heidi
: tidy data in Haskell
This library aims to bridge the gap between Haskell's precise but inflexible type discipline and the dynamic world of dataframes.
More specifically, heidi
aims to make it easy to analyze collections of Haskell values; users encode
their data (lists, maps and so on) into dataframes, and use functions provided by heidi
for manipulation.
If this sounds interesting, read on!
Introduction
A "dataframe" is conceptually a table of data that can be manipulated with a computer program; it potentially contains numbers, text and anything else that can be rendered as text.
In scientific practice, a "tidy" dataframe is a specific way of arranging the data in which each row represents a distinct observation ("data point") and each column a "feature" (i.e. some observable aspect) of the data.
Nowadays, data science is a very established practice and many software libraries offer excellent functionality for working with such dataframes. R
has tidyverse
, Python has pandas
, and so on.
What about Haskell?
The Frames
[1] library offers rigorous type safety and good runtime performance, at the cost of some setup overhead. Heidi
's main design goal instead is to have minimal overhead and possibly very low cognitive load to data science practitioners, at the cost of some type safety.
Quickstart
The following snippet demonstrates the minimal setup necessary to use heidi
:
{-# language DeriveGeneric #-} (1)
module MyDataScienceTask where
import GHC.Generics (2)
import Heidi
data Sales = Sales { item :: String, amount :: Int } deriving (Eq, Show, Generic) (3)
instance Heidi Sales (4)
All datatypes that are meant to be used within dataframes must be in the Heidi
typeclass, which in turn requires a Generic
instance.
The DeriveGeneric
language extension (1) enables the compiler to automatically write the correct incantations (3), as long as the user also imports the GHC.Generics
module (2) from base
.
The automatic dataframe encoding mechanism is made possible by the empty Heidi
instance (4).
It is also convenient to use DeriveAnyClass
to avoid writing the empty typeclass instance :
{-# language DeriveGeneric, DeriveAnyClass #-}
data Foo = Foo Int String deriving (Generic, Heidi)
Rationale
Out of the box, Haskell offers record types, e.g.
data Row a = MkRow { column1 :: Int, column2 :: String } deriving (Eq, Show)
which is handy because in one declaration you get a constructor method MkRow
and accessors column1
, column2
, so a simple "data table" could be constructed as a list of such records, simply enough.
One thing that the language doesn't natively support is lookup by accessor name. For example column1 :: Row -> Int
can only access a value of type Row
, since the column1
name is globally unique (for a discussion on modern techniques to deal with this, see the Advanced section below).
In addition to lookup, many data tasks require relational operations across pairs of data tables; algorithmically, these require lookups both across rows and columns, and there's nothing in Haskell's implementation of records that supports this.
There are a number of additional tasks that are routine in data analysis such as plotting, rendering the dataset to various tabular formats (CSV, database ...), and this library aims to support those too with a convenient syntax.
Advanced
Haskell offers a number of advanced workarounds for manipulating types, such as generic traversals, lookups, etc. A brief list of keywords is given in the following, for those inclined to dive into the rabbit hole.
Row polymorphism
Elm, Purescript etc.
OverloadedRecordFields
[1]
Row types
As you might know, the "row types" problem is well understood and has been explored in practice; discussing the various tradeoffs between approaches would be lengthy and quite technical (and your humble author is not too qualified to do full justice to the topic either).
In Haskell , the Frames [2] library and related ecosystem stands out as a full-featured dataframe implementation that does not compromise on type safety.
Heidi instead offers generic transformations from the source datatypes to uni-typed values (conceptually, each row is a Map String T
where data T = TInt Int | TChar Char
etc.), a domain in which it's convenient to perform lookups and similar operations.
Exploring further : vinyl [3], heterogeneous lists, sums-of-products ...
References
[1] OverloadedRecordFields : https://downloads.haskell.org/ghc/latest/docs/html/users_guide/glasgow_exts.html#record-field-selector-polymorphism
[2] Frames : https://hackage.haskell.org/package/Frames
[3] vinyl : https://hackage.haskell.org/package/vinyl
[4] generics-sop : https://hackage.haskell.org/package/generics-sop