🗺 Embedding Insights

[mikeldking]

Goals

• Provide a way for users to visualize embeddings of a dataset in a UMAP point cloud. The point cloud should provide a visual representation of the embeddings' vector space. In addition, the point cloud should be colored by various heuristics such that the user can visually identify areas that are performing badly or that contain a high amount of anomalies.

• Provide Embedding drift analysis using UMAP projection of one or two datasets. These two datasets are crafted by a sample of the primary dataset compared against the full (down-sampled) baseline dataset.

Euclidean Distance

Sub-goal: display temporal drift over time so that we can perform timeseries based analysis when there are two datasets.

Backend

• Dataset bookends plumbed through to GQL #172
• Euclidean distance over time (timestamp indexes) #188
• Traffic (cardinality) over time #242
• [gql] Euclidean distance over time resolver #198
• Make the reference dataset optional in the data model #218

Metrics

• Euclidean distance over time (timestamp indexes) #188
• [metrics] Timeseries granularity parameterization #241

Datasets

• [datasets] vector column normalization #171
• [Datasets] sort records by time, add timestamp index #230

Components

• [components] Breadcrumbs #190

UMAP and Clustering

Backend

• [gql] timeseries interface #235
• Embedding GraphQL Node #168
• Run HDBSccan in series w/ UMAP #169
• GraphQL Plumbing for clusters and inferences #170
• [gql] UMAP points strawman in streawberry #195
• GraphQL Plumbing for UMAP and cluster hyperparameters (n neighbors, min distance) #227
• [UMAP] event summary field on each umap point #226
• Record level retrieval of event data (features, predictions, actuals) #229
• [gql] add data_quality_metrics to EmbeddingDimension #270
• [point-cloud] point visibility controls #282
• [point-cloud] Gallery / List view of selection #323
• [UMAP] plumb through prediction_id on embeddingMetadata #376

App

• Dataset time presets selector #243

• ☔ Timeseries Graph #192

• UMAP Viz #193

• [point-cloud] resizable panel system #251

• [point-cloud] Hyperparameter controls #341

• [point-cloud] Clustering Visualization #250

• [point-cloud] Color by feature #304

• [point-cloud] Color by tag #342

• [point-cloud]Color by correctness #223

• [ui] single dataset store dependency injection #345

• [Embeddings] Single-dataset timeseries of count #439

• [Embeddings] Tabular view with resizable column headers #438

• [Embeddings] Grid view of raw text and images #437

• [point-cloud] Point Selection #234

• [point-cloud] Showing point metadata (Event details) #305

• [point-cloud] Calculate cluster purity score (e.g. % distribution of data in a cluster) #244

• [point-cloud] [clustering] display cluster listing #343

• [point-cloud] [clustering] color clusters by purity score #344