Sequential Pattern Mining

Unlike standard association rule mining (e.g., market basket analysis), Sequential Pattern Mining explicitly discovers patterns where the order of events is significant. A pattern <{A}, {B}, {C}> means event A occurred, then later B, then later C. This is crucial for analyzing user sessions, process logs, DNA sequences, and financial transactions where timing matters.

• Example: In web clickstream analysis, the pattern <{Homepage}, {Search}, {Product Page}, {Checkout}> is meaningful, whereas an unordered set is not.

Algorithms incorporate constraints to make discovered patterns meaningful and computationally feasible.

• Time Constraints: Define maximum/minimum gaps between consecutive elements in a sequence (e.g., events B must follow A within 30 seconds).
• Sliding Window: Events occurring within a specified time window can be considered part of the same element in the sequence.
• Granularity: Analysis can be performed at different temporal resolutions (e.g., seconds, days, sessions), which dramatically changes the patterns found.

Pattern significance is measured statistically.

• Support: The percentage of input sequences that contain the candidate pattern. A high support indicates a common temporal pathway.
• Confidence: For a rule derived from a pattern (e.g., <{A}, {B}> → {C}), confidence measures the probability that C occurs given the prior sequence A then B.

These metrics filter out spurious correlations and identify robust, recurring temporal behaviors.

Key algorithms define the field's methodology.

• GSP (Generalized Sequential Patterns): An Apriori-based, breadth-first search algorithm. It uses a candidate generation-and-test approach, pruning the search space using the downward closure property (all subsequences of a frequent sequence must also be frequent).
• PrefixSpan (Prefix-Projected Sequential Pattern Mining): A pattern-growth, depth-first search algorithm. It avoids candidate generation by recursively projecting the database based on frequent prefixes, which is typically more efficient for long sequences.
• SPADE: Uses vertical id-list data formats for efficient lattice traversal.

In Agentic Memory and Context Management, this technique is foundational for Temporal Memory Sequencing.

• Predicting Agent Behavior: Mining an agent's own action histories to predict its next likely tool call or API execution.
• Anomaly Detection in Logs: Identifying deviations from normal operational sequences in multi-agent system orchestration.
• Workflow Discovery: Automatically discovering common procedural patterns from event streams in clinical workflow automation or autonomous supply chains.
• Enhancing Memory Retrieval: Informing time-aware retrieval strategies by understanding which past events typically co-occur in temporal proximity.

Sequential Pattern Mining interacts closely with other concepts in Temporal Memory Sequencing.

• Input: Operates on Event Streams and Time-Series data.
• Representation: Discovered patterns can populate an Event Causality Graph or Temporal Knowledge Graph.
• Mechanism: Relies on efficient Time-Series Indexing for scalable processing.
• Output: Patterns enable Sequence Prediction and inform Temporal Reasoning.
• Contrast: Differs from Event Correlation, which finds statistical relationships but not necessarily frequent ordered subsequences.

A continuous, time-ordered sequence of discrete events or state changes that serves as the foundational data source for temporal memory in autonomous agents. It is the raw input for Sequential Pattern Mining algorithms.

• Characteristics: Append-only, immutable, and high-velocity.
• Examples: User clickstreams, IoT sensor readings, financial transaction logs, and API call histories.
• Role in Memory: Acts as the persistent source of truth from which patterns are extracted and episodic memories are constructed.

A knowledge graph where facts (entities and relationships) are associated with timestamps or valid time intervals. This enables querying over evolving knowledge states and is a powerful structure for storing mined sequential patterns.

• Structure: Extends standard triples (subject, predicate, object) to include a temporal dimension (e.g., quadruples).
• Use Case: Represents discovered patterns like (User_A, purchased, Product_B, [2024-01-01, 2024-01-07]) as temporal facts.
• Advantage: Supports complex temporal reasoning queries, such as "What patterns were active during the last quarter?"

The computational process of mapping and comparing two or more temporal sequences to identify correspondences, similarities, or differences in their event order. It is critical for evaluating and generalizing mined patterns.

• Core Algorithm: Dynamic Time Warping (DTW) is a common technique that finds an optimal alignment between sequences of different lengths and speeds.
• Application: Used to measure similarity between a candidate pattern and a new event stream, or to cluster similar behavioral sequences.
• Outcome: Determines if a new sequence instance matches a known pattern despite minor variations in timing or event spacing.

The process of transforming low-level, time-stamped data into higher-level, interval-based concepts or states that are meaningful for reasoning. It creates the semantic building blocks for pattern mining.

• Process: Converts raw events (e.g., sensor_reading=72, click=/home) into abstract states (e.g., state=HIGH_TEMP, intent=NAVIGATE_HOME).
• Benefit: Reduces noise and dimensionality, allowing pattern mining algorithms to discover meaningful, domain-relevant sequences instead of spurious low-level correlations.
• Example: In healthcare, converting vital sign readings into abstract states like HYPERTENSIVE_EPISODE enables mining of patterns in patient deterioration.

A directed graph structure where nodes represent events and edges represent inferred causal or strong temporal precedence relationships. It moves beyond correlation to model potential chains of influence discovered through pattern mining.

• Construction: Often built by applying causal discovery algorithms (e.g., PCMCI, Granger causality) to sequential pattern results.
• Difference from Pattern: A sequential pattern (A -> B -> C) shows frequent co-occurrence; a causality graph infers A *causes* B.
• Utility: Enables counterfactual reasoning and predictive interventions ("If we prevent event A, will pattern C still occur?").