dimensionality reduction is `Principal Component Analysis (PCA)`. PCA identifies the principal components linear combinations of the original features capturing the maximum variance within the data. By selecting a subset of these components, we can achieve a reduced-dimensional representation that retains a significant portion of the dataset's variability.



Initially, PCA is fitted to the training data without specifying the number of components. The cumulative explained variance is then calculated, and the number of components needed to retain at least 95% of the variance is determined. Subsequently, a new PCA object is created with the selected number of components, and the transformation is applied to both the training and test datasets.



After applying Principal Component Analysis (PCA) to reduce the dimensionality of the dataset, an intriguing aspect is the ability to reverse this transformation and visualize the reconstructed images.



![image](https://github.com/majkeloess/fashionMnist/assets/118011581/2a2a98c6-f63a-4256-9d50-99072dcde9dd)



This visualization provides a tangible glimpse into how the dimensionality-reduced representation impacts the appearance of the original images. It highlights the ability of PCA to capture essential features while reducing the dataset's dimensionality, enabling a more efficient representation for subsequent tasks.



# Clustering using K-Means Algorithm



`K-Means` is a partitioning algorithm that divides a dataset into 'k' clusters. It operates iteratively, assigning data points to clusters and updating cluster centroids until convergence. The number of clusters, 'k,' is a crucial parameter and often needs to be specified a priori.



In the context of the Fashion-MNIST dataset, K-Means clustering proves to be a suitable choice for several reasons. The dataset consists of grayscale images of fashion items, and K-Means is particularly effective when dealing with relatively well-defined and spherical clusters in the feature space. Given the nature of clothing items, such as T-shirts, trousers, and dresses, K-Means can potentially capture inherent patterns in the pixel space.



The algorithm's simplicity, scalability, and efficiency make it well-suited for large datasets like Fashion-MNIST. Additionally, the interpretability of K-Means clusters can provide insights into how different fashion items group together, potentially revealing interesting relationships and similarities. The K-Means clustering model is initialized with 10 clusters, aligning with the number of unique labels in the dataset. This choice is made with the expectation that the clusters may align with the distinct classes of fashion items, allowing for an informative analysis of the dataset's structure.



Upon applying K-Means clustering to the Fashion-MNIST dataset, an analysis of the clustering accuracy yields a result of `0.56`. This metric is obtained by comparing the assigned cluster labels to the true labels of the dataset. A cluster analysis accuracy of 0.56 indicates that approximately `56%` of the instances were correctly assigned to their corresponding clusters.



While cluster analysis accuracy provides a quantitative measure of the algorithm's performance, it's important to interpret this result in the context of the dataset's complexity and inherent challenges. Fashion-MNIST, with its diverse array of fashion items and intricate patterns, presents a non-trivial clustering task.



# Classification using K-Nearest Neighbors



`K-Nearest Neighbors (KNN)` is a versatile and intuitive classification algorithm that makes predictions based on the majority class of its 'k' nearest neighbors in the feature space. KNN is particularly effective for datasets with discernible decision boundaries and instances of local patterns.



In the progression of this data science project, the inclusion of classification techniques becomes paramount. Classification allows us to go beyond clustering and assign specific labels to data instances, enabling the model to make predictions on unseen data. In the context of Fashion-MNIST, the goal is to train a classifier that can accurately categorize images of clothing items into their respective classes. KNN classifier is implemented using the reduced-dimensional data obtained after applying Principal Component Analysis (PCA) to the Fashion-MNIST dataset. This approach leverages the insights gained from dimensionality reduction to enhance the efficiency and interpretability of the classification model.



The classification report provides a detailed evaluation of the KNN classifier's performance on the Fashion-MNIST test set. The table below summarizes key metrics such as **precision**, **recall**, and **F1-score** for each class, along with overall accuracy: 



## Classification table



| Class | Precision | Recall | F1-score | Support |

|-------|-----------|--------|----------|---------|

| 0     | 0.75      | 0.85   | 0.80     | 1232    |

| 1     | 0.99      | 0.97   | 0.98     | 1174    |

| 2     | 0.73      | 0.80   | 0.76     | 1200    |

| 3     | 0.89      | 0.87   | 0.88     | 1242    |

| 4     | 0.77      | 0.73   | 0.75     | 1185    |

| 5     | 0.99      | 0.89   | 0.94     | 1141    |

| 6     | 0.66      | 0.58   | 0.62     | 1243    |

| 7     | 0.92      | 0.97   | 0.94     | 1224    |

| 8     | 0.98      | 0.97   | 0.98     | 1149    |

| 9     | 0.93      | 0.97   | 0.95     | 1210    | 



| Metric        | Precision | Recall | F1-score | Support |

|---------------|-----------|--------|----------|---------|

| Accuracy      |           |        | 0.86     | 12000   |

| Macro avg     | 0.86      | 0.86   | 0.86     | 12000   |

| Weighted avg  | 0.86      | 0.86   | 0.86     | 12000   |



## Classification Report



- **Precision:** The proportion of true positive predictions among all positive predictions.

- **Recall:** The proportion of true positive predictions among all actual positives.

- **F1-score:** The harmonic mean of precision and recall, providing a balance between the two metrics.

- **Support:** The number of actual occurrences of the class in the test set.



The table is organized by class (represented by numbers 0 to 9 in the Fashion-MNIST dataset). The "accuracy" row indicates the overall accuracy of the model on the entire test set. The "macro avg" row represents the average performance across all classes, while the "weighted avg" row considers the proportion of instances for each class.



## Classification Heatmap 

![image](https://github.com/majkeloess/fashionMnist/assets/118011581/747f917a-50b6-493c-8883-8c6d25f9ff91) 



The classification report provides a detailed evaluation of the KNN classifier's performance on the Fashion-MNIST test set. The **confusion matrix**, a tabular representation of true positive, true negative, false positive, and false negative predictions, offers further insights into the classifier's behavior. The confusion matrix allows us to assess how well the classifier performs for each class, revealing patterns of misclassifications and highlighting areas for improvement. True positives, true negatives, false positives, and false negatives contribute to the overall understanding of the model's strengths and weaknesses. The heatmap visualization of the confusion matrix enhances interpretability by providing a visual representation of the classification performance. Intensity of color indicates the number of instances falling into each category. 

 

The overall accuracy of the KNN classifier is reported to be `0.86`, indicating that approximately `86%` of the instances in the Fashion-MNIST test set were correctly classified. This metric serves as a general measure of the model's effectiveness in making accurate predictions across all classes. 



# Conclusion

In comparing the results of clustering and classification, the K-Means clustering algorithm yielded an accuracy of `0.56`, indicating a moderate success in grouping similar instances within the Fashion-MNIST dataset. However, it's important to note that clustering operates in an unsupervised manner, and the clusters may not align precisely with the true class labels. On the other hand, the K-Nearest Neighbors (KNN) classifier achieved an overall accuracy of `0.86`, showcasing a substantial success in categorizing clothing items based on reduced-dimensional features obtained through Principal Component Analysis (PCA). The supervised nature of classification, with explicit class labels, allows for a more direct evaluation of model performance on unseen data. The higher accuracy in classification suggests the effectiveness of supervised learning techniques in capturing the diverse patterns and styles present in the Fashion-MNIST dataset.



# References



- [Fashion MNIST GitHub Repository](https://github.com/zalandoresearch/fashion-mnist)

- [Fashion MNIST Kaggle Dataset](https://www.kaggle.com/datasets/zalando-research/fashionmnist)

- [YouTube: "MNIST Dataset Explained"](https://www.youtube.com/watch?v=3bHxOYczJn0)

- [YouTube: "Introduction to MNIST"](https://www.youtube.com/watch?v=TKfTKA5EmmY&t=302s)

- [Medium: "K-Means Clustering for Image Classification"](https://medium.com/@joel_34096/k-means-clustering-for-image-classification-a648f28bdc47)

- Data Science Lectures 



Feel free to check out these resources for more information on Fashion MNIST dataset, tutorials, and related topics.