AgentSkillsCN

user-pattern-matcher

利用机器学习将用户与成功的自动化模式进行匹配。适用场景如下：(1) 从数据库中查找相似用户，为推荐方案提供参考；(2) 将拟议方案与成功用户的典型模式进行对比验证；(3) 从现有用户规则中提取用户意图或目标；(4) 训练或更新模式匹配模型。该技能主要负责地图生成中的机器学习与数据科学相关环节。

SKILL.md
--- frontmatter
name: user-pattern-matcher
description: Match users to successful automation patterns using ML. Use when: (1) Finding similar users from the database to inform recommendations, (2) Validating a proposed plan against successful user patterns, (3) Extracting intent/goals from existing user rules, (4) Training or updating the pattern matching model. This skill handles the ML/data science aspects of map generation.

User Pattern Matcher

Find similar successful users and extract patterns to inform automation recommendations.

Overview

This skill uses machine learning to match new users to patterns from 2,500+ existing Sequence users. It analyzes user profiles and their automation rules to find relevant templates and validate proposed plans.

Data Sources

User Profiles (enrichment_data.csv)

Features available:

  • •ANNUALINCOME - Income bracket (UP_TO_10K, BETWEEN_10K_AND_25K, BETWEEN_25K_AND_50K, BETWEEN_50K_AND_100K, BETWEEN_100K_AND_250K, OVER_250K)
  • •OCCUPATION - Job category
  • •PRODUCTGOAL - Stated primary goal
  • •USER_TYPE - INDIVIDUAL or BUSINESS
  • •CURRENT_SUBSCRIPTION_NAME - Starter, Pro, Business, Growth
  • •DEBIT_CARD_SPENDER - true/false
  • •ACTIVATED_PORTS - Number of income sources connected
  • •ACCOUNTS_CONNECTED - Total connected accounts
  • •AGE_GROUP - Age bracket

User Rules (itaytestfinal.csv)

  • •organization_id - Links to profile
  • •description - Human-readable rules (parseable for pattern extraction)

Matching Process

1. Feature Engineering

Categorical encoding:

  • •Income bracket → ordinal (0-5)
  • •User type → binary
  • •Occupation → one-hot or embedding
  • •Goal → one-hot or embedding

Numerical features:

  • •Accounts connected (normalize)
  • •Activated ports (normalize)

Derived features:

  • •Complexity score = accounts_connected * activated_ports
  • •Business indicator = USER_TYPE == BUSINESS
  • •Goal alignment score (if comparing to a specific pattern)

2. Similarity Matching

For finding similar users:

code
Input: New user profile
Output: Top-K most similar existing users

Approach options:
1. KNN on encoded features (cosine similarity)
2. Embedding-based similarity
3. Rule-based filtering + similarity

Feature weights (suggested starting point):

  • •PRODUCTGOAL: High weight (primary intent signal)
  • •USER_TYPE: High weight (personal vs business very different)
  • •ANNUALINCOME: Medium weight (affects scale of automations)
  • •ACCOUNTS_CONNECTED: Medium weight (complexity indicator)
  • •OCCUPATION: Low weight (weak signal)

3. Pattern Extraction from Rules

Parse human-readable rules to extract:

Trigger patterns:

  • •"When funds are received" → INCOMING_FUNDS
  • •"When [Pod] balance is at least $X" → BALANCE_THRESHOLD
  • •Scheduled patterns → SCHEDULED

Action patterns:

  • •"X% of incoming funds" → PERCENTAGE
  • •"$X moves from" → FIXED
  • •"Anything above $X moves" → TOP_UP/overflow
  • •"Funds move from" → REMAINDER

Strategy patterns:

  • •Multiple liabilities with same source → debt payoff strategy
  • •Percentage splits from income → budget allocation
  • •Threshold-based moves → savings/reserve building

Use Cases

Find Similar Users

code
Input:
- User profile (income, occupation, goal, type)

Process:
1. Encode profile features
2. Calculate similarity to all users in database
3. Filter by same USER_TYPE
4. Return top 10 most similar

Output:
- List of similar user IDs
- Their rule descriptions
- Common patterns among them

Validate Proposed Plan

code
Input:
- Proposed automation plan
- User profile

Process:
1. Find similar successful users
2. Extract patterns from their rules
3. Compare proposed plan to common patterns
4. Flag deviations or missing elements

Output:
- Validation score
- Suggestions based on what similar users do
- Missing patterns to consider

Extract Intent from Rules

code
Input:
- Raw rule descriptions (from itaytestfinal.csv)

Process:
1. Parse trigger types
2. Parse action types
3. Identify node relationships
4. Classify into goal categories

Output:
- Inferred PRODUCTGOAL
- Automation complexity score
- Primary patterns used

Model Training Approach

Supervised: Goal Prediction

Train model to predict PRODUCTGOAL from rules.

  • •Input: Rule descriptions (text)
  • •Output: Goal category
  • •Use for: Inferring intent when goal not stated

Unsupervised: User Clustering

Cluster users by rule patterns + profile.

  • •Features: Profile + rule pattern features
  • •Method: K-means or hierarchical
  • •Use for: Finding user archetypes

Similarity: User Matching

Build similarity index for fast lookup.

  • •Encode all users
  • •Index with FAISS or similar
  • •Query with new user profile

Integration Points

With sequence-map-generator

  1. •Generator requests similar users for a profile
  2. •Matcher returns top matches + their patterns
  3. •Generator incorporates patterns into plan

With map-json-converter

  1. •Converter can request validation
  2. •Matcher checks if JSON structure matches successful patterns

References