Automatic feature engineering

A typical problem: to write a Metarank config file with event to feature mapping, you need to read the docs on feature extraction and well-understand your click-through input dataset:

Which fields do items have? Which values does each field have?
Do these values look like categories?
How many unique values are there per field?

Nobody likes reading docs and writing YAML, so Metarank has an AutoML level-4 style generator of typical feature extractors based on the historical click-through dataset you already have.

Running the autofeature generator

Use the autofeature sub-command from the main binary:

                __                              __    
  _____   _____/  |______ ____________    ____ |  | __
 /     \_/ __ \   __\__  \\_  __ \__  \  /    \|  |/ /
|  Y Y  \  ___/|  |  / __ \|  | \// __ \|   |  \    < 
|__|_|  /\___  >__| (____  /__|  (____  /___|  /__|_ \
      \/     \/          \/           \/     \/     \/ ver:None
Usage: metarank <subcommand> <options>

Subcommand: autofeature - generate reference config based on existing data
  -d, --data  <arg>      path to a directory with input files
  -f, --format  <arg>    input file format: json, snowplow, snowplow:tsv,
                         snowplow:json (optional, default=json)
  -o, --offset  <arg>    offset: earliest, latest, ts=1663161962, last=1h
                         (optional, default=earliest)
      --out  <arg>       path to an output config file
  -r, --ruleset  <arg>   set of rules to generate config: stable, all (optional,
                         default=stable, values: [stable, all])
  -c, --cat-threshold  <arg>   min threshold of category frequency, when its
                               considered a catergory (optional, default=0.003)
  -h, --help             Show help message

For all other tricks, consult the docs on https://docs.metarank.ai

An example minimal command to generate the config file for your dataset:

java -jar metarank.jar autofeature --data /path/to/events.json --out /path/to/config.yaml

For a RankLens dataset, for example, it will emit the following:

15:32:11.284 INFO  a.metarank.main.command.AutoFeature$ - Generating config file
15:32:11.524 INFO  ai.metarank.source.FileEventSource - path=/home/shutty/code/metarank/src/test/resources/ranklens/events/events.jsonl.gz is a file
15:32:11.537 INFO  ai.metarank.source.FileEventSource - file /home/shutty/code/metarank/src/test/resources/ranklens/events/events.jsonl.gz selected=true (timeMatch=true formatMatch=true)
15:32:11.538 INFO  ai.metarank.source.FileEventSource - reading file /home/shutty/code/metarank/src/test/resources/ranklens/events/events.jsonl.gz (with gzip decompressor)
15:32:15.686 INFO  a.metarank.main.command.AutoFeature$ - Event model statistics collected
15:32:15.691 INFO  a.m.m.c.a.r.InteractionFeatureRule$ - generated interacted_with feature for interaction 'click' over field 'writer'
15:32:15.692 INFO  a.m.m.c.a.r.InteractionFeatureRule$ - generated interacted_with feature for interaction 'click' over field 'tags'
15:32:15.692 INFO  a.m.m.c.a.r.InteractionFeatureRule$ - generated interacted_with feature for interaction 'click' over field 'director'
15:32:15.693 INFO  a.m.m.c.a.r.InteractionFeatureRule$ - generated interacted_with feature for interaction 'click' over field 'title'
15:32:15.693 INFO  a.m.m.c.a.r.InteractionFeatureRule$ - generated interacted_with feature for interaction 'click' over field 'genres'
15:32:15.693 INFO  a.m.m.c.a.r.InteractionFeatureRule$ - generated interacted_with feature for interaction 'click' over field 'actors'
15:32:15.700 INFO  a.m.m.c.a.r.NumericalFeatureRule$ - generated `number` feature for item field popularity in the range 0.6..967.351
15:32:15.704 INFO  a.m.m.c.a.r.NumericalFeatureRule$ - generated `number` feature for item field vote_avg in the range 3.2..8.7
15:32:15.705 INFO  a.m.m.c.a.r.NumericalFeatureRule$ - generated `number` feature for item field release_date in the range 3.18816E8..1.56168E9
15:32:15.707 INFO  a.m.m.c.a.r.NumericalFeatureRule$ - generated `number` feature for item field budget in the range 0.0..3.8E8
15:32:15.708 INFO  a.m.m.c.a.r.NumericalFeatureRule$ - generated `number` feature for item field vote_cnt in the range 15.0..30232.0
15:32:15.709 INFO  a.m.m.c.a.r.NumericalFeatureRule$ - generated `number` feature for item field runtime in the range 3.0..242.0
15:32:15.719 INFO  a.m.m.c.a.rules.StringFeatureRule - field writer is not looking like a categorial value, skipping
15:32:15.726 INFO  a.m.m.c.a.rules.StringFeatureRule - field tags is not looking like a categorial value, skipping
15:32:15.728 INFO  a.m.m.c.a.rules.StringFeatureRule - field director is not looking like a categorial value, skipping
15:32:15.730 INFO  a.m.m.c.a.rules.StringFeatureRule - field title is not looking like a categorial value, skipping
15:32:15.731 INFO  a.m.m.c.a.rules.StringFeatureRule - item field genres has 19 distinct values, generated 'string' feature with index encoding for top 11 items
15:32:15.734 INFO  a.m.m.c.a.rules.StringFeatureRule - field actors is not looking like a categorial value, skipping
15:32:15.735 INFO  a.m.m.c.a.rules.RelevancyRule$ - skipped generating relevancy feature: non_zero=0 min=Some(0.0) max=Some(0.0)
15:32:15.851 INFO  ai.metarank.main.Main$ - My job is done, exiting.

Process finished with exit code 0

Supported heuristics

Metarank has multiple sets of heuristics to generate feature configuration, toggled by the --ruleset CLI option:

stable: a default one, ruleset with less agressive heuristics, proven to be safe in production use.
all: generates all features it can, even the problematic ones (like CTR, which may introduce biases).

The following stable heuristics are supported:

Numeric: all numerical item fields are encoded as a number feature. So for a numeric field budget describing a movie budget in dollars, it will generate the following feature extractor defitition:

- source: item.budget
  type: number
  name: budget
  scope: item

String: string item fields with low-cardinality are encoded as a string feature. So movie genres field is a good candidate for this type of heuristic due to its low cardinality:

- name: genres
  type: string
  source: item.genres
  scope: item
  encode: index
  values:
  - drama
  - comedy
  - thriller
  - action
  - adventure
  - romance
  - crime
  - science fiction
  - fantasy
  - family
  - horror

If you have a lot of distinct categories, and Metarank does not pick them up (e.g. decides that a category is too infrequent, and you get much less possible categories than expected), you can lower the category frequency threshold with a --cat-threshold flag.

The default --cat-threshold value of 0.003 means that only categories with frequencies above 0.3% are included.

InteractedWith: all interaction over low-cardinality fields are translated to interacted_with feature. So if a user clicked on an item with horror genre, other horror movies may get extra points:

- name: click_genres
  type: interacted_with
  scope: user
  interaction: click
  field: item.genres

Relevancy: if rankings with non-zero relevancy are present, then a feature relevancy is built:

- name: relevancy
  type: relevancy

Vector: all numerical vectors are transformed into statically-sized features. Vectors of static size are passed through as-is, and variable-length vectors are reduced into a quadruplets of [min, max, size, avg] values:

- name: embedding
  type: vector
  field: item.embedding // must be a singular number or a list of numbers
  scope: item
  # which reducers to use. optional. Default: [min, max, size, avg]
  reduce: [vector16]

The all ruleset contains all stable heuristics with an addition of a couple of extra ones:

Rate: For all interaction types a rate feature is generated over multiple typical time windows:

- name: click_rate
  type: rate
  top: click
  bucket: 1d
  bottom: impression
  scope: item
  periods:
  - 3
  - 7
  - 14
  - 30

InteractionCount: all interaction types are wrapped into interaction_count feature.

- name: count_click
  type: window_count
  bucket: 1d
  scope: item
  interaction: click
  periods:
  - 3
  - 7
  - 14
  - 30

Why stable ruleset has no counters?

The main difference between two rulesets is the lack of rate/window_count features, which is made deliberately:

rate/window_count features usually introduce a popularity bias to the final ranking: as people tend to click more on popular items, ML model may attempt to always put popular items on top just because they're popular.
this behavior may be not that bad from the business KPI standpoint, but may make your ranking more static and less affected by past visitor actions.

PreviousAutomated ML model retraining NextRunning in production

Last updated 2 years ago