HNYC Spatial Research on Spatial Linkages

Summer 2020

1. Developed confidence score tuning process (see disambiguation_1880 folder)
2. Developed interpolation process (see interpolation folder for code and interpolation_notebooks folder for jupyter notebooks that document process)
3. For overview of process and next steps see google drive HNYC_Project/Projects/spatial_linkage/Spatial Linkage & Interpolation: Summer 2020.ppt and HNYC_Project/Projects/spatial_linkage/Spatial Linkage and Interpolation Workflow.doc

Spring 2020 Actions

Accomplished:

1. Updated confidence score to include census conflicts (see disambiguation_2.ipynb)
2. Merged lat lng data to matched data -> produced matched.csv
3. Experimented with methods to add spatial weights, including graph-based and cluster-based approaches on a subset of the data (see spatial-disambiguation.ipynb)
4. Outlined overall workflow for disambiguation using bipartite graph matching algorithm (linkage-disambiguation.ipynb).
5. Ran algorithms on full dataset & obtained initial performance metrics
   1. Functionalized process using disambiguation module
6. Tuned algorithms and compared metrics + recommendations
7. Updated ES matching process to reflect metaphone matching See Google Drive Documentation > disambiguation > Spatial Linkage: Spring 2020 for slides

Summary of Processes Taken in Fall 2019

1880 Process

Two sources:

• City Directory: name (first, last, initial), address, ID, occupation, ED, ward, block number (constructed)
• Census: address (hidden during match process), ID (different from CD), occupation, ED, ward, age, gender, dwelling, household characteristics

Steps:
0. Preprocessing: changing names to their phonetic index

1. Initial entity link (to generate possible matches): criteria = same ED + JW dist < 2 on indexed name
2. Disambiguation (to choose between non-unique matches):
   a. Generate a confidence score based on occupation (having occupation), age > 12 (not implemented), JW dist of name, relative probability (number of non-unique matches)

1850

Similar, but no ED and address data, only ward data in the census.

Sitemap

• /ES_matching contains all work related to elastic search
  • readme.md gives guidelines on how to implement the process
  • fall_2019_analysis.md describes the work done up to fall 2019
  • /doc: relevant documentation of the process
  • /src: source code for running elastic search on our data
  • explore_output.ipynb quick validation of the ES output (targeted at understanding whether metaphone matching was implemented in ES process)
• /disambiguation_1850 contains all work related to disambiguation of 1880 data
  • disambiguation_1850_v1.ipynb runs disambiguation process on 1850 ES output
• /disambiguation_1880 contains all work related to disambiguation of 1880 data
  • Confidence_Score_Tuning.ipynb: Documents confidence score tuning process
  • Confidence_Score_Tuning_v02.ipynb: Confidence score tuning results used for 1850 v02 disambiguation run (10/2020), uses old version of 1880 data because of data issues with new 1880
  • Confidence_Score_Tuning_new_1880_data_draft.ipynb: Attempt to tune confidence score with new 1880 data, revealed that there was an issue with the data
  • /_archived: archived scripts
    • /confidence_score
      • preprocess.ipynb: preprocessing of data including generation of metaphones
      • get_confidence_score.ipynb: outlines process for generating confidence score, including calculation of jaro-wrinkler
      • disambiguation_analysis.ipynb: EDA on confidence scores -- generates match_results_confidence_score.csv and fall_2019_disambiguation_report.md
    • /linkage_eda
      • spatial-disambiguation.ipynb: documentation of different spatial weight algorithms
      • linkage-disambiguation.ipynb: outline record linkage approach (conceptually valid but code is outdated)
      • linkage_full_run_v1.ipynb: applies basic algorithm to the whole dataset + initial performance analysis
      • linkage_eda.ipynb: applies various iterations of algorithm to the whole dataset + conclusions
      • linkage_eda_v2.ipynb: applies updated geocodes on best 2 algorithms + improved benchmarking
  • run_link_records.ipynb: implemented record matching using pyspark
  • confidence_score_latlng.ipynb: adding of census conflicts to confidence score, merging of lat lng data (contains most updated confidence score formula)
  • linkage_full_run_SPRING_LATEST.ipynb: informed by linkage EDA (see archive), generates latest disambiguated output from ES matching (with metaphone issue fixed)
• /interpolation_notebooks
  • Process_Documentation
    • Disambiguation_Analysis_v01.ipynb: Resolves dwelling conflicts, calculates statistics,explores distance based sequences, and interpolation between known dwellings
    • Interpolation_v01.ipynb: Runs through current version of predicting unknown records
    • Disambiguation_Analysis_v02.ipynb: Same information as v01, for new data (10/2020)
    • Interpolation_v02.ipynb: Same information as v01, for new data (10/2020)
  • Concepts_and_Development:
    • Block and Centroid Prediction with Analysis.ipynb: Walks through approaches to predicting block numbers directly, and then clusters (tests different clustering algorithms)
    • Block Centroids and What They Represent, 1850.ipynb: Creates block centroids and illustrates them with visualizations
    • Dwelling Addresses Fill In and Conflict Resolution Development.ipynb: Development of conflict resolution within dwelling process
    • Developing Distance Based Sequences.ipynb: Process of developing distance based sequences
    • Model Comparison.ipynb: Tests a few different model options (no in depth tuning)
    • Sequences Exploration.ipynb: Tests different iterations of sequence identification
    • Model Exploration.ipynb: Brief experimentation with using neural networks, incomplete because of preprocessing necessary
  • Archived
    • 1880_1850_for_Interpolation.ipynb: Explores 1880 and 1850 census datasets
    • Feature_Exploration.ipynb: Explores some of the columns in 1880 and 1850 datasets in order to determine what they represents and if they can be used for modelling
    • Interpolation Pilots.ipynb: Working notebook for starting explorations of options for interpolation (often moved into a separate notebook when they seem worth looking at in more depth)
    • Linear_Model.ipynb: Creates and tests linear models for house number interpolation
    • Modeling Comparison.ipynb: Tests different modeling approaches for house numbers (currently linear model and gradient boosting) -- includes haversine sequences and block numbers as features
    • Block_Numbers Early Exploration.ipynb: Explore block numbers distributions/data analysis and try using them as feature to predict house number
    • Street_Dictionaries.ipynb: Tried out looking at street dictionaries for dwellings in between
    • Block Number Prediction.ipynb: Initial experiment with predicting block numbers
    • Interpolation between known address development.ipynb: Process of looking at values between known dwellings
• /interpolation See read me within this folder for details
• /disambiguation is a python module containing wrapper functions needed in the disambiguation process
  • init.py contains a Disambiguator object, when instantiated can be used to run entire disambiguation process, calling functions from below (see linkage_eda.ipynb for example on usage)
  • preprocess.py contains functions needed before applying disambiguation algorithms, including confidence score generation
  • disambiguation.py contains functions needed for disambiguation
  • analysis.py wrapper functions to produce performance metrics
  • confidence_score_tuning.py contains functions needed for the confidence tuning process
  • benchmarking.py contains Benchmark objects, to run benchmarking process in confidence tuning for 1880
• /matching_viz visualization web app to understand disambiguation output, see readme in folder for guidance on how to run

Data

Data is available in the HNYC Spatial Linkage Google Drive HNYC_Project/Projects/spatial_linkage/Data

1850 /1850

• 1850 disambiguated output: 1850_disambiguated.csv
• 1850 disambiguated output 10/2020 (current): 1850_mn_match_v02.csv
• 1850 ES matches: es-1850-22-9-2020.csv

1880

• Matches with confidence score (raw input for 1880 disambiguation processes): matches.csv
  • this is based off Fall 2019's Spark matching output
• Latest ES matches: es-1880-21-5-2020.csv
• Latest Disambiguated Output: disambiguated-21-5-2020.csv