Model training best practices
There are things you can do to improve the performance of Neptune ML models.
Choose the right node property
Not all the properties in your graph may be meaningful or relevant to your machine learning tasks. Any irrelevant properties should be excluded during data export.
Here are some best practices:
Use domain experts to help evaluate the importance of features and the feasibility of using them for predictions.
Remove the features that you determine are redundant or irrelevant to reduce noise in the data and unimportant correlations.
Iterate as you build your model. Adjust the features, feature combinations, and tuning objectives as you go along.
Feature Processing in the Amazon Machine Learning Developer Guide provides additional guidelines for feature processing that are relevant to Neptune ML.
Handle outlier data points
An outlier is a data point that is significantly different from the remaining data. Data outliers can spoil or mislead the training process, resulting in longer training time or less accurate models. Unless they are truly important, you should eliminate outliers before exporting the data.
Remove duplicate nodes and edges
Graphs stored in Neptune may have duplicate nodes or edges. These redundant elements will introduce noise for ML model training. Eliminate duplicate nodes or edges before exporting the data.
Tune the graph structure
When the graph is exported, you can change the way features are processed and how the graph is constructed, to improve the model performance.
Here are some best practices:
When an edge property has the meaning of categories of edges, it is worth turning it into edge types in some cases.
The default normalization policy used for a numerical property is
min-max
, but in some cases other normalization policies work better. You can preprocess the property and change the normalization policy as explained in Elements of a model-HPO-configuration.json file.The export process automatically generates feature types based on property types. For example, it treats
String
properties as categorical features andFloat
andInt
properties as numerical features. If you need to, you can modify the feature type after export (see Elements of a model-HPO-configuration.json file).
Tune the hyperparameter ranges and defaults
The data-processing operation infers hyperparameter configuration ranges from the graph. If the generated model hyperparameter ranges and defaults don't work well for your graph data, you can edit the HPO configuration file to create your own hyperparameter tuning strategy.
Here are some best practices:
When the graph goes large, the default hidden dimension size may not be large enough to contain all the information. You can change the
num-hidden
hyperparameter to control the hidden dimension size.-
For knowledge graph embedding (KGE) models, you may want to change the specific model being used according to your graph structure and budget.
TrainsE
models have difficulty in dealing with one-to-many (1-N), many-to-one (N-1), and many-to-many (N-N) relations.DistMult
models have difficulty in dealing with symmetric relations.RotatE
is good at modeling all kinds of relations but is more expensive thanTrainsE
andDistMult
during training. In some cases, when both node identification and node feature information are important, you should use
`concat-node-embed`
to tell the Neptune ML model to get the initial representation of a node by concatenating its features with its initial embeddings.When you are getting reasonably good performance over some hyperparameters, you can adjust the hyperparameter search space according to those results.
Early stopping of the model training process in Neptune ML
Early stopping can significantly reduce the model-training run time and associated costs without degrading model performance. It also prevent the model from overfitting on the training data.
Early stopping depends on regular measurements of validation-set performance. Initially, performance improves as training proceeds, but when the model starts overfitting, it starts to decline again. The early stopping feature identifies the point at which the model starts overfitting and halts model training at that point.
Neptune ML monitors the validation metric calls and compares the most recent validation
metric to the average of validation metrics over the last n
evaluations, where n
is a number set using the
window-for-early-stop
parameter. As soon as the validation metric is worse than
that average, Neptune ML stops the model training and saves the best model so far.
You can control early stopping using the following parameters:
-
window-for-early-stop
– The value of this parameter is an integer that specifies the number of recent validation scores to average when deciding on an early stop. The default value is3
. -
enable-early-stop
– Use this Boolean parameter to turn off the early stop feature. By default, its value istrue
.
Early stopping of the HPO process in Neptune ML
The early stop feature in Neptune ML also stops training jobs that are not performing well compared to other training jobs, using the SageMaker HPO warm-start feature. This too can reduce costs and improve the quality of HPO.
See Run
a warm start hyperparameter tuning job
Warm start provides the ability to pass information learned from previous training jobs to subsequent training jobs and provides two distinct benefits:
First, the results of previous training jobs are used to select good combinations of hyperparameters to search over in the new tuning job.
Second, it allows early stopping to access more model runs, which reduces tuning time.
This feature is enabled automatically in Neptune ML, and allows you strike a balance between model training time and performance. If you are satisfied with the performance of the current model, you can use that model. Otherwise, you run more HPOs that are warm-started with the results of previous runs so as to discover a better model.
Get professional support services
Amazon offers professional support services to help you with problems in your
machine learning on Neptune projects. If you get stuck, reach out to Amazon support