saturn

This dataset is a 2,000,072-row movie tag log from MovieLens (ml-32m/tags.csv) with four columns: a free-text tag, a timestamp, a userId, and a movieId. The tag column is the most interesting feature — it has only 140,981 unique values across 2M rows (a 92.95% duplicate rate) and 52.47% of tags are a single word, with 'sci-fi', 'atmospheric', and 'action' leading the list. The timestamp column is left-skewed (skew −1.22) toward more recent activity, suggesting tagging picked up in later years, and userId shows that tagging is concentrated among a subset of users (only 15,848 distinct userIds for 2M rows). Start by looking at the top tags and the timestamp distribution to understand what users tag and when.

This is almost certainly a user identifier stored as an integer, with 15,848 unique values spread across roughly 2M rows (so ~126 rows per user on average). Values range from 22 to 162,279 with no nulls or zeros, and the distribution is near-symmetric (skew 0.12, kurtosis -0.26), consistent with sparsely allocated account IDs rather than a meaningful numeric quantity. The 157,830 flagged outliers (7.9%) are an artefact of treating IDs as numeric — they are not statistical anomalies in any business sense.

Integer movieId spanning 1 to 292629 across 2,000,072 rows with 51,323 unique values and no nulls. The cardinality and bounded integer range indicate a foreign key into a movies table rather than a measured quantity; summary statistics like mean (71893.26) and skew (0.82) reflect ID allocation, not a meaningful distribution.

This column holds short descriptive tags applied to items (likely films), with frequent values like 'sci-fi', 'atmospheric', 'action', and 'comedy'. Entries are tiny — mean length 11.1 characters and word_mean 1.67 — and 52.5% are a single word. Despite 2,000,072 rows there are only 140,981 distinct tags and a 92.95% duplicate rate (1,859,091 duplicates), so the same labels recur heavily across items.

Unix epoch timestamps spanning roughly 1135429210 (late 2005) to 1697154983 (late 2023), with a median of 1574071062 (late 2019). The distribution is left-skewed (skew -1.22) and 6.39% of rows fall outside the IQR fence, indicating a long tail of older records before activity concentrated post-2018. With 1291250 unique values across 2000072 rows there is meaningful repetition but no nulls or zeros.