Validated income-modeling pipeline and planned HGB sweep

2 min read

Validated income-modeling pipeline and planned HGB sweep

  • Day: 2026-05-17
  • Time: 11:10 to 11:20
  • Project: Dev
  • Workspace: WP 2: Operational
  • Status: Completed
  • Priority: HIGH
  • Assignee: Matías Nehuen Iglesias
  • Tags: Ml-Pipeline, Diagnostics, Ridge, Lasso, Hgb, Artifact-Management

Description

Session Goal

Validate the income-modeling pipeline end-to-end, inspect diagnostics from the debug run, and decide the next experimental step for the thesis regression workflow.

Key Activities

  • Converted the project runbook into a concrete operational sequence: clean install, dataset verification, debug run, diagnostics generation, and guarded full baseline run.
  • Treated the run directory as the canonical source of truth for artifacts and validation outputs.
  • Reviewed debug-run results for the linear and ridge regression baselines.
  • Evaluated residual diagnostics and identified the main failure mode as income-compression bias.
  • Designed a more rigorous regularization sweep for Ridge and Lasso using log-spaced alpha grids, train/CV diagnostics, coefficient norms, and sparsity metrics.
  • Reframed HistGradientBoostingRegressor as the main nonlinear probe and outlined a staged debug/sweep plan with targeted hyperparameters, plots, and thesis-oriented questions.

Achievements

  • Confirmed the local ML experiment lifecycle works end-to-end from input checks through artifact archiving and diagnostics generation.
  • Established that Linear and Ridge behave nearly identically in debug mode, suggesting limited benefit from weak regularization in the current setup.
  • Clarified the key diagnostic insight: low incomes are overpredicted while high incomes are underpredicted, indicating strong compression toward the center.
  • Identified a small infrastructure issue where split-specific diagnostic plots can be overwritten, and recommended fixing artifact naming before broader runs.
  • Determined that the pipeline is ready for a full baseline run if runtime is acceptable, and that HGB should be assessed as the next nonlinear candidate.

Pending Tasks

  • Fix split-specific diagnostic plot output so artifacts do not overwrite each other.
  • Run a dedicated Ridge/Lasso regularization sweep instead of relying on a two-point alpha check.
  • Execute the full baseline run after confirming runtime and artifact handling.
  • Implement and validate the HistGradientBoostingRegressor debug/sweep experiment plan.

Evidence

  • source_file=2026-05-17.sessions.jsonl, line_number=1, event_count=0, session_id=d9459ec01b08aa6f213e686ab03be28376b782bcaea1b8b4715649815ba6891d
  • event_ids: []

Graph View

Backlinks