General ML Pipeline
Machine Learning (ML) / Artificial Intelligent (AI) pipeline consists of multiple steps for training and using a model including data collection, data preprocessing, model training, model prediction, and model evaluation.
Beta version: The platform is in a development stage of the final version, which may be less stable than usual. The efficiency of platform access and usage might be limited. For example, the platform might crash, some features might not work properly, or some data might be lost.
.png?alt=media&token=59054d59-00b2-4a84-af06-d557f8918c0b)
¶ Data collection
These are basic steps for collecting a dataset:
- Before developing a model, you have to consider its purpose; then, think about the data needed for achieving the purpose. Then, you have to gather your existing and external data sources as required.
- Data exploration is a step after you collect data. You need to fully understand the characteristics of the data by reviewing the type, distribution, and relationship between data, trends, etc.
¶ Data preprocessing
Data preprocessing is a step to transform and clean data into the format suitable for a model. Preprocessing relies on the type of data, which is required differently depending on a machine learning algorithm. Please follow the guidelines below:
- The data used in a development period should be representative of the future, or real data in production in terms of attributes, distribution, and domain.
- Data leakage is the situation when the training data contains information that you are predicting, which can bring about inaccurate model performance. This can be solved by splitting your dataset in an appropriate way.
- Data cleansing is a step to deal with inconsistencies, outliers, anomalies, missing data, and other issues. All of these issues could affect model performance. This step would allow the model to focus on meaningful patterns in the data.
- The quality of data is measured by the consistency and accuracy of labeled data. Consistency means similar data should be labeled with the same label. Accuracy means the correctness of labeling is close to true conditions. Labelers need to control labeling across a dataset by using one standard guideline and constantly monitoring labels.
- Data structuring is a step to ensure that your data is formatted in the form that fits the selected algorithm.
¶ Train/Test split
The data split procedure is used to evaluate a predictive or machine learning model. Generally, the data is separated into 2 sets: training and testing data.
The purpose of each set is as follows:
- Training data – for building up a machine learning model.
- Testing data– for estimating the performance of a model after training to confirm that the model is trained effectively by using an unseen dataset
The simplest method for splitting data is to shuffle the dataset and split it into an arbitrary percentage, for example, 80% training, and 20% testing. The shuffling is carried out to make sure that each split has similar characteristics. Although this is generally fine for most datasets, some datasets need more caution in splitting. The number of samples in training and test sets should be large and diverse enough to ensure that the datasets are representative of the original dataset.
¶ Model training
After splitting data into training and testing sets, a model is trained by the algorithm using the training data. There are various algorithms depending on the expected purpose and outcome. Each algorithm has its own following mandatory parameters to config. Our ACP provides automated hyperparameters tuning in the training process. This would find the best parameters that give a model the best results.
¶ Model prediction
After training a model, you will get the final model to make predictions on new data. Then, you need to give input data to the model to predict the expected output.
¶ Model evaluation
Model performances can be evaluated by predicting the test data, which is unseen data. To calculate the performance measure, you can compare the prediction’s results with the actual results with evaluation metrics or statistic techniques that suit each algorithm.
- About evaluation metrics, please go to the section