استخراج الميزات الزمنية

A raw datetime64 column is opaque to most machine learning algorithms — they cannot understand that 2024-12-25 is a holiday or that Friday drives higher retail sales. Temporal feature engineering extracts meaningful components from datetime columns — year, month, day, weekday, hour — as separate numeric or categorical columns that models and analysis functions can act on directly.

When a DataFrame column has datetime64 dtype, the .dt accessor exposes all datetime properties and methods vectorially. Instead of applying a Python function row by row, you write df['date'].dt.year and Pandas computes the year for every row in a single fast operation. The same accessor works on timedelta64 columns with different properties.

The most commonly used .dt properties for date components are: .dt.year, .dt.month, .dt.day, .dt.day_of_week (0=Monday through 6=Sunday), .dt.day_of_year, .dt.week or .dt.isocalendar().week, .dt.quarter, and .dt.days_in_month. Each returns a Series of integers that can be used directly as features.

For datetime columns that include time information (not just date), use .dt.hour, .dt.minute, .dt.second, and .dt.microsecond. These are critical for intraday analysis, user behaviour modelling (morning vs evening patterns), and any system where the time of day drives the target variable.

For human-readable labels, use .dt.day_name() and .dt.month_name() to get string names like 'Monday' and 'January'. These are useful for display in reports and for creating categorical features that encode weekday/month effects. You can also pass a locale to get names in other languages.

A very common feature is a boolean flag indicating whether a date falls on a weekend. day_of_week >= 5 gives True for Saturday (5) and Sunday (6). For more sophisticated business calendar features, numpy.busdaycalendar handles holidays, but for most cases the simple weekday/weekend split captures most of the effect.

The .dt accessor also provides boolean properties that identify whether a date is at the start or end of a period: .dt.is_month_start, .dt.is_month_end, .dt.is_quarter_start, .dt.is_quarter_end, .dt.is_year_start, and .dt.is_year_end. These are useful for detecting reporting periods, billing cycles, and seasonal boundary effects.

Month and hour are cyclical — December (12) is close to January (1), and 23:00 is close to 00:00. Treating them as linear integers (1-12 or 0-23) misleads most models. Encode cyclical features using sine and cosine transforms: sin(2π × value / max_value) and cos(2π × value / max_value). This embeds the circular structure in two numeric columns that models can use correctly.

Subtracting two datetime columns produces a timedelta64 Series. The .dt accessor on timedelta Series exposes .dt.days, .dt.seconds, and .dt.total_seconds(). This lets you compute event durations, time since last purchase, and age in days — all as numeric columns ready for further analysis.

When the datetime is the DataFrame's index (a DatetimeIndex) rather than a column, access the same properties directly on the index without .dt: df.index.year, df.index.month, df.index.day_of_week, etc. You can add these as new columns directly: df['month'] = df.index.month.

Here is a complete temporal feature engineering function that takes a DataFrame with a datetime column and returns it enriched with the most useful temporal features. This pattern is directly applicable at the start of any machine learning or analysis pipeline working with time-stamped event data.

Test your understanding of extracting temporal features from datetime columns.

In this lesson you learned: the .dt accessor gives vectorised access to all datetime properties; common extracted features include year, month, day, day_of_week, hour, quarter; cyclical encoding with sine/cosine handles the circular nature of hours and months; and timedelta arithmetic lets you compute durations in days or seconds. Next up we explore Matplotlib's figure and axes architecture for creating publication-quality charts.