Home > Tasks > Reasoning > DecisionTree

DecisionTreeTask

Constructs an LLM-driven symbolic decision tree classifier. Proposes semantic splitting rules using LLMs and validates them against data using Information Gain.

Category: Reasoning Output: Kotlin Code Model: GPT-4 Preferred
⚙️ Execution Config
{
  "task_type": "DecisionTree",
  "data_file": "data/customer_churn.csv",
  "target_column": "churn_risk", 
  "max_depth": 3,
  "candidate_rules": 5
}
→
👁️ Generated DecisionTree.kt
fun predict(record: Map<String, String>): String {
  // Probability: 0.89 (n=142)
  if ((record["usage_hours"]?.toDoubleOrNull() ?: 0.0) > 50.0) {
    if (record["support_tickets"] == "0") {
      return "Low Risk"
    } else {
      return "Medium Risk"
    }
  } else {
    return "High Risk"
  }
}

Test Workspace Browser

Explore generated decision trees and training logs in the task workspace.

TaskExecutionConfig Parameters

Field Type Description
data_file* String The data file to analyze (CSV or JSONL). Must exist in the workspace.
target_column* String The target column name to predict.
max_depth Int Maximum depth of the tree. Default: 3.
candidate_rules Int Number of candidate rules the LLM generates per node. Default: 5.

Task Lifecycle

  1. Initialization: Validates data_file existence and target_column presence in headers.
  2. Rule Proposal: LLM analyzes data samples to propose semantic rules (==, !=, >, <, contains, matches).
  3. Validation: Calculates Information Gain for each candidate rule against the full dataset.
  4. Recursive Split: Selects the optimal rule and recurses until max_depth, purity > 95%, or minimum record count is reached.
  5. Code Generation: Compiles the symbolic tree into a standalone, executable Kotlin function.

Embedded Execution

To invoke this task programmatically using the Cognotik library:

import com.simiacryptus.cognotik.plan.tools.reasoning.DecisionTreeTask
import com.simiacryptus.cognotik.plan.tools.reasoning.DecisionTreeTask.Companion.DecisionTree
fun generateModel(harness: UnifiedHarness, workspace: File) {
    val config = DecisionTreeTask.DecisionTreeTaskExecutionConfigData(
        data_file = "data/training_set.csv",
        target_column = "label",
        max_depth = 4,
        candidate_rules = 10
    )
    harness.runTask(
        taskType = DecisionTree,
        typeConfig = TaskTypeConfig(), // Default static config
        executionConfig = config,
        workspace = workspace,
        autoFix = true
    )
}

CLI / CI Integration

Run via the command line by passing the instruction to the orchestrator:

java -jar cognotik-cli.jar \
  --instruction "Build a decision tree for 'churn' using 'customers.csv' with depth 3" \
  --workspace ./project

Prompt Segment

The following context is injected into the orchestrator to describe this capability:

DecisionTree - Build an LLM-driven symbolic decision tree
  ** Specify the data file (CSV)
  ** Specify the target column to predict
  ** Configure max depth and candidate rules
  ** Uses LLM to propose semantic splitting rules
  ** Validates rules using Information Gain
  ** Generates executable code
display: grid;