bsovs / winhacks-ml-pipeline

recommendation encoder for home data using Spotify's annoy package

winhacks ml pipeline

This is a reference encoder for a home recommendation system created for winhacks 2021.

About

Encoder

• Pulls down all home data stored on a bigQuery instance and transforms it to a dataframe
• All text based attributes are assigned to a label encoder (saved to pkl files for query reference later)
• Each row is converted to a vector and added to the annoy index
• X trees are created with these vectors to the n nearest neighbours can be looked up very quick in memory

FastApi App

Endpoints Available

• /run/model/update (update annoy model for all homes)
• /run/model/fit (update collaborative rating in bigQuery) *needs payed deployment
• /run/profile/update (return an average vector for a user based on the homes they like)
• /query/by-embed (query n best fit homes for a vector)
• /query/by-attributes (query n best fit homes by given attributes)

BigQuery

• example query to insert homes:

INSERT INTO `winhacks-308216.homes_data.sample` (
        id,
        created_on,
        operation,
        property_type,
        place_name,
        place_with_parent_names,
        country_name,
        state_name,
        geonames_id,
        lat_lon,
        lat,
        lon,
        price,
        currency,
        price_aprox_local_currency,
        price_aprox_usd,
        surface_total_in_m2,
        surface_covered_in_m2,
        price_usd_per_m2,
        price_per_m2,
        floor,
        rooms,
        expenses,
        properati_url,
        description,
        title,
        images
)
 VALUES 
    (GENERATE_UUID(),
     CURRENT_DATE(),
     'sell',
        'house',
        '67 Shelborne Ave',
        null,
        'Canada',
        'Ontario',
        null,
        '43.7675433, -79.2738102',
        43.7675433,
        -79.2738102,
        4199000,
        'CAD',
        4199000,
        3311111,
        48009,
        48009,
        null,
        null,
        null,
        6,
        null,
        'https://www.zillow.com/homedetails/67-Shelborne-Ave-Toronto-ON-M5N-1Z2/2076679623_zpid/?',
        'A Truly Spectacular New Custom Home With High Quality Finishes Throughout. The Attention To Detail And Fine Craftsmanship Is Very Noticeable Upon Inspection. A Chef\'s Kitchen With Two Dishwashers, Two Sinks, Make Meal Prep An Occasion. Tarion New Home Warranty Included. High Ceilings And Heated Floors In All Tiled Areas, Main Floor, Bsmt And All Bathrooms. Snow Melt System In Driveway. Built In Sound System. Control 4 Electronic Lighting System.(Programmable)',
        '67 Shelborne Ave, Toronto, ON',
        [
        'https://photos.zillowstatic.com/fp/5538f14b835aaf729b65aa2cc27c6747-cc_ft_768.jpg',
        'https://photos.zillowstatic.com/fp/04c6dcb90326ed32c513cc3e65058e74-uncropped_scaled_within_1536_1152.webp', 
        'https://photos.zillowstatic.com/fp/7b58101189039bef8749b80d63ecc8e3-uncropped_scaled_within_1536_1152.webp',
        'https://photos.zillowstatic.com/fp/9ceb2ab6e26918c50abcaab346977873-uncropped_scaled_within_1536_1152.webp',
        'https://photos.zillowstatic.com/fp/93315d75c9a3484dc704f5330e00271d-uncropped_scaled_within_1536_1152.webp'
        ]
)

• query by distance and filter

WITH params AS (
  SELECT ST_GeogPoint(@longitude, @latitude) AS center,
         @limit AS maxn_homes,
         @radius AS maxdist_km
),
distance_from_center AS (
  SELECT
    id,
    created_on,
    operation,
    property_type,
    place_name,
    place_with_parent_names,
    country_name,
    state_name,
    geonames_id,
    lat_lon,
    lat,
    lon,
    price,
    currency,
    price_aprox_local_currency,
    price_aprox_usd,
    surface_total_in_m2,
    surface_covered_in_m2,
    price_usd_per_m2,
    price_per_m2,
    floor,
    rooms,
    expenses,
    properati_url,
    description,
    title,
    images,
    ST_GeogPoint(lon, lat) AS loc,
    ST_Distance(ST_GeogPoint(lon, lat), params.center) AS dist_meters
  FROM
    `winhacks-308216.homes_data.sample`,
    params
  WHERE ST_DWithin(ST_GeogPoint(lon, lat), params.center, params.maxdist_km*1000)
),
nearest_homes AS (
  SELECT 
    *, 
    RANK() OVER (ORDER BY dist_meters ASC) AS rank
  FROM 
    distance_from_center
),
filtered_homes AS (
  SELECT 
    station.*
  FROM 
    nearest_homes AS station, params
  WHERE 
    id NOT IN UNNEST(@filter)
),
nearest_nhomes AS (
  SELECT 
    station.* 
  FROM 
    filtered_homes AS station, params
  WHERE 
    rank <= params.maxn_homes
)
SELECT * from nearest_nhomes

• run fit on user interactions

CREATE OR REPLACE MODEL `winhacks-308216.ml.model`
OPTIONS
 (model_type='matrix_factorization',
  feedback_type='implicit',
  user_col='profile_id',
  item_col='home_id',
  rating_col='rating',
  l2_reg=30,
  num_factors=15) AS
SELECT
 profile_id,
 home_id,
 rating
FROM `winhacks-308216.profiles_data.sample`