Pattern Mining

The PatternMiningService discovers common execution patterns, anomalous traces, and pattern clusters from agent trace data. It implements sequential pattern mining (PrefixSpan-style), graph-based pattern discovery, and anomaly detection algorithms.

Overview

Pattern mining enables the platform to learn from agent behavior over time. By identifying common successful patterns, unusual execution sequences, and clusters of similar behavior, the system can optimize agent routing, detect problems early, and provide pattern-based recommendations.

Source: data-plane/ai-service/src/context_graph/services/pattern_mining_service.py

Pattern Types

Sequential Patterns

Discovered from the ordered sequence of action types within traces:

Anomaly Scores

Each trace receives an anomaly score based on deviation from established patterns:

Pattern Clusters

Groups of similar patterns identified by clustering algorithms:

Mining Configuration

from context_graph.models.process import PatternMiningConfig
 
config = PatternMiningConfig(
    min_support=5,            # Minimum traces for a valid pattern
    min_confidence=0.3,       # Minimum pattern confidence
    max_pattern_length=20,    # Maximum steps in a pattern
    anomaly_threshold=0.7,    # Score above which a trace is anomalous
)

Mining Operations

Run Pattern Mining

result = await service.mine_patterns(
    tenant_id="acme",
    traces=traces,
    config=config,
)
# MiningResult(
#     patterns_discovered=15,
#     traces_analyzed=500,
#     anomalies_detected=12,
#     mining_duration_ms=1234.5,
# )

Detect Anomalies

anomalies = await service.detect_anomalies(
    tenant_id="acme",
    traces=traces,
    threshold=0.7,
)

REST API

Use Cases