ML Models

Log predictions and actuals for classification, regression, ranking, and object detection models. Monitor drift, performance, and data quality.

Key Capabilities

Stream or batch logging
Embedding features for drift detection
SHAP values for explainability
Tags and metadata for segmentation
Support for delayed actuals
Export data for offline analysis

Stream Logging

Log individual predictions in real-time as they occur in production. log_stream sends the request asynchronously and returns a concurrent.futures.Future. Call .result() on the future to block until the request completes.

from arize.ml.types import ModelTypes, Environments, Embedding

future = client.ml.log_stream(
    space_id="your-space-id",
    model_name="fraud-detection-v1",
    model_type=ModelTypes.SCORE_CATEGORICAL,
    environment=Environments.PRODUCTION,
    prediction_id="unique-prediction-id",
    prediction_label=("not fraud", 0.85),
    actual_label=("fraud", 1.0),
    features={
        "transaction_amount": 150.0,
        "merchant_category": "online_retail",
        "user_age": 32,
    },
    embedding_features={
        "user_embedding": Embedding(
            vector=[0.1, 0.2, 0.3],
            data="user_123",
        ),
    },
)

print(f"Logged prediction: {future.result().status_code}")

With Tags and Metadata

Add custom tags for segmentation and filtering.

future = client.ml.log_stream(
    space_id="your-space-id",
    model_name="fraud-detection-v1",
    model_type=ModelTypes.SCORE_CATEGORICAL,
    environment=Environments.PRODUCTION,
    prediction_id="unique-prediction-id",
    prediction_label=("not fraud", 0.85),
    features={"transaction_amount": 150.0},
    tags={"risk_level": "high", "merchant_type": "new"},
)

Batch Logging

Log bulk predictions from historical data or batch processing.

from arize.ml.types import Schema, EmbeddingColumnNames
import pandas as pd

# Define schema to map DataFrame columns
schema = Schema(
    prediction_id_column_name="prediction_id",
    timestamp_column_name="prediction_ts",
    prediction_label_column_name="predicted_label",
    actual_label_column_name="actual_label",
    feature_column_names=["feature_1", "feature_2", "feature_3"],
    embedding_feature_column_names={
        "text_embedding": EmbeddingColumnNames(
            vector_column_name="text_vector",
            link_to_data_column_name="text_content",
        ),
    },
)

# Log batch data
response = client.ml.log(
    space_id="your-space-id",
    model_name="fraud-detection-v1",
    model_type=ModelTypes.SCORE_CATEGORICAL,
    environment=Environments.PRODUCTION,
    dataframe=prod_df,
    schema=schema,
    model_version="1.2.0",
)

print(f"Log status: {response.status_code}")

With SHAP Values

Include SHAP values for model explainability.

schema = Schema(
    prediction_id_column_name="prediction_id",
    prediction_label_column_name="predicted_label",
    feature_column_names=["feature_1", "feature_2"],
    shap_values_column_names={
        "feature_1": "shap_feature_1",
        "feature_2": "shap_feature_2",
    },
)

response = client.ml.log(
    space_id="your-space-id",
    model_name="fraud-detection-v1",
    model_type=ModelTypes.SCORE_CATEGORICAL,
    environment=Environments.PRODUCTION,
    dataframe=df_with_shap,
    schema=schema,
)

Export Data

Export ML model data for offline analysis, custom processing, or archival.

from datetime import datetime
from arize.ml.types import Environments

start_time = datetime.strptime("2024-01-01", "%Y-%m-%d")
end_time = datetime.strptime("2026-01-01", "%Y-%m-%d")

# Export to DataFrame
df = client.ml.export_to_df(
    space_id="your-space-id",
    model_name="fraud-detection-v1",
    environment=Environments.PRODUCTION,
    model_version="1.2.0",
    start_time=start_time,
    end_time=end_time,
)

print(f"Exported {len(df)} records")

Export to Parquet

client.ml.export_to_parquet(
    space_id="your-space-id",
    model_name="fraud-detection-v1",
    environment=Environments.PRODUCTION,
    start_time=start_time,
    end_time=end_time,
    path="./model_data_export.parquet",
)

Export capabilities:

Time-range filtering
DataFrame or Parquet output
Efficient Arrow Flight transport for large exports
Progress bars for long-running exports

Supported Model Types

Model Type	Use Case
`SCORE_CATEGORICAL`	Multi-class classification with scores
`MULTI_CLASS`	Multi-class classification
`BINARY_CLASSIFICATION`	Binary classification
`NUMERIC`, `REGRESSION`	Regression tasks
`RANKING`	Ranking and recommendation systems
`OBJECT_DETECTION`	Computer vision object detection
`GENERATIVE_LLM`	Not supported here — use `client.spans` instead for LLMs

Supported Environments

Environment	Description
`PRODUCTION`	Live production traffic
`TRAINING`	Training dataset
`VALIDATION`	Validation/test dataset (requires `batch_id`)
`CORPUS`	Document corpus logging (`log` only, with `CorpusSchema`)

API reference

API clients

Key Capabilities

Stream Logging

With Tags and Metadata

Batch Logging

With SHAP Values

Export Data

Export to Parquet

Supported Model Types

Supported Environments

​Key Capabilities

​Stream Logging

​With Tags and Metadata

​Batch Logging

​With SHAP Values

​Export Data

​Export to Parquet

​Supported Model Types

​Supported Environments

Key Capabilities

Stream Logging

With Tags and Metadata

Batch Logging

With SHAP Values

Export Data

Export to Parquet

Supported Model Types

Supported Environments