Changelog

2.0-beta-47

• UUIDs are now supported as datatypes. This includes parsing them from strings out of csv and xlsx files and as a fully supported object in mapseq pathways.

2.0-beta-46

• Conversion of instant->zoned-date-time uses UTC zone instead of system zone.

2.0-beta-45

• issue-76 - quartiles for datetime types in descriptive stats
• issue-71 - Added shape function to main dataset api. Returns shape in row major format.
• issue-69 - Columns elide missing values during print operation.

2.0-beta-44

• drop-rows on an empty set is a noop.
• tech.ml.dataset.column/stats was wrong for columns with missing values.

2.0-beta-43

• tech.v2.datatype was causing double-read on boolean readers.
• issue-72 - added max-num-columns because csv and tsv files with more than 512 columns were failing to parse. New default is 8192.
• issue-70 - The results of any join have two maps in their metadata - :left-column-names - map of original left-column-name->new-column-name. :right-column-names - map of original right-column-name->new-column-name.

2.0-beta-42

• n-initial-skip-rows works with xlsx spreadsheets.
• assoc, dissoc implemented in the main dataset namespace.

2.0-beta-41

• issue-67 - Various tech.v2.datatype.functional functions are updated to be more permissive about their inputs and cast the result to the appropriate datatype.

2.0-beta-40

• issue-65 - datetimes in mapseqs were partially broken.
• tech.v2.datatype.functional will now change the datatype appropriately on a lot of unary math operations. So for instance calling sin, cos, log, or log1p on an integer reader will now return a floating point reader. These methods used to throw.
• subtle bug in the ->reader method defined for object arrays meant that sometimes attempting math on object columns would fail.
• tech.ml.dataset/column-cast - Changes the column datatype via a an optionally privided cast function. This function is powerful - it will correctly convert packed types to their string representation, it will use the parsing system on string columns and it uses the same complex datatype argument as tech.ml.dataset.column/parse-column:

user> (doc ds/column-cast)
-------------------------
tech.ml.dataset/column-cast
([dataset colname datatype])
  Cast a column to a new datatype.  This is never a lazy operation.  If the old
  and new datatypes match and no cast-fn is provided then dtype/clone is called
  on the column.

  colname may be a scalar or a tuple of [src-col dst-col].

  datatype may be a datatype enumeration or a tuple of
  [datatype cast-fn] where cast-fn may return either a new value,
  the :tech.ml.dataset.parse/missing, or :tech.ml.dataset.parse/parse-failure.
  Exceptions are propagated to the caller.  The new column has at least the
  existing missing set if no attempt returns :missing or :cast-failure.
  :cast-failure means the value gets added to metadata key :unparsed-data
  and the index gets added to :unparsed-indexes.


  If the existing datatype is string, then tech.ml.datatype.column/parse-column
  is called.

  Casts between numeric datatypes need no cast-fn but one may be provided.
  Casts to string need no cast-fn but one may be provided.
  Casts from string to anything will call tech.ml.dataset.column/parse-column.
user> (def stocks (ds/->dataset "test/data/stocks.csv" {:key-fn keyword}))

#'user/stocks
user> (ds/head stocks)
test/data/stocks.csv [5 3]:

| :symbol |      :date | :price |
|---------+------------+--------|
|    MSFT | 2000-01-01 |  39.81 |
|    MSFT | 2000-02-01 |  36.35 |
|    MSFT | 2000-03-01 |  43.22 |
|    MSFT | 2000-04-01 |  28.37 |
|    MSFT | 2000-05-01 |  25.45 |
user> (ds/head stocks)
test/data/stocks.csv [5 3]:

| :symbol |      :date | :price |
|---------+------------+--------|
|    MSFT | 2000-01-01 |  39.81 |
|    MSFT | 2000-02-01 |  36.35 |
|    MSFT | 2000-03-01 |  43.22 |
|    MSFT | 2000-04-01 |  28.37 |
|    MSFT | 2000-05-01 |  25.45 |
user> (take 5 (stocks :price))
(39.81 36.35 43.22 28.37 25.45)
user> (take 5 ((ds/column-cast stocks :price :string) :price))
("39.81" "36.35" "43.22" "28.37" "25.45")
user> (take 5 ((ds/column-cast stocks :price [:int32 #(Math/round (double %))]) :price))
(40 36 43 28 25)
user>

2.0-beta-29

• renamed 'column-map' to 'column-name->column-map'. This is a public interface change and we do apologize!
• added 'column-map' which maps a function over one or more columns. The result column has a missing set that is the union of the input columns' missing sets:

user> (-> (ds/->dataset [{:a 1} {:b 2.0} {:a 2 :b 3.0}])
          (ds/column-map
           :summed
           (fn ^double [^double lhs ^double rhs]
             (+ lhs rhs))
           :a :b))
_unnamed [3 3]:

| :a |    :b | :summed |
|----+-------+---------|
|  1 |       |         |
|    | 2.000 |         |
|  2 | 3.000 |   5.000 |
user> (tech.ml.dataset.column/missing
       (*1 :summed))
#{0,1}

2.0-beta-38

• [issue-64] - more tests revealed more problems with concat with different column types.
• added tech.v2.datatype/typed-reader-map where the result datatype is derived from the input datatypes of the input readers. The result of map-fn is unceremoniously coerced to this datatype -

user> (-> (ds/->dataset [{:a 1.0} {:a 2.0}])
               (ds/update-column
                :a
                #(dtype/typed-reader-map (fn ^double [^double in]
                                           (if (< in 2.0) (- in) in))
                                         %)))
_unnamed [2 1]:

|     :a |
|--------|
| -1.000 |
|  2.000 |

• Cleaned up the tech.datatype widen datatype code so it models a property type graph with clear unification rules (where the parent are equal else :object).

2.0-beta-37

• [issue-64] - concat columns with different datatypes does a widening. In addition, there are tested pathways to change the datatype of a column without changing the missing set.
• unroll-column takes an optional argument :indexes? that will record the source index in the entry the unrolled data came from.

2.0-beta-36

• generic column data lists now support .addAll

2.0-beta-35

• tech.datatype - all readers are marked as sequential.
• unroll-column - Given a column that may container either iterable or scalar data, unroll it so it only contains scalar data duplicating rows.
• Issue 61 - Empty bitsets caused exceptions.
• Issue 62 - IP addresses parsed as durations.

2.0-beta-34

• Major speed (100x+) improvements to tech.ml.dataset.column/unique and especially `tech.ml.dataset.pipeline/string->number.

2.0-beta-33

• tech.v2.datatype namespace has a new function - make-reader - that reifies a reader of the appropriate type. This allows you to make new columns that have nontrivial translations and datatypes much easier than before.
• tech.v2.datatype namespace has a new function - ->typed-reader - that typecasts the incoming object into a reader of the appropriate datatype. This means that .read calls will be strongly typed and is useful for building up a set of typed variables before using make-reader above.
• Some documentation on the implications of columns, readers, and datatypes.

2.0-beta-32

• Issue 52 - CSV columns with empty column names get named after their index. Before they would cause an exception.
• tech.datatype added a method to transform a reader into a persistent-vector-like object that derives from clojure.lang.APersistentVector and thus gains benefit from the excellent equality and hash semantics of persistent vectors.

2.0-beta-31

• Fixed #38 - set-missing/remove-rows can take infinite seqs - they are trimmed to dataset length.
• Fixed #47 - Added columnwise-concat which is a far simpler version of dplyr's https://tidyr.tidyverse.org/reference/pivot_longer.html. This is implemented efficiently in terms of indexed reader concatentation and as such should work on tables of any size.
• Fixed #57 - BREAKING PUBLIC API CHANGES - We are getting more strict on the API - if a function is dataset-last (thus appropriate for ->>) then any options must be passed before the dataset. Same is true for the set of functions that are dataset first. We will be more strict about this from now on.

2.0-beta-30

• Parsing datetime types now works if the column starts with missing values.
• An efficient formulation of java.util.map is introduced for when you have a bitmap of keys and a single value: tech.v2.datatype.bitmap/bitmap-value->bitmap-map. This is used for replace-missing type operations.

2.0-beta-29

• brief now does not return missing values. Double or float NaN or INF values from a mapseq result in maps with fewer keys.
• Set of columns used for default descriptive stats is reduced to original set as this fits on a small repl nicely. Possible to override. brief overrides this to provide defaults to get more information.
• unique-by returns indexes in order.
• Fixed #51 - mapseq parsing now follows proper number tower.

2.0-beta-28

• Fixed #36 - use key-fn uniformly across all loaded datatypes
• Fixed #45 - select can take a map. This does a selection and a projection to new column names.
• Fixed #41 - boolean columns failed to convert to doubles.
• Fixed #44 - head,tail,shuffle,rand-nth,sample all implemented in format appropriate for ->> operators.

2.0-beta-27

• Update tech.datatype with upgraded and fewer dependencies.
  • asm 7.1 (was 7.0)
  • org.clojure/math.combinatorics 1.6 (was 1.2)
  • org.clojure/test.check 1.0.0

2.0-beta-25

• Optimized filter. Record of optimization is on zulip. Synopsis is a speedup of like 10-20X depending on how much work you want to do :-). The base filter pathway has a speedup of around 2-4X.

2.0-beta-23

• Updated description stats to provide list of distinct elements for categorical columns of length less than 21.
• Updated mapseq system to provide nil values for missing data as opposed to the specific column datatype's missing value indicator. This can be overridden by passing in :missing-nil? false as an option.
• Added brief function to main namespace so you can get a nice brief description of your dataset when working from the REPL. This prints out better than descriptive-stats.

2.0-beta-21

• loading jsons files found issues with packing.
• optimized conversion to/from maps.

2.0-beta-20

• sort-by works with generic comparison fns.

2.0-beta-19

• descriptive stats works with mixed column name types
• argsort is now used for all sort functions
• -> versions of sort added so you can sort in -> pathways
• instants and such can used for sorting

Added Functions

• column->dataset - map a transform function over a column and return a new dataset from the result. It is expected the transform function returns a map.
• drop-rows, select-rows, drop-columns - more granular select calls.
• append-columns - append a list of columns to a dataset. Used with column->dataset.
• column-labeled-mapseq - Create a sequence of maps with a :value and :label members. this flattens the dataset by producing Y maps per row instead of 1 map per row where the maps themselves are labeled with the value in their :value member. This is useful to building vega charts.
• ->distinct-by-column - take the first row where a given key is present. The arrow form of this indicats the dataset is the first argument.
• ->sort-by, ->sort-by-column - Forms of these functions for using in (->) dataflows.
• interpolate-loess - Produce a new column from a given pair of columns using loess interpolation to create the column. The interpolator is saved as metadata on the new column.

2.0-beta-16

• Missing a datetime datatype for parse-str and add-to-container! means a compile time error. Packed durations can now be read from mapseqs.

2.0-beta-15

• Descriptive stats now works with instants.

2.0-beta-14

• Descriptive stats now works with datetime types.

2.0-beta-12

• Support for parsing and working with durations. Strings that look like times - "00:00:12" will be parsed into hh:mm:ss durations. The value can have a negative sign in front. This is in addition to the duration's native serialization string type.
• Added short test for tensors in datasets. This means that the venerable print-table is no longer enough as it doesn't account for multiline strings and thus datatets with really complex things will not print correctly for a time.

2.0-beta-11

• Various fixes related to parsing and working with open data.
• tech.ml.dataset.column/parse-column - given a string column that failed to parse for some reason, you can force the system to attempt to parse it using, for instance, relaxed parsing semantics where failures simply record the failure in metadata.
• relaxed parsing in general is supported across all input types.

0.26

Added

• rolling (rolling windows of computed functions - math operation)
• dataset/dssort-by
• dataset/ds-take-nth

0.22

Added

• PCA