This project evaluates the performance of models trained on UC Irvine’s Abalone dataset to model the age and sex of new abalone specimens. We compare key metrics for both clustering and classification models; to see if both agree on which attribute is most closely linked to the data features.

The data was sourced UC Irvine’s Machine Learning Repository, and was created in 1994 by the Marine Resources division of Primary Industry & Fisheries in Tasmania, before being donated to the university in 1995.

Total abalone samples: 4176

• 1307 males

• 1342 females

• 1527 infants (unknown sex)

Data features (input):

1. Length (mm)

2. Diameter (mm)

3. Height (mm)

4. Whole weight (grams)

5. Shucked weight (grams)

6. Viscera (grams)

7. Shell (grams)

8. Rings (integer count)

Output features:

1. Sex (classification)

2. Cluster assignment (clustering)

1.Data pre-processing

A. Data was split into 4 separate dataframes:

1. age_data

2. sex_data

3. age_labels

4. sex_labels

B. In age_labels, records labeled as male or female are relabeled as “adult”.

This data preprocessing will allow us to evaluate the models performance on the test data and to compare models trained on sex and age.

C. Both age_data & sex_data are scaled using sci-kit learn’s StandardScaler to prevent feature dominance.

D. A 2D projection of age_data is created so that cluster assignments can be visualized through scatter diagrams later

2.Metric Evaluation

A. Models trained using age & sex labels are evaluated using standard classification metrics

B. Clustering performance is measured via silhouette scores and how well clusters reflect the actual distribution of age and sex labels.

K-Means Clustering

Silhouette Scores

1. 2 Clusters = 0.4695

2. 3 Clusters = 0.3848

Multi-Layer Perceptron Classifier

Age Data

• Accuracy = 80.14%

• Precision Score: 78.12%

• Recall Score: 78.79%

• F1 Score: 78.42%

Sex Data:

• Accuracy = 52.99%

• Precision Score: 52.32%

• Recall Score: 52.86%

• F1 Score: 52.27%

The data seems to be best separated into 2 groups, with age (adult vs. infant) being most accurately modeled using the input features, as opposed to sex (male, female, infant). This is supported by the high accuracy value of the age-based classifier ~80%, the fair silhouette score from the 2-Means clustering analysis, and broad alignment of the cluster point cloud with the actual distribution of the projected data. The physical characteristics of an abalone specimen appear to be a good predictor of whether they are an adult or infant. More fine grain prediction of age may be possible using regression techniques.

1. Lack of cross-validation

   • classification metrics are highly subject to sampling bias

2. Average silhouette scores lack nuance

   1. Silhouette plots could show the degree of definition for each cluster

3. Age modeling can only separate adults from infants

   • More precise age prediction may be possible via binning or use of regression techniques on ring count