Pipelines, Pandas and Flask the solution for parsing string user Inputs to your machine learning model.

Here is a situation, I have a trained ML model and I want to parse both strings and numbers as features to my model and the reason is; imaginatively speaking, I hope the Hotel manager-who doesn’t know much about tech will be able to use my trained model by selecting their concerns from the html file.

Assuming that you have knowledge of ML modelling then this article promises to help you with the basics of parsing strings and numbers to you machine learning model for prediction.

Therefore, in this tutorial we shall look at the following:

1.       Building a Machine learning model that accepts strings and passthrough’ numbers for predictions.

2.       Building an interactive user interface to parse form inputs to the trained model to make predictions.

3.       Using Flask to build a REST API to deploy the model locally.

 

By ensuring these three concerns are met then any authorized user can interact with the model from the local machines or cloud.

What will our model do? Our model will be able to predict the type of food that will be ordered (DV-Dependent Variable) based on Age, Nationality, Gender, Dessert, Juice (IV-independent Variable). The model is based on this dataset from Kaggle, I had to modify it a bit-added some more classes to our target variable.

Let’s go for commercial break please!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

The main purpose of this Tutorial is demonstrating how your model can accept string user inputs especially, transform them and then make Predictions.

Step 1. Building a machine learning model.

There are 5 steps that I feel will be key in this section:

A. We start by reading the csv file and then using Matplotlib we plot out the relevant relationships to help visualize variables better. This process will help to determine which features will influence your model performance considering your target variable.

 

In this dataset there is only one continuous variable and that is Age, my target variable being  a categorical data-Food, then you might notice that the distribution of the observations might inspire the use of modelling algorithms such us logistic regression, Ensemble techniques, Naïve Bayes and Decision trees as opposed to say Linear regression. However, it is important that we try different modeling algorithms as we perform parameter tuning with the aim of getting the best optimized model using the R2 score.

B. Dividing the dataset into training and Testing and storing the categorical features in a variable for preprocessing.

The focus here after splitting the dataset into train and test is to store the categorical features in a variable that will be passed to the column transformer for imputing and encoding.

      C. Pipelines and Column Transformers.

We are going to create a pipeline that contains preprocessing steps and use it within the column transformer to process the categorical columns that we are representing as “xcols”. We shall impute missing values and one-hot encode them with the parameter “add_indicator” and “handle_unknown” set to “False” and “ ignore” respectively because we don’t need to encode the automatically generated columns holding Boolean values indicating where the transformation occurred and if case any other column is generated after encoding we are simply just ignore it. We transform all categorical features/columns and we pass “Age” column since it’s a number already.

The difference between how pipelines and columns transformers preprocess features is with a column transformer different features are transformed using different or same preprocessing steps but, in a pipeline, multiple transformations are applied on the same column(s).

Then lastly, we transform the target variable which in this case Is “Food” using the label encoder.

      D. Model selection, fitting and predicting.

DecisionTreeRegressor model gave out the best coefficient of determination even without any feature engineering applied. According to scikit-learn 0.24.1 pipelines offer the same API as a regular estimator meaning it has functions for training and making predictions.

     E. Using Pandas Data Frame to create out of sample data.

This section and section D form the foundation of this Tutorial.

We expect to capture user input and parse it to our model to make predictions, therefore, we need to create an unseen observation and pass' it to our model. I have done this by using a dictionary then converted the dictionary to pandas data frame with the index parameter set to “289” because my rows end at “288”.

The index parameter accepts any value from zero it will not affect the preprocessing.

Step 2. Developing a bootstrap user interface for user interaction.

Basically, the import bit here is pointing your form data to the correct decorator url function using “url_for()” method and lastly, rendering the output of the model is facilitated by Jinja templates render_template method with “{{_}}” delimiters put in the html page for model output.

Jinja template engine is simply Flask’s solution for generating html within python securely and automatically.

Step 3. Creating a Rest API using flask to serve user requests to the model.

To put it in a very simple way Flask is the “engine” upon which the entire vehicle is built on. You can safely say that both the Html page and the model runs on Flask.

 So, we use the “request. form” which means that the request object has an attribute called form that accesses form data using the keys(xcols) and store it in the “qwert” variables. Then, we use these variables to from a dictionary called “damp” afterwards, we convert the dictionary into a panda’s data frame called “test” consequently, parsing the “test” sample to the predictive model. Finally, storing the outcome of the prediction in a variable called “result”. I have written a for loop because the model outputs the value in a list hence, once I get the float value I run it through the conditions in the if_statements that I created by comparing the transformed labels to the classes generated from the categorical target “Food".

Summary.

Basically speaking, there are only two concepts that you need to wrap your head around:

1.     Use pipelines to chain the pre-processing and the model building steps together, we do this because the user inputs are strings so we will need to transform them before we predict them.

2.      Since user inputs are out of sample features then we need to create a dictionary then convert it into a pandas data frame which we shall parse to our model for predictions.

 

Thank you!