"*The `pandas` library provides high-performance, easy-to-use data structures and data analysis tools. The main data structure is the `DataFrame`, which you can think of as an in-memory 2D table (like a spreadsheet, with column names and row labels). Many features available in Excel are available programmatically, such as creating pivot tables, computing columns based on other columns, plotting graphs, etc. You can also group rows by column value, or join tables much like in SQL. Pandas is also great at handling time series.*\n",
"* `Series` objects, that we will discuss now. A `Series` object is 1D array, similar to a column in a spreadsheet (with a column name and row labels).\n",
"* `DataFrame` objects. This is a 2D table, similar to a spreadsheet (with column names and row labels).\n",
"* `Panel` objects. You can see a `Panel` as a dictionary of `DataFrame`s. These are less used, so we will not discuss them here."
"Each item in a `Series` object has a unique identifier called the *index label*. By default, it is simply the rank of the item in the `Series` (starting at `0`) but you can also set the index labels manually:"
"To make it clear when you are accessing by label or by integer location, it is recommended to always use the `loc` attribute when accessing by label, and the `iloc` attribute when accessing by integer location:"
"But remember that you can access elements by integer location using the `iloc` attribute. This illustrates another reason why it's always better to use `loc` and `iloc` to access `Series` objects:"
"The resulting `Series` contains the union of index labels from `s2` and `s3`. Since `\"colin\"` is missing from `s2` and `\"charles\"` is missing from `s3`, these items have a `NaN` result value. (ie. Not-a-Number means *missing*).\n",
"Automatic alignment is very handy when working with data that may come from various sources with varying structure and missing items. But if you forget to set the right index labels, you can have surprising results:"
"Pandas makes it easy to plot `Series` data using matplotlib (for more details on matplotlib, check out the [matplotlib tutorial](tools_matplotlib.ipynb)). Just import matplotlib and call the `plot` method:"
"There are *many* options for plotting your data. It is not necessary to list them all here: if you need a particular type of plot (histograms, pie charts, etc.), just look for it in the excellent [Visualization](http://pandas.pydata.org/pandas-docs/stable/visualization.html) section of pandas' documentation, and look at the example code."
"Let's start by creating a time series using `timerange`. This returns a `DatetimeIndex` containing one datetime per hour for 12 hours starting on October 29th 2016 at 5:30pm."
"Note how the values have automatically been aggregated into 2-hour periods. If we look at the 6-8pm period, for example, we had a value of `5.1` at 6:30pm, and `6.1` at 7:30pm. After resampling, we just have one value of `5.6`, which is the mean of `5.1` and `6.1`. Computing the mean is the default behavior, but it is also possible to use a different aggregation function, for example we can decide to keep the minimum value of each period:"
"One solution is to fill the gaps by interpolating. We just call the `interpolate` method. The default is to use linear interpolation, but we can also select another method, such as cubic interpolation:"
"By default datetimes are *naive*: they are not aware of timezones, so 2016-10-30 02:30 might mean October 30th 2016 at 2:30am in Paris or in New York. We can make datetimes timezone *aware* by calling the `tz_localize` method:"
"Note that `-04:00` is now appended to all the datetimes. This means that these datetimes refer to [UTC](https://en.wikipedia.org/wiki/Coordinated_Universal_Time) - 4 hours.\n",
"\n",
"We can convert these datetimes to Paris time like this:"
"You may have noticed that the UTC offset changes from `+02:00` to `+01:00`: this is because France switches to winter time at 3am that particular night (time goes back to 2am). Notice that 2:30am occurs twice! Let's go back to a naive representation (if you log some data hourly using local time, without storing the timezone, you might get something like this):"
"Fortunately using the `ambiguous` argument we can tell pandas to infer the right DST (Daylight Saving Time) based on the order of the ambiguous timestamps:"
"The `asfreq` method lets us change the frequency of the `PeriodIndex`. All periods are lengthened or shortened accordingly. For example, let's convert all the quarterly periods to monthly periods (zooming in):"
"We can convert periods to timestamps by calling `to_timestamp`. By default this will give us the first day of each period, but by setting `how` and `freq`, we can get the last hour of each period:"
"Pandas also provides many other time-related functions that we recommend you check out in the [documentation](http://pandas.pydata.org/pandas-docs/stable/timeseries.html). To whet your appetite, here is one way to get the last business day of each month in 2016, at 9am:"
"A DataFrame object represents a spreadsheet, with cell values, column names and row index labels. You can define expressions to compute columns based on other columns, create pivot-tables, group rows, draw graphs, etc. You can see `DataFrame`s as dictionaries of `Series`.\n",
"* the `Series` were automatically aligned based on their index,\n",
"* missing values are represented as `NaN`,\n",
"* `Series` names are ignored (the name `\"year\"` was dropped),\n",
"* `DataFrame`s are displayed nicely in Jupyter notebooks, woohoo!"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"You can access columns pretty much as you would expect. They are returned as `Series` objects:"
]
},
{
"cell_type": "code",
"execution_count": 50,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people[\"birthyear\"]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"You can also get multiple columns at once:"
]
},
{
"cell_type": "code",
"execution_count": 51,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people[[\"birthyear\", \"hobby\"]]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"If you pass a list of columns and/or index row labels to the `DataFrame` constructor, it will guarantee that these columns and/or rows will exist, in that order, and no other column/row will exist. For example:"
"Another convenient way to create a `DataFrame` is to pass all the values to the constructor as an `ndarray`, or a list of lists, and specify the column names and row index labels separately:"
"As you may have noticed, the `stack` and `unstack` methods do not modify the object they apply to. Instead, they work on a copy and return that copy. This is true of most methods in pandas."
"The `loc` attribute lets you access rows instead of columns. The result is `Series` object in which the `DataFrame`'s column names are mapped to row index labels:"
]
},
{
"cell_type": "code",
"execution_count": 68,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people.loc[\"charles\"]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"You can also access rows by integer location using the `iloc` attribute:"
]
},
{
"cell_type": "code",
"execution_count": 69,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people.iloc[2]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"You can also get a slice of rows, and this returns a `DataFrame` object:"
]
},
{
"cell_type": "code",
"execution_count": 70,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people.iloc[1:3]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Finally, you can pass a boolean array to get the matching rows:"
]
},
{
"cell_type": "code",
"execution_count": 71,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people[np.array([True, False, True])]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"This is most useful when combined with boolean expressions:"
"Having to create a temporary variable `d6` is not very convenient. You may want to just chain the assigment calls, but it does not work because the `people` object is not actually modified by the first assignment:"
"But fear not, there is a simple solution. You can pass a function to the `assign` method (typically a `lambda` function), and this function will be called with the `DataFrame` as a parameter:"
"Assignment expressions are also supported, and contrary to the `assign` method, this does not create a copy of the `DataFrame`, instead it directly modifies it:"
"You can sort a `DataFrame` by calling its `sort_index` method. By default it sorts the rows by their index label, in ascending order, but let's reverse the order:"
]
},
{
"cell_type": "code",
"execution_count": 87,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people.sort_index(ascending=False)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Note that `sort_index` returned a sorted *copy* of the `DataFrame`. To modify `people` directly, we can set the `inplace` argument to `True`. Also, we can sort the columns instead of the rows by setting `axis=1`:"
]
},
{
"cell_type": "code",
"execution_count": 88,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people.sort_index(axis=1, inplace=True)\n",
"people"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"To sort the `DataFrame` by the values instead of the labels, we can use `sort_values` and specify the column to sort by:"
"Just like for `Series`, pandas makes it easy to draw nice graphs based on a `DataFrame`.\n",
"\n",
"For example, it is trivial to create a line plot from a `DataFrame`'s data by calling its `plot` method:"
]
},
{
"cell_type": "code",
"execution_count": 90,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"people.plot(kind = \"line\", x = \"body_mass_index\", y = [\"height\", \"weight\"])\n",
"plt.show()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"You can pass extra arguments supported by matplotlib's functions. For example, we can create scatterplot and pass it a list of sizes using the `s` argument of matplotlib's `scatter` function:"
]
},
{
"cell_type": "code",
"execution_count": 91,
"metadata": {
"collapsed": false,
"scrolled": true
},
"outputs": [],
"source": [
"people.plot(kind = \"scatter\", x = \"height\", y = \"weight\", s=[40, 120, 200])\n",
"plt.show()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Again, there are way too many options to list here: the best option is to scroll through the [Visualization](http://pandas.pydata.org/pandas-docs/stable/visualization.html) page in pandas' documentation, find the plot you are interested in and look at the example code."
"You can apply NumPy mathematical functions on a `DataFrame`: the function is applied to all values:"
]
},
{
"cell_type": "code",
"execution_count": 93,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"np.sqrt(grades)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Similarly, adding a single value to a `DataFrame` will add that value to all elements in the `DataFrame`. This is called *broadcasting*:"
]
},
{
"cell_type": "code",
"execution_count": 94,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"grades + 1"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Of course, the same is true for all other binary operations, including arithmetic (`*`,`/`,`**`...) and conditional (`>`, `==`...) operations:"
]
},
{
"cell_type": "code",
"execution_count": 95,
"metadata": {
"collapsed": false,
"scrolled": false
},
"outputs": [],
"source": [
"grades >= 5"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Aggregation operations, such as computing the `max`, the `sum` or the `mean` of a `DataFrame`, apply to each column, and you get back a `Series` object:"
]
},
{
"cell_type": "code",
"execution_count": 96,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"grades.mean()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"The `all` method is also an aggregation operation: it checks whether all values are `True` or not. Let's see during which months all students got a grade greater than `5`:"
]
},
{
"cell_type": "code",
"execution_count": 97,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"(grades > 5).all()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Most of these functions take an optional `axis` parameter which lets you specify along which axis of the `DataFrame` you want the operation executed. The default is `axis=0`, meaning that the operation is executed vertically (on each column). You can set `axis=1` to execute the operation horizontally (on each row). For example, let's find out which students had all grades greater than `5`:"
]
},
{
"cell_type": "code",
"execution_count": 98,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"(grades > 5).all(axis = 1)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"The `any` method returns `True` if any value is True. Let's see who got at least one grade 10:"
]
},
{
"cell_type": "code",
"execution_count": 99,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"(grades == 10).any(axis = 1)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"If you add a `Series` object to a `DataFrame` (or execute any other binary operation), pandas attempts to broadcast the operation to all *rows* in the `DataFrame`. This only works if the `Series` has the same size as the `DataFrame`s rows. For example, let's substract the `mean` of the `DataFrame` (a `Series` object) from the `DataFrame`:"
"We substracted `7.75` from all September grades, `8.75` from October grades and `7.50` from November grades. It is equivalent to substracting this `DataFrame`:"
"Similar to `Series`, when operating on multiple `DataFrame`s, pandas automatically aligns them by row index label, but also by column names. Let's create a `DataFrame` with bonus points for each person from October to December:"
"Looks like the addition worked in some cases but way too many elements are now empty. That's because when aligning the `DataFrame`s, some columns and rows were only present on one side, and thus they were considered missing on the other side (`NaN`). Then adding `NaN` to a number results in `NaN`, hence the result.\n",
"Dealing with missing data is a frequent task when working with real life data. Pandas offers a few tools to handle missing data.\n",
" \n",
"Let's try to fix the problem above. For example, we can decide that missing data should result in a zero, instead of `NaN`. We can replace all `NaN` values by a any value using the `fillna` method:"
]
},
{
"cell_type": "code",
"execution_count": 105,
"metadata": {
"collapsed": false,
"scrolled": true
},
"outputs": [],
"source": [
"(grades + bonus_points).fillna(0)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"It's a bit unfair that we're setting grades to zero in September, though. Perhaps we should decide that missing grades are missing grades, but missing bonus points should be replaced by zeros:"
]
},
{
"cell_type": "code",
"execution_count": 106,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"fixed_bonus_points = bonus_points.fillna(0)\n",
"fixed_bonus_points.insert(0, \"sep\", 0)\n",
"fixed_bonus_points.loc[\"alice\"] = 0\n",
"grades + fixed_bonus_points"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"That's much better: although we made up some data, we have not been too unfair.\n",
"\n",
"Another way to handle missing data is to interpolate. Let's look at the `bonus_points` `DataFrame` again:"
]
},
{
"cell_type": "code",
"execution_count": 107,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"bonus_points"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Now let's call the `interpolate` method. By default, it interpolates vertically (`axis=0`), so let's tell it to interpolate horizontally (`axis=1`)."
]
},
{
"cell_type": "code",
"execution_count": 108,
"metadata": {
"collapsed": false,
"scrolled": false
},
"outputs": [],
"source": [
"bonus_points.interpolate(axis=1)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Bob had 0 bonus points in October, and 2 in December. When we interpolate for November, we get the mean: 1 bonus point. Colin had 1 bonus point in November, but we do not know how many bonus points he had in September, so we cannot interpolate, this is why there is still a missing value in October after interpolation. To fix this, we can set the September bonus points to 0 before interpolation."
"Great, now we have reasonable bonus points everywhere. Let's find out the final grades:"
]
},
{
"cell_type": "code",
"execution_count": 110,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"grades + better_bonus_points"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"It is slightly annoying that the September column ends up on the right. This is because the `DataFrame`s we are adding do not have the exact same columns (the `grades` `DataFrame` is missing the `\"dec\"` column), so to make things predictable, pandas orders the final columns alphabetically. To fix this, we can simply add the missing column before adding:"
]
},
{
"cell_type": "code",
"execution_count": 111,
"metadata": {
"collapsed": false,
"scrolled": true
},
"outputs": [],
"source": [
"grades[\"dec\"] = np.nan\n",
"final_grades = grades + better_bonus_points\n",
"final_grades"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"There's not much we can do about December and Colin: it's bad enough that we are making up bonus points, but we can't reasonably make up grades (well I guess some teachers probably do). So let's call the `dropna` method to get rid of rows that are full of `NaN`s:"
"Similar to the SQL language, pandas allows grouping your data into groups to run calculations over each group.\n",
"\n",
"First, let's add some extra data about each person so we can group them, and let's go back to the `final_grades` `DataFrame` so we can see how `NaN` values are handled:"
"That was easy! Note that the `NaN` values have simply been skipped when computing the means."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Pivot tables\n",
"Pandas supports spreadsheet-like [pivot tables](https://en.wikipedia.org/wiki/Pivot_table) that allow quick data summarization. To illustrate this, let's create a simple `DataFrame`:"
"Now we can call the `pivot_table` function for this `DataFrame`, asking to group by the `name` column. By default, `pivot_table` computes the `mean` of each numeric column:"
"We can change the aggregation function by setting the `aggfunc` attribute, and we can also specify the list of columns whose values will be aggregated:"
"When dealing with large `DataFrames`, it is useful to get a quick overview of its content. Pandas offers a few functions for this. First, let's create a large `DataFrame` with a mix of numeric values, missing values and text values. Notice how Jupyter displays only the corners of the `DataFrame`:"
"Pandas can save `DataFrame`s to various backends, including file formats such as CSV, Excel, JSON, HTML and HDF5, or to a SQL database. Let's create a `DataFrame` to demonstrate this:"
"Note that the index is saved as the first column (with no name) in a CSV file, as `<th>` tags in HTML and as keys in JSON.\n",
"\n",
"Saving to other formats works very similarly, but some formats require extra libraries to be installed. For example, saving to Excel requires the openpyxl library:"
"As you might guess, there are similar `read_json`, `read_html`, `read_excel` functions as well. We can also read data straight from the Internet. For example, let's load all U.S. cities from [simplemaps.com](http://simplemaps.com/):"
"There are more options available, in particular regarding datetime format. Check out the [documentation](http://pandas.pydata.org/pandas-docs/stable/io.html) for more details."
"One powerful feature of pandas is it's ability to perform SQL-like joins on `DataFrame`s. Various types of joins are supported: inner joins, left/right outer joins and full joins. To illustrate this, let's start by creating a couple simple `DataFrame`s:"
"Note that both `DataFrame`s have a column named `state`, so in the result they got renamed to `state_x` and `state_y`.\n",
"\n",
"Also, note that Cleveland, Salt Lake City and Houston were dropped because they don't exist in *both* `DataFrame`s. This is the equivalent of a SQL `INNER JOIN`. If you want a `FULL OUTER JOIN`, where no city gets dropped and `NaN` values are added, you must specify `how=\"outer\"`:"
"Of course `LEFT OUTER JOIN` is also available by setting `how=\"left\"`: only the cities present in the left `DataFrame` end up in the result. Similarly, with `how=\"right\"` only cities in the right `DataFrame` appear in the result. For example:"
"If the key to join on is actually in one (or both) `DataFrame`'s index, you must use `left_index=True` and/or `right_index=True`. If the key column names differ, you must use `left_on` and `right_on`. For example:"
"Note that this operation aligned the data horizontally (by columns) but not vertically (by rows). In this example, we end up with multiple rows having the same index (eg. 3). Pandas handles this rather gracefully:"
"Notice that when a column does not exist in a `DataFrame`, it acts as if it was filled with `NaN` values. If we set `join=\"inner\"`, then only columns that exist in *both* `DataFrame`s are returned:"
"In this case it really does not make much sense because the indices do not align well (eg. Cleveland and San Francisco end up on the same row, because they shared the index label `3`). So let's reindex the `DataFrame`s by city name before concatenating:"
"This looks a lot like a `FULL OUTER JOIN`, except that the `state` columns were not renamed to `state_x` and `state_y`, and the `city` column is now the index."
"It is quite frequent to have values that represent categories, for example `1` for female and `2` for male, or `\"A\"` for Good, `\"B\"` for Average, `\"C\"` for Bad. These categorical values can be hard to read and cumbersome to handle, but fortunately pandas makes it easy. To illustrate this, let's take the `city_pop` `DataFrame` we created earlier, and add a column that represents a category:"
"Right now the `eco_code` column is full of apparently meaningless codes. Let's fix that. First, we will create a new categorical column based on the `eco_code`s:"
"As you probably noticed by now, pandas is quite a large library with *many* features. Although we went through the most important features, there is still a lot to discover. Probably the best way to learn more is to get your hands dirty with some real-life data. It is also a good idea to go through pandas' excellent [documentation](http://pandas.pydata.org/pandas-docs/stable/index.html), in particular the [Cookbook](http://pandas.pydata.org/pandas-docs/stable/cookbook.html)."
