Prediction Tasks - Classification

Introduction

Each app is a specialized intelligence to perform a single prediction task, trained from one or more datasets. Each app provides a space for the user to define the task, and to train and compare several models to achieve the goal. The "classification" app, as its name suggests, is tasked with performing the classification task.

Data Preparation for Classification

Apart from the actions committed in previous stages of the "data preparation module", you must also have a target column in the dataset. The target column must be of categorical type.

Example:

Predicting whether a tumour is malignant (M) or benign (B) by looking at its measurements. A sample of the dataset for this task is shown below, with diagnosis being the target column:

	diagnosis	radius_mean	texture_mean	perimeter_mean	area_mean	smoothness_mean	compactness_mean	concavity_mean	concave points_mean	symmetry_mean	fractal_dimension_mean	radius_se	texture_se	perimeter_se	area_se	smoothness_se	compactness_se	concavity_se	concave points_se	symmetry_se	fractal_dimension_se	radius_worst	texture_worst	perimeter_worst	area_worst	smoothness_worst	compactness_worst	concavity_worst	concave points_worst	symmetry_worst	fractal_dimension_worst
0	B	13.62	23.23	87.19	573.2	0.09246	0.06747	0.02974	0.02443	0.1664	0.05801	0.346	1.336	2.066	31.24	0.005868	0.02099	0.02021	0.009064	0.02087	0.002583	15.35	29.09	97.58	729.8	0.1216	0.1517	0.1049	0.07174	0.2642	0.06953
1	B	11.66	17.07	73.7	421	0.07561	0.0363	0.008306	0.01162	0.1671	0.05731	0.3534	0.6724	2.225	26.03	0.006583	0.006991	0.005949	0.006296	0.02216	0.002668	13.28	19.74	83.61	542.5	0.09958	0.06476	0.03046	0.04262	0.2731	0.06825
2	B	12.76	13.37	82.29	504.1	0.08794	0.07948	0.04052	0.02548	0.1601	0.0614	0.3265	0.6594	2.346	25.18	0.006494	0.02768	0.03137	0.01069	0.01731	0.004392	14.19	16.4	92.04	618.8	0.1194	0.2208	0.1769	0.08411	0.2564	0.08253
3	M	21.71	17.25	140.9	1546	0.09384	0.08562	0.1168	0.08465	0.1717	0.05054	1.207	1.051	7.733	224.1	0.005568	0.01112	0.02096	0.01197	0.01263	0.001803	30.75	26.44	199.5	3143	0.1363	0.1628	0.2861	0.182	0.251	0.06494
4	B	12.18	14.08	77.25	461.4	0.07734	0.03212	0.01123	0.005051	0.1673	0.05649	0.2113	0.5996	1.438	15.82	0.005343	0.005767	0.01123	0.005051	0.01977	0.0009502	12.85	16.47	81.6	513.1	0.1001	0.05332	0.04116	0.01852	0.2293	0.06037

The Engine's data preparation combined with automated feature engineering embedded within model templates help you with many tasks but there are some tasks that must be handled by you prior to uploading the data to the platform. See:

Task	Handled by the Engine	Handled by user
Categorical columns "most frequent" imputation	Yes
Categorical columns one hot encoding	Yes
Numerical columns scaling	Yes
Text columns "constant" imputation	Yes
Text columns TF-IDF vectorizer	Yes
Outliers removal		Yes
Cleanup of bad target labels		Yes
Cleanup of duplicated values		Yes
Domain knowledge enabled feature transformations		Yes
Cleanup of data leakages		Yes

This section shows how to build an application of problem type classification on the AI & Analytics Engine

Using the GUI
Using API access through SDK

Using the GUI

Creating an app for a dataset can be accomplished in one of two methods. Either create it from the dashboard by clicking on the "+ New App" rectangle, or, from within the project or dataset page, hover the mouse above the '+' icon in the bottom right corner of the screen and then click on the "New App" button:

First method	Second method

Once any of these options are selected, the "New App" menu will appear. It requires the user to first assign an app name, the related dataset to the app and the name of the target column for that app. The type of the app (classification/regression) is automatically determined (unlike the API where it's explicitly stated) by the type of the target column.

In the second step, the app can be created with the default configuration, which means an 80/20 train/test split or with an advanced configuration that allows the user to change the train/test split size. In the advanced configuration, a value of 90 represents a 90:10 train/test split.

Step 1	Step 2

Once finished, the app is created and the app ID can be found in the browser's address bar.

Using API access through SDK

To access the API functions, you must first authenticate into the platform by

from aiaengine import api

client = api.Client()

Importing `app`

Next you need to import app in order to call functions involved in this module.

from aiaengine.api import app

Creating an app for a dataset

Now you can add a new app by specifying the required parameter values as follows.

create_app_response = client.apps.CreateApp(
    app.CreateAppRequest(
        name='App Name',
        description='What is this app about',
        dataset_id='id_of_dataset_app_is_created_for',
        problem_type='classification',
        target_columns=['target_column'],
        extra_columns={},
        training_data_proportion=0.8
    )
)

Apart from the name and description of the application, you need to inform the platform the dataset id that the application is built for. You also need to specify the problem type, which is 'classification' for this task, and a single target column ('target_column'in the above example). For a classification task, you can keep extra_columns (only used in a forecasting task) as an empty dictionary {}. At last, you can set up a ratio of train-test split using training_data_proportion which indicates the proportion of data used for training over the whole dataset.

app_id = create_app_response.id

Once created, an application is assigned with a unique id, which is frequently used in the related functions.