Contents<\/h3>\n\n- Application type<\/a>.<\/li>\n
- Data set<\/a>.<\/li>\n
- Neural network<\/a>.<\/li>\n
- Training strategy<\/a>.<\/li>\n
- Model selection<\/a>.<\/li>\n
- Testing analysis<\/a>.<\/li>\n
- Model deployment<\/a>.<\/li>\n<\/ol>\n<\/section>\n
\n1. Application type<\/h2>\n
This is an approximation<\/a> project since the variable to be predicted is continuous (wine quality).<\/p>\nThe primary objective is to model the quality of a wine as a function of its characteristics.<\/p>\n<\/section>\n\n2. Data set<\/h2>\n
The data file wine_quality.csv<\/a> contains a total of 1599 rows and 12 columns.<\/p>\nVariables<\/h3>\n
The data set contains the following variables<\/a>:<\/p>\n\n- fixed_acidity<\/strong> \u2013 Non-volatile acids in wine (e.g., tartaric acid).<\/li>\n
- volatile_acidity<\/strong> \u2013 Volatile acids that can affect aroma (e.g., acetic acid).<\/li>\n
- citric_acid<\/strong> \u2013 Citric acid content, which can add freshness.<\/li>\n
- residual_sugar<\/strong> \u2013 Sugar remaining after fermentation.<\/li>\n
- chlorides<\/strong> \u2013 Salt content (g\/L of sodium chloride).<\/li>\n
- free_sulfur_dioxide<\/strong> \u2013 Free SO\u2082, prevents microbial growth.<\/li>\n
- total_sulfur_dioxide<\/strong> \u2013 Total SO\u2082 (bound + free).<\/li>\n
- density<\/strong> \u2013 Density of the wine (g\/cm\u00b3).<\/li>\n
- pH<\/strong> \u2013 Measure of acidity (scale of 0\u201314).<\/li>\n
- sulfates<\/strong> \u2013 Sulfate content, contributes to SO\u2082 levels.<\/li>\n
- alcohol<\/strong> \u2013 Alcohol percentage by volume.<\/li>\n<\/ul>\n
Instances<\/h3>\n
On the other hand, the instances are randomly divided into training (60%), selection (20%), and testing (20%) subsets.<\/p>\n
Distributions<\/h3>\n
Once the dataset page is edited, we run several related tasks to verify the quality of the provided information.<\/p>\n
The following figure\u00a0illustrates the\u00a0distribution<\/a> of the quality variable<\/span>.<\/p>\n![](\"https:\/\/www.neuraldesigner.com\/images\/wine-quality-quality-distribution.webp\")
<\/p>\n
The target variable is unbalanced, with most scores clustered around 5 and 6 and very few near 0 or 10, which can degrade model performance.<\/p>\n
Input-target correlations<\/h3>\n
The following chart illustrates the correlations between the physicochemical characteristics and the quality.<\/p>\n
![](\"https:\/\/www.neuraldesigner.com\/images\/wine-quality-inputs-targets-correlations.webp\")
<\/p>\n
As you can see, alcohol has the highest positive correlation with quality, and volatile acidity has the highest negative correlation.<\/p>\n<\/section>\n\n3. Neural network<\/h2>\n
The second step is to set up a neural network <\/a><\/span>that learns the relationship between the wine\u2019s laboratory measurements and its quality score.<\/p>\nThe network has ten input variables (the wine\u2019s chemical properties) and one output (the predicted quality).<\/p>\n
Scaling layer<\/h3>\n
The scaling layer<\/a> normalizes the input data so that all variables (such as acidity, alcohol, or pH) are comparable.<\/p>\nHidden dense layer<\/h3>\n
The first dense layer, or hidden layer, consists of three neurons that combine the ten input variables to capture complex patterns, much like an oenologist evaluating acidity, aroma, and body together.<\/p>\n
Output dense layer<\/h3>\n
The second dense layer, or output layer, has one neuron that produces the final predicted quality score, just as the oenologist gives an overall judgment of the wine.<\/p>\n
Unscaling layer<\/h3>\n
Next, the unscaling layer<\/a> unnormalizes the output, so the prediction is expressed in the same quality range as the data.<\/p>\nBounding layer<\/h3>\n
Finally, the bounding layer ensures that the predictions fall between 1 and 10, which aligns with the way wine quality is evaluated.<\/p>\n
Neural network graph<\/h3>\n
You can represent the neural network for this example in the following diagram:<\/p>\n
![](\"https:\/\/www.neuraldesigner.com\/images\/wine-quality-neural-network.webp\")
<\/p>\n
In total, the network contains 37 parameters, which are adjusted during training to provide accurate predictions of wine quality.<\/p>\n<\/section>\n
4. Training strategy<\/h2>\n
The next step is\u00a0to select an appropriate\u00a0training strategy<\/a> that defines<\/span>\u00a0what the neural network will learn. A general training strategy for approximation consists of two components:<\/p>\n\n- A loss index.<\/li>\n
- An optimization algorithm.<\/li>\n<\/ul>\n
Loss index<\/h3>\n
The loss index<\/a> chosen for this problem is the normalized squared error<\/a> between the neural network’s outputs and the dataset’s targets. On the other hand, we apply L2 regularization with a weak weight in this case.<\/p>\nOptimization algorithm<\/h3>\n
The selected optimization algorithm<\/a> is the quasi-Newton method<\/a>.<\/p>\nTraining<\/h3>\n
![](\"https:\/\/www.neuraldesigner.com\/images\/wine-quality-training.webp\")
<\/p>\n
The key training result is the final selection error, which reflects the network\u2019s generalization ability.<\/p>\n
In this case, the final selection error is 0.678 NSE.<\/p>\n\n5. Model selection<\/h2>\n
The objective of model selection<\/a> is to find the network architecture with the best generalization properties, which minimizes the error on the selected instances<\/a> of the data set.<\/p>\nMore specifically, we\u00a0aim to develop a neural network with a selection error of less than\u00a00.678 NSE<\/strong>, which is the current value we have <\/span>achieved.<\/p>\nOrder selection<\/a> algorithms train several network architectures with different numbers of neurons and select the one with the smallest selection error.<\/p>\nThe incremental order<\/a> method starts with a few neurons and increases the complexity at each iteration. The following chart shows the training error (blue) and the selection error (orange) as a function of the number of neurons.<\/p>\n![](\"https:\/\/www.neuraldesigner.com\/images\/wine-quality-model-selection.webp\")
<\/p>\n
The final selection error achieved is 0.661<\/b> for an optimal number of neurons of 1.<\/p>\n
![](\"https:\/\/www.neuraldesigner.com\/images\/wine-quality-final-architecture.webp\")
<\/p>\n
The graph above represents the architecture of the final neural network.<\/p>\n<\/section>\n\n6. Testing analysis<\/h2>\n
A standard method for testing the prediction capabilities is to compare the neural network outputs with an independent dataset.<\/p>\n
Goodnes-of-fit<\/h3>\n
1. Application type<\/h2>\n
This is an approximation<\/a> project since the variable to be predicted is continuous (wine quality).<\/p>\n The primary objective is to model the quality of a wine as a function of its characteristics.<\/p>\n<\/section>\n The data file wine_quality.csv<\/a> contains a total of 1599 rows and 12 columns.<\/p>\n The data set contains the following variables<\/a>:<\/p>\n On the other hand, the instances are randomly divided into training (60%), selection (20%), and testing (20%) subsets.<\/p>\n Once the dataset page is edited, we run several related tasks to verify the quality of the provided information.<\/p>\n The following figure\u00a0illustrates the\u00a0distribution<\/a> of the quality variable<\/span>.<\/p>\n The target variable is unbalanced, with most scores clustered around 5 and 6 and very few near 0 or 10, which can degrade model performance.<\/p>\n The following chart illustrates the correlations between the physicochemical characteristics and the quality.<\/p>\n As you can see, alcohol has the highest positive correlation with quality, and volatile acidity has the highest negative correlation.<\/p>\n<\/section>\n The second step is to set up a neural network <\/a><\/span>that learns the relationship between the wine\u2019s laboratory measurements and its quality score.<\/p>\n The network has ten input variables (the wine\u2019s chemical properties) and one output (the predicted quality).<\/p>\n The scaling layer<\/a> normalizes the input data so that all variables (such as acidity, alcohol, or pH) are comparable.<\/p>\n The first dense layer, or hidden layer, consists of three neurons that combine the ten input variables to capture complex patterns, much like an oenologist evaluating acidity, aroma, and body together.<\/p>\n The second dense layer, or output layer, has one neuron that produces the final predicted quality score, just as the oenologist gives an overall judgment of the wine.<\/p>\n Next, the unscaling layer<\/a> unnormalizes the output, so the prediction is expressed in the same quality range as the data.<\/p>\n Finally, the bounding layer ensures that the predictions fall between 1 and 10, which aligns with the way wine quality is evaluated.<\/p>\n You can represent the neural network for this example in the following diagram:<\/p>\n In total, the network contains 37 parameters, which are adjusted during training to provide accurate predictions of wine quality.<\/p>\n<\/section>\n The next step is\u00a0to select an appropriate\u00a0training strategy<\/a> that defines<\/span>\u00a0what the neural network will learn. A general training strategy for approximation consists of two components:<\/p>\n The loss index<\/a> chosen for this problem is the normalized squared error<\/a> between the neural network’s outputs and the dataset’s targets. On the other hand, we apply L2 regularization with a weak weight in this case.<\/p>\n The selected optimization algorithm<\/a> is the quasi-Newton method<\/a>.<\/p>\n The key training result is the final selection error, which reflects the network\u2019s generalization ability.<\/p>\n In this case, the final selection error is 0.678 NSE.<\/p>\n The objective of model selection<\/a> is to find the network architecture with the best generalization properties, which minimizes the error on the selected instances<\/a> of the data set.<\/p>\n More specifically, we\u00a0aim to develop a neural network with a selection error of less than\u00a00.678 NSE<\/strong>, which is the current value we have <\/span>achieved.<\/p>\n Order selection<\/a> algorithms train several network architectures with different numbers of neurons and select the one with the smallest selection error.<\/p>\n The incremental order<\/a> method starts with a few neurons and increases the complexity at each iteration. The following chart shows the training error (blue) and the selection error (orange) as a function of the number of neurons.<\/p>\n The final selection error achieved is 0.661<\/b> for an optimal number of neurons of 1.<\/p>\n The graph above represents the architecture of the final neural network.<\/p>\n<\/section>\n A standard method for testing the prediction capabilities is to compare the neural network outputs with an independent dataset.<\/p>\n2. Data set<\/h2>\n
Variables<\/h3>\n
\n
Instances<\/h3>\n
Distributions<\/h3>\n
<\/p>\nInput-target correlations<\/h3>\n
<\/p>\n3. Neural network<\/h2>\n
Scaling layer<\/h3>\n
Hidden dense layer<\/h3>\n
Output dense layer<\/h3>\n
Unscaling layer<\/h3>\n
Bounding layer<\/h3>\n
Neural network graph<\/h3>\n
<\/p>\n4. Training strategy<\/h2>\n
\n
Loss index<\/h3>\n
Optimization algorithm<\/h3>\n
Training<\/h3>\n
<\/p>\n5. Model selection<\/h2>\n
<\/p>\n<\/p>\n6. Testing analysis<\/h2>\n
Goodnes-of-fit<\/h3>\n
![](\"https:\/\/www.neuraldesigner.com\/images\/wine-quality-linear-regression-analysis.webp\")
<\/p>\n