Data Manifest Specification

This document describes the format of data manifests.

Data manifests are written as YAML files. Only the JSON-compatible semantics of YAML are used, so in the future they may be representable in formats such as JSON or TOML.

Overall Layout

A manifest file represents either a list or a map.

If it is a map and contains at least one of the keys file or type, then the map is taken to describe a single data file (see below).

Otherwise, the file describes a collection of data files. If it is a map, then those files are labelled with their map keys; otherwise, they are labeled with their positions (starting from 0).

For example:

ratings:
  file: "ratings.csv"
  format: csv
  entity: rating
  header: true
items:
  file: "items.csv"
  entity: movie
  header: true
  columns:
    movieId: id
    title: name

defines a data source consisting of ratings data, read from ratings.csv, and movie titles read from items.csv. The files have labels. The same manifest can be written with numeric labels as follows:

- file: "ratings.csv"
  format: csv
  entity: rating
  header: true
- file: "items.csv"
  entity: movie
  header: true
  columns:
    movieId: id
    title: name

Data Source Description

Individual data sources are described with the following schema.

type

The data source type. Currently only textfile is supported.

The default source type is textfile.

The remainder of the keys are defined by the particular data source.

textfile data sources

textfile sources read data from text files, with one entity per line. A text file may have one or more lines of header data.

file

the file to read.

format

The file format. Can be one of:

  • delimited — delimited, columnar text (default delimiter is \t)

  • csv — delimited with a delimiter of ,

  • tsv — tab-separated (delimited with delimiter of \t)

delimiter

The delimiter string for files with the delimited format.

entity_type

The name of the entity type contained in this file. The entity type is also used to provide defaults for the columns. The default is rating.

builder

The entity builder to be used for these entities. The entity type may provide a default; otherwise, org.lenskit.data.entities.BasicEntityBuilder is used. The keyword basic can be used to refer to the basic entity builder, to override a default entity builder if desired.

header

Whether the file has a header. If true, the file has a single-line header; if false, no header is assumed. If an integer, it is the number of header lines. The default is false.

columns

A list describing the columns in the file (for columnar formats). A column descriptor can be either a string, giving a column name, or a map with keys name (the column name) and type (the column type, see [attribute data types](#data-types)).

If header is true, then this can be a map whose keys are column header labels and whose values are column descriptors.

If the id column is not specified, then entity IDs are synthesized from the line numbers in the file.

indexes

A list of attribute names to be indexed for fast lookup. If no indexes are specified, item and user if they are present on the entities.

meta

Metadata about the data, such as the domain for rating values.` Some of the common metadata:

domain

The valid range of rating values. A map with the keys minimum, maximum, and precision; for example, 0.5-5 star data with 1/2 star precision would be described as follows:

meta:
  domain:
    minimum: 0.5
    maximum: 5
    precision: 0.5

Attribute Data Types

The following types are supported for attributes:

int or Integer

Java integer.

long or Long

Java long.

double, real, or Double

Java double.

string or String

Java String.

Java class name

The corresponding class. Must be convertible with Joda-Convert.

The entity type may provide default types for various attribute names, in addition to providing a default set of columns if columns is missing entirely. If no default is available and the type is not specified, attributes are assumed to be strings.

results matching ""

    No results matching ""