Manifold

This project is stable and being incubated for long-term support.

<img alt="Manifold" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/Manifold_Header.jpg" width="600">

Manifold is a model-agnostic visual debugging tool for machine learning.

Understanding ML model performance and behavior is a non-trivial process, given the intrisic opacity of ML algorithms. Performance summary statistics such as AUC, RMSE, and others are not instructive enough for identifying what went wrong with a model or how to improve it.

As a visual analytics tool, Manifold allows ML practitioners to look beyond overall summary metrics to detect which subset of data a model is inaccurately predicting. Manifold also explains the potential cause of poor model performance by surfacing the feature distribution difference between better and worse-performing subsets of data.

Table of contents

• Prepare your data
• Interpret visualizations
• Using the demo app
• Using the component
• Contributing
• Versioning
• License

Prepare Your Data

There are 2 ways to input data into Manifold:

• csv upload if you use the Manifold demo app, or
• convert data programatically if you use the Manifold component in your own app.

In either case, data that's directly input into Manifold should follow this format:

const data = {
  x:     [...],         // feature data
  yPred: [[...], ...]   // prediction data
  yTrue: [...],         // ground truth data
};

Each element in these arrays represents one data point in your evaluation dataset, and the order of data instances in x, yPred and yTrue should all match. The recommended instance count for each of these datasets is 10000 - 15000. If you have a larger dataset that you want to analyze, a random subset of your data generally suffices to reveal the important patterns in it.

x: {Object[]}

A list of instances with features. Example (2 data instances):

[{feature_0: 21, feature_1: 'B'}, {feature_0: 36, feature_1: 'A'}];

yPred: {Object[][]}

A list of lists, where each child list is a prediction array from one model for each data instance. Example (3 models, 2 data instances, 2 classes ['false', 'true']):

[
  [{false: 0.1, true: 0.9}, {false: 0.8, true: 0.2}],
  [{false: 0.3, true: 0.7}, {false: 0.9, true: 0.1}],
  [{false: 0.6, true: 0.4}, {false: 0.4, true: 0.6}],
];

yTrue: {Number[] | String[]}

A list, ground truth for each data instance. Values must be numbers for regression models, must be strings that match object keys in yPred for classification models. Example (2 data instances, 2 classes ['false', 'true']):

['true', 'false'];

Interpret visualizations

This guide explains how to interpret Manifold visualizations.

Manifold consists of:

• Performance Comparison View which compares prediction performance across models, across data subsets
• Feature Attribution View which visualizes feature distributions of data subsets with various performance levels

Performance Comparison View

This visualization is an overview of performance of your model(s) across different segments of your data. It helps you identify under-performing data subsets for further inspection.

Reading the chart

<img alt="performance comparison view" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/performance_comparison_1.png" width="600">

1. X axis: performance metric. Could be log-loss, squared-error, or raw prediction.
2. Segments: your dataset is automatically divided into segments based on performance similarity between instances, across models.
3. Colors: represent different models.

<img alt="performance comparison view unit" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/performance_comparison_2.png" width="600">

1. Curve: performance distribution (of one model, for one segment).
2. Y axis: data count/density.
3. Cross: the left end, center line, and right end are the 25th, 50th and 75th percentile of the distribution.

Explanation

Manifold uses a clustering algorithm (k-Means) to break prediction data into N segments based on performance similarity.

The input of the k-Means is per-instance performance scores. By default, that is the log-loss value for classification models and the squared-error value for regression models. Models with a lower log-loss/squared-error perform better than models with a higher log-loss/squared-error.

If you're analyzing multiple models, all model performance metrics will be included in the input.

Usage

• Look for segments of data where the error is higher (plotted to the right). These are areas you should analyze and try to improve.

• If you're comparing models, look for segments where the log-loss is different for each model. If two models perform differently on the same set of data, consider using the better-performing model for that part of the data to boost performance.

• After you notice any performance patterns/issues in the segments, slice the data to compare feature distribution for the data subset(s) of interest. You can create two segment groups to compare (colored pink and blue), and each group can have 1 or more segments.

Example

<img alt="performance comparison view example" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/performance_comparison_3.png" width="600">

Data in Segment 0 has a lower log-loss prediction error compared to Segments 1 and 2, since curves in Segment 0 are closer to the left side.

In Segments 1 and 2, the XGBoost model performs better than the DeepLearning model, but DeepLearning outperforms XGBoost in Segment 0.

<br/>

Feature Attribution View

This visualization shows feature values of your data, aggregated by user-defined segments. It helps you identify any input feature distribution that might correlate with inaccurate prediction output.

Reading the chart

<img alt="feature attribution view" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/feature_attribution_1.png" width="600">

1. Histogram / heatmap: distribution of data from each data slice, shown in the corresponding color.
2. Segment groups: indicates data slices you choose to compare against each other.
3. Ranking: features are ranked by distribution difference between slices.

<img alt="feature attribution view unit" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/feature_attribution_2.png" width="600">

1. X axis: feature value.
2. Y axis: data count/density.
3. Divergence score: measure of difference in distributions between slices.

Explanation

After you slice the data to create segment groups, feature distribution histograms/heatmaps from the two segment groups are shown in this view.

Depending on the feature type, features can be shown as heatmaps on a map for geo features, distribution curve for numerical features, or distribution bar chart for categorical features. (In bar charts, categories on the x-axis are sorted by instance count difference. Look for differences between the two distributions in each feature.)

Features are ranked by their KL-Divergence - a measure of difference between the two contrasting distributions. The higher the divergence is, the more likely this feature is correlated with the factor that differentiates the two Segment Groups.

Usage

• Look for the differences between the two distributions (pink and blue) in each feature. They represent the difference in data from the two segment groups you selected in the Performance Comparison View.

Example

<img alt="feature attribution view example" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/feature_attribution_3.png" width="600">

Data in Groups 0 and 1 have obvious differences in Features 0, 1, 2 and 3; but they are not so different in features 4 and 5.

Suppose Data Groups 0 and 1 correspond to data instances with low and high prediction error respectively, this means that data with higher errors tend to have lower feature values in Features 0 and 1, since peak of pink curve is to the left side of the blue curve.

<br/>

Geo Feature View

If there are geospatial features in your dataset, they will be displayed on a map. Lat-lng coordinates and h3 hexagon ids are currently supoorted geo feature types.

Reading the chart

<img alt="geo feature view lat-lng" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/geo_feature_1.png" width="600">

1. Feature name: when multiple geo features exist, you can choose which one to display on the map.
2. Color-by: if a lat-lng feature is chosen, datapoints are colored by group ids.
3. Map: Manifold defaults to display the location and density of these datapoints using a heatmap.

<img alt="geo feature view hex id" src="https://d1a3f4spazzrp4.cloudfront.net/manifold/docs/geo_feature_2.png" width="600">

1. Feature name: when choosing a hex-id feature to display, datapoints with the same hex-id are displayed in aggregate.
2. Color-by: you can color the hexagons by: average model performance, percentage of segment group 0, or total count per hexagon.
3. Map: all metrics that are used for coloring are also shown in tooltip