Analysis of Pitchfork Music Reviews
Introduction
Hello! Welcome to an analysis of 18 years of everyone’s favorite cool music website, pitchfork.com. This data was generously uploaded by Nolan Conway at Kaggle and can be found at https://www.kaggle.com/nolanbconaway/pitchfork-data.
There’s a lot to explore here. Have album scores changed overtime? Do the sages of Pitchfork betray preferences for different genres, artists or labels? Do different classes of writers dole out different judgment? With more than 18,000 rows in this dataset, we’ll soon find out!
A note on the data: The fact that this was originally stored in a SQLite file meant that when the tables were merged into the dataframe, there was a fair number of duplicate rows, as a new row was created if an album had more than one label or genre. That sometimes resulted in single albums containing dozens of rows. (One of Radiohead’s albums had 32 rows lol.) Thus, there could be a slight loss of resolution when analyses include genres or labels.
With that said, let’s begin.
Summary statistics
Here’s a bird’s-eye view of our dataset:
## 'data.frame': 18389 obs. of 18 variables:
## $ reviewid : int 1 6 7 8 10 11 12 15 16 17 ...
## $ title : Factor w/ 17832 levels "",">>","$","*",..: 17703 11392 10562 1999 277 6473 284 8635 6369 326 ...
## $ artist : Factor w/ 8715 levels "","!!!","+/-",..: 55 47 47 47 52 54 57 74 74 74 ...
## $ url : Factor w/ 18389 levels "http://pitchfork.com/reviews/albums/10000-steingarten/",..: 10058 15999 16667 17392 1015 2028 3039 6169 7156 8089 ...
## $ score : num 3.4 7.4 8.1 7.2 8 7 7 6.7 7.5 7.3 ...
## $ best_new_music: int 0 0 0 0 0 0 0 0 0 0 ...
## $ author : Factor w/ 423 levels "aaron leitko",..: 247 132 83 367 283 193 68 29 215 114 ...
## $ author_type : Factor w/ 16 levels "","assistant editor",..: 15 8 1 1 1 8 1 8 8 1 ...
## $ pub_date : Factor w/ 3950 levels "1999-01-05","1999-01-12",..: 944 618 402 1040 395 522 174 664 592 174 ...
## $ pub_weekday : int 0 1 3 6 1 0 6 6 6 6 ...
## $ pub_day : int 30 8 16 6 7 4 31 15 2 31 ...
## $ pub_month : int 8 4 5 2 5 11 12 6 3 12 ...
## $ pub_year : int 2004 2003 2002 2005 2002 2002 2000 2003 2003 2000 ...
## $ genre : Factor w/ 10 levels "","electronic",..: 10 2 2 2 6 10 7 9 9 10 ...
## $ label : Factor w/ 3404 levels "","00:02:59",..: 2550 2674 731 731 3101 350 2707 766 2377 2377 ...
## $ clean : Factor w/ 18376 levels "[]","['0101', '0103', '0107', '0108', '0113', 'and', '0115', 'since', 'all', 'jj', 'choose', 'to', 'show', 'of', 'themselves', 'is',"| __truncated__,..: 60 2469 4814 47 59 6034 14724 2203 134 1705 ...
## $ diversity : num 0.59 0.594 0.625 0.666 0.634 ...
## $ length : int 549 672 635 365 320 499 600 811 550 843 ...This is a reasonably large dataframe. It has 18 columns, and many of the rows have hundreds or thousands of unique values. We’ll take summary values of everything but the ‘content’ column, which gives us very large, very unquantitative criticism.
## reviewid title artist
## Min. : 1 dj-kicks : 20 various artists : 687
## 1st Qu.: 7444 ii : 17 guided by voices: 23
## Median :12806 live : 9 david bowie : 21
## Mean :12346 iii : 8 the beatles : 21
## 3rd Qu.:17659 greatest hits: 7 mogwai : 20
## Max. :22745 dj kicks : 6 of montreal : 20
## (Other) :18322 (Other) :17597
## url
## http://pitchfork.com/reviews/albums/10000-steingarten/ : 1
## http://pitchfork.com/reviews/albums/10001-touch-up/ : 1
## http://pitchfork.com/reviews/albums/10002-seven-sisters/ : 1
## http://pitchfork.com/reviews/albums/10003-pagoda/ : 1
## http://pitchfork.com/reviews/albums/10004-all-things-forests/ : 1
## http://pitchfork.com/reviews/albums/10005-scribble-mural-comic-journal/: 1
## (Other) :18383
## score best_new_music author
## Min. : 0.000 Min. :0.00000 joe tangari : 815
## 1st Qu.: 6.400 1st Qu.:0.00000 stephen m. deusner: 725
## Median : 7.200 Median :0.00000 ian cohen : 699
## Mean : 7.006 Mean :0.05128 brian howe : 500
## 3rd Qu.: 7.800 3rd Qu.:0.00000 mark richardson : 476
## Max. :10.000 Max. :1.00000 stuart berman : 445
## (Other) :14729
## author_type pub_date pub_weekday
## contributor :12420 2000-03-31: 15 Min. :0.000
## : 3904 2000-04-30: 14 1st Qu.:1.000
## senior editor : 486 2001-03-31: 13 Median :2.000
## executive editor : 475 1999-04-20: 12 Mean :2.107
## senior staff writer: 439 2001-02-20: 11 3rd Qu.:3.000
## contributing editor: 210 1999-06-08: 10 Max. :6.000
## (Other) : 455 (Other) :18314
## pub_day pub_month pub_year genre
## Min. : 1.00 Min. : 1.000 Min. :1999 rock :7815
## 1st Qu.: 8.00 1st Qu.: 3.000 1st Qu.:2005 electronic :2900
## Median :15.00 Median : 6.000 Median :2009 :2365
## Mean :15.53 Mean : 6.283 Mean :2009 rap :1413
## 3rd Qu.:23.00 3rd Qu.: 9.000 3rd Qu.:2013 experimental:1141
## Max. :31.00 Max. :12.000 Max. :2017 pop/r&b :1128
## (Other) :1627
## label diversity length
## self-released: 419 Min. :0.0000 Min. : 0.0
## drag city : 263 1st Qu.:0.5448 1st Qu.: 498.0
## sub pop : 261 Median :0.5778 Median : 604.0
## thrill jockey: 241 Mean :0.5771 Mean : 650.1
## merge : 231 3rd Qu.:0.6108 3rd Qu.: 746.0
## warp : 210 Max. :0.8889 Max. :3688.0
## (Other) :16764Already, some points arise:
Compliations abound, with “Various Artists” having 687 albums. The single most reviewed group/artist is Guided By Voices—I don’t even know what that is.
An album’s inclusion in Pitchfork’s annual end-of-year Best New Music is indicated by a 1, rather than by the ranking in the BNM list, which is too bad, but good on Mr. Conway to think of including this in the scraping.
All the top publishing days for the website in the summary function come up from the early years of Pitchfork, which was founded in January 1999.
There are individual writers who have contributed hundreds and hundreds of reviews.
The DJ Kicks mixtapes are the most reviewed series of album.
Too many artists make album series titled with Roman numerals.
Of Monteal’s 20 albums seem excessive.
Let’s take a look at how the scores fall:
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.000 6.400 7.200 7.006 7.800 10.000Pitchfork, it seems, grades on a curve—if you release an album, you’re likely to get something near a C.
Breaking down scoring factors
Genre
The fun of the dataset is how many angles it provides to look at scores. Let’s start by taking a look at genres.
There doesn’t seem to be any screamingly conscpicuous prejudice for or against any particular genre in terms of score, just a clear signal that rock predominates in terms of sheer number of albums reviewed. Here’s the same info broken down by year:
##
## Attaching package: 'plotly'## The following object is masked from 'package:MASS':
##
## select## The following object is masked from 'package:ggplot2':
##
## last_plot## The following object is masked from 'package:stats':
##
## filter## The following object is masked from 'package:graphics':
##
## layout
## # A tibble: 10 x 3
## genre median_scores mean_scores
## <fctr> <dbl> <dbl>
## 1 7.20 7.017463
## 2 electronic 7.20 6.958034
## 3 experimental 7.50 7.325416
## 4 folk/country 7.40 7.233116
## 5 global 7.75 7.431250
## 6 jazz 7.60 7.459273
## 7 metal 7.40 7.046050
## 8 pop/r&b 7.10 6.905940
## 9 rap 7.10 6.926327
## 10 rock 7.20 6.957223However, the first time I made a pass at this data, I lazily dismissed the possibility that there was any real distinction in genre scores—the mean scores don’t diverge by more than a single standard deviation from one another. But despite the ostensibly small distinctions, they’re still worth testing. The challenge is that there are more albums of types other than jazz or global by orders of magnitude. But, with the magic of R, we can sample the dataset such that we get a number of reviews that are equal to the number of ‘global’ albums, 144, and run some tests.
## # A tibble: 9 x 3
## genre median.score mean.score
## <fctr> <dbl> <dbl>
## 1 electronic 7.10 6.945833
## 2 experimental 7.60 7.362500
## 3 folk/country 7.50 7.336806
## 4 global 7.75 7.431250
## 5 jazz 7.60 7.459273
## 6 metal 7.40 6.889583
## 7 pop/r&b 7.20 6.946528
## 8 rap 7.00 6.685417
## 9 rock 7.05 6.927778These are the descriptive stats of our sample, which are close enough to the unsampled data that we can test them comfortably. What do we get?
##
## Call:
## lm(formula = score ~ genre, data = testing)
##
## Residuals:
## Min 1Q Median 3Q Max
## -6.8465 -0.5593 0.2375 0.8104 3.0722
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 6.9458333 0.1020924 68.035 < 2e-16 ***
## genreexperimental 0.4166667 0.1443805 2.886 0.003962 **
## genrefolk/country 0.3909722 0.1443805 2.708 0.006852 **
## genreglobal 0.4854167 0.1443805 3.362 0.000794 ***
## genrejazz 0.5134394 0.1260184 4.074 4.87e-05 ***
## genremetal -0.0562500 0.1443805 -0.390 0.696894
## genrepop/r&b 0.0006944 0.1443805 0.005 0.996163
## genrerap -0.2604167 0.1443805 -1.804 0.071493 .
## genrerock -0.0180556 0.1443805 -0.125 0.900497
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.225 on 1418 degrees of freedom
## Multiple R-squared: 0.04874, Adjusted R-squared: 0.04337
## F-statistic: 9.081 on 8 and 1418 DF, p-value: 3.106e-12## Analysis of Variance Table
##
## Response: score
## Df Sum Sq Mean Sq F value Pr(>F)
## genre 8 109.04 13.6302 9.0814 3.106e-12 ***
## Residuals 1418 2128.26 1.5009
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1It looks like genre matters. And if we look at the plot of residuals…
…we see a pretty normal distribution, meaning that our linear model isn’t clearly violating any rules.
##
## Call:
## lm(formula = score ~ genre, data = testing)
##
## Coefficients:
## (Intercept) genreexperimental genrefolk/country
## 6.9458333 0.4166667 0.3909722
## genreglobal genrejazz genremetal
## 0.4854167 0.5134394 -0.0562500
## genrepop/r&b genrerap genrerock
## 0.0006944 -0.2604167 -0.0180556The intercept is what an album would receive if it were theoretically genre-less, pop nets you .02 points less for being pop, jazz .58 points more for being jazz, and so on. The question is how reliable this kind of thing is in this case. The p-value indicates there’s a one-in-gajillion chance that the difference in scores between genres was random, and yet p-values are understandably falling out of fashion. If this were an intro stats class, we might be happy with our p-value, deem genre significant, and call it a day. But there isn’t more than about a half-point difference between genres in the their aggregate scores, and our sample size is quite small.
I would say it’s too ambiguous to conclusively determine that there’s a truly meaningful distinction here. What I AM curious to look at is how the genres of albums Pitchfork has reviewed has changed over the years. Someone recently remarked to me that Pitchfork has “a pre-woke phase and a woke phase”, so I’m wondering if this bears out in the genre allotment.
## # A tibble: 1 x 1
## n
## <int>
## 1 34596
This data doesn’t include albums for which there was no genre info, and of course, overlapping genres were removed when we did the cleaning. Still, what remains is considered a very large sample of the dataset. Between this graph and the last, we can see a few things:
- Rock predominantes, although
- every other genre appears to be growing vis-a-vis 2000, particularly rap and electronic.
- Pitchfork reviews so few jazz albums that it probably should not review jazz.
- Speaking of which, there’s no contemporary classical (maybe it fell under ‘experimental’ lol).
- Metal should probably just be merged into the rock category, so if we mentally do that, then rock is even bigger.
- Pitchfork has peaked maybe?
Author Type
We can also examine whether differnet types of writers are more exacting in their critiques.
This is a pretty but unintelligible graph—the only thing that comes through the noise is how much Pitchfork relies on freelancers. Let’s look only at each individual category.
Similarly, the problem with this graph is that Pitchfork’s masthead, like that of every other publication, is so clogged with honorifics and petty editorial distinctions that RStudio practically runs out of space on the color spectrum to represent every editorial role. Viewed glancingly, the graph makes the associate reviews editor ca. 2016 look far more generous than th editor-in-chief ca. 2000. The problem, however, is that this graph doesn’t really incorporate the sense of scale of author type while reminaing readable. We can do that as such:
It’s not reasonable to compare the average scores of anyone but contributors, and their average scores never fall beyond even a single standard deviation from the mean of 7 because their sample size is so large every year. Conversely, staffers are all over the place. Let’s take a look at just the top few rows based on frequency:
## author_type pub_year mean_score count
## 19 assistant editor 2015 7.200000 1
## 20 assistant editor 2016 7.553333 15
## 21 associate editor 2008 6.800000 21
## 22 associate editor 2009 8.000000 1
## 23 associate editor 2015 6.520000 5
## 24 associate editor 2016 6.620000 10
## 25 associate features editor 2015 5.600000 1
## 26 associate features editor 2016 6.500000 10
## 27 associate reviews editor 2011 7.500000 1
## 28 associate reviews editor 2012 7.784615 13In many cases, author types have but a single data point for every year. So while there may be a correlation between author type and review scores, the samples are too small to be able to identify it here.
Frequency
Let’s look at whether authors’ scores change over the course of their own careers—i.e., whether writing more articles results in higher, lower or stagnant scores.
This is a histogram of authors based on how many articles they’ve written:
So, quite a bit of variation, with some people writing more than 800 reviews and others contributing only one. This is the total number of articles written by the top 20 percent of writers:
## [1] 14755So the top 20 percent of writers (i.e., those who were in the 80th percentile of number of articles written) wrote 14,872 articles. How much of the total output is that?
## n
## 1 0.8023819Almost exactly 80 percent! It’s the Pareto Principle before our eyes. Now, what follows is the average score of every writer as a function of number of articles written—in other words, a single point is a single writer, their positions along the x-axis are determined by the number of articles they wrote, and their positions on the y-axis by the average score they gave to albums.
See how quickly the scores regress to the mean? It’s the Central Limit Theorem! Recall that the CLT states that over sufficient time, a sample of averages of a data set of any shape will ultimately plot along a normal distribution. This looks pretty normal to me! (Note: this chart excluded the top 5 percent of writers for the sake of scale, but they also fall into the long center of the mean.)
Review Length
Could longer reviews be associated with different types of scores? Looking at the stats for length, we get:
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0 498.0 604.0 650.1 746.0 3688.0First, we’ll indulge ourselves by discovering out which album was critiqued for 3700 words—a lot of words by Pitchfork standards.
## reviewid title artist
## 17756 22044 dangerous michael jackson
## url score
## 17756 http://pitchfork.com/reviews/albums/22044-dangerous/ 8.6
## best_new_music author author_type pub_date pub_weekday pub_day
## 17756 0 jeff weiss contributor 2016-08-07 6 7
## pub_month pub_year genre label diversity length
## 17756 8 2016 pop/r&b epic 0.4118764 3688Now let’s chart our data and see what happens.
It looks like we have found something! Here are some descriptive stats on the top 20 percent and bottom 20 percent of review scores, respectively:
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0 604.0 743.0 806.8 923.0 3688.0## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0 482.0 583.0 616.2 717.0 1858.0As we can see, the differences are substantial. If we run a test for correlation on the two variables, we get:
##
## Pearson's product-moment correlation
##
## data: reviews$length and reviews$score
## t = 31.195, df = 18387, p-value < 2.2e-16
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
## 0.2104259 0.2378803
## sample estimates:
## cor
## 0.2241976In other words, there’s a correlation. Logically, this makes sense: A writer will gush about an album he/she has fallen in love with, while a bad album is more likely (though certainly not certain) to be quickly dispatched. For example, see that one way down in the bottom left corner? If I recall correctly, that would be faux-leather Top 40 mistress Jet, which Pitchfork treated with a 0-point review and nothing more than a .gif of a chimpanzee pissing into its own mouth.
subset(reviews, score == 0 & length == 0)## reviewid title artist
## 6054 9464 shine on jet
## url score
## 6054 http://pitchfork.com/reviews/albums/9464-shine-on/ 0
## best_new_music author author_type pub_date pub_weekday pub_day
## 6054 0 ray suzuki contributor 2006-10-02 0 2
## pub_month pub_year genre label diversity length
## 6054 10 2006 electronic elektra 0 0LOL
Time of Review
Have reviews as a whole gotten more or less generous over time, or do they change depending on which month or where in the month they happen?
Doesn’t seem like it. Next.
Artist scores over time
One hypothesis is that artists peak and then fall. Let’s see if this is obvious just from a graph of artists who have produced 12 or more albums.
Nope! Putting aside nonquantitative points from this graphic, like why there are more than 12 Belle and Sebastian albums in existence or what value a Pitchfork review on Gucci Mane could hold, this is too illegibile to tell us anything. We can do a few things to fix it. First, we can manipulate our dataframe a bit so that albums are not only ordered by release date, but by where in the sequence of an artist’s reptertoire a given album came.
This is cool, but there’s a lot of white space. Let’s try looking at artists that have exactly 10 albums.
That also looks cool, and yet it’s still hard to tell if there’s a trend. The right side looks a bit bluer. One thing we can do is see look at some summary statistics of to see if the nth album on average has a lower or higher score than albums that came earlier.
## Warning: Removed 2 rows containing missing values (geom_path).
It doesn’t look like it, but it’s sort of impossible to tell with this. There are so few artists putting out 15+ albums that it’s not a decent sample. Plus, this data is arguably too aggregate to draw any meaningful conclusions from.
## -------------------------------------------------------------------------## You have loaded plyr after dplyr - this is likely to cause problems.
## If you need functions from both plyr and dplyr, please load plyr first, then dplyr:
## library(plyr); library(dplyr)## -------------------------------------------------------------------------##
## Attaching package: 'plyr'## The following objects are masked from 'package:plotly':
##
## arrange, mutate, rename, summarise## The following objects are masked from 'package:dplyr':
##
## arrange, count, desc, failwith, id, mutate, rename, summarise,
## summarize
The first two bars on the X axis are scores for individual albums, and from the third bar onward, the color is the rolling mean of the three albums inclusively. This is just artists who have put out 10 albums. We can also look at artists who have released 9:
or 20
or however many.
So do albums get worse over time? It looks like, maybe??? I like to save regression analyses to the end; it forces you to find phenomena clearly visible if in fact they exist. But since we’ve been unable to do that, let’s just do a regression analysis and see what comes up, wher each color indicates a different artist.
This is so horizontal that I actually have to check to make sure I was using the right code. But no, if we actually run the regression formula, we get
##
## Call:
## lm(formula = reviews$score ~ reviews$album.number)
##
## Coefficients:
## (Intercept) reviews$album.number
## 6.997581 0.000536which means, no, there’s nothing to see here, either.
Conclusion
I noticed something while playing with the dataset, and what I love about it is that after all the analysis and confronation with ggplot and facet-wrapping and tricked out multivariate scatterplots and heat maps that show nearly no measurable correlations, after all the labored deciphering of R’s cryptic, obscurantist documentation, perhaps the sole important insight from an 800 megabyte file of 18,000 rows of data can be gleaned from a single, one-line function of just four words and two numbers,
head(subset(reviews, score == 10), 20)## artist albums reviewid
## 606 ...and you will know us by the trail of dead 10 178
## 972 a tribe called quest 2 21218
## 1291 beastie boys 9 12671
## 1793 boards of canada 8 838
## 1837 bob dylan 16 22485
## 1902 bonnie prince billy 16 699
## 2067 brian eno 9 22061
## 2175 bruce springsteen 11 7728
## 2349 can 8 16075
## 3347 dangelo 1 17407
## 3477 david bowie 21 21487
## 3483 david bowie 21 21478
## 4028 dj shadow 11 2377
## 4596 elvis costello & the attractions 2 1615
## 5087 fleetwood mac 3 17499
## 5655 glenn branca 2 909
## 6010 gza 3 16857
## 6811 james brown 11 976
## 6983 j dilla 9 17510
## 7254 john coltrane 3 1550
## title
## 606 source tags and codes
## 972 people's instinctive travels and the paths of rhythm
## 1291 paul's boutique
## 1793 music has the right to children
## 1837 blood on the tracks
## 1902 i see a darkness
## 2067 another green world
## 2175 born to run: 30th anniversary edition
## 2349 tago mago [40th anniversary edition]
## 3347 voodoo
## 3477 "heroes"
## 3483 low
## 4028 endtroducing... [deluxe edition]
## 4596 this year's model
## 5087 rumours
## 5655 the ascension
## 6010 liquid swords: chess box deluxe edition
## 6811 live at the apollo [expanded edition]
## 6983 donuts (45 box set)
## 7254 the olatunji concert: the last live recording
## url
## 606 http://pitchfork.com/reviews/albums/178-source-tags-and-codes/
## 972 http://pitchfork.com/reviews/albums/21218-peoples-instinctive-travels-and-the-paths-of-rhythm/
## 1291 http://pitchfork.com/reviews/albums/12671-pauls-boutique/
## 1793 http://pitchfork.com/reviews/albums/838-music-has-the-right-to-children/
## 1837 http://pitchfork.com/reviews/albums/22485-blood-on-the-tracks/
## 1902 http://pitchfork.com/reviews/albums/699-i-see-a-darkness/
## 2067 http://pitchfork.com/reviews/albums/22061-another-green-world/
## 2175 http://pitchfork.com/reviews/albums/7728-born-to-run-30th-anniversary-edition/
## 2349 http://pitchfork.com/reviews/albums/16075-tago-mago-40th-anniversary-edition/
## 3347 http://pitchfork.com/reviews/albums/17407-voodoo/
## 3477 http://pitchfork.com/reviews/albums/21487-heroes/
## 3483 http://pitchfork.com/reviews/albums/21478-low/
## 4028 http://pitchfork.com/reviews/albums/2377-endtroducing-deluxe-edition/
## 4596 http://pitchfork.com/reviews/albums/1615-this-years-model/
## 5087 http://pitchfork.com/reviews/albums/17499-rumours/
## 5655 http://pitchfork.com/reviews/albums/909-the-ascension/
## 6010 http://pitchfork.com/reviews/albums/16857-liquid-swords-chess-box-deluxe-edition/
## 6811 http://pitchfork.com/reviews/albums/976-live-at-the-apollo-expanded-edition/
## 6983 http://pitchfork.com/reviews/albums/17510-donuts-45-box-set/
## 7254 http://pitchfork.com/reviews/albums/1550-the-olatunji-concert-the-last-live-recording/
## score best_new_music author author_type pub_date
## 606 10 0 matt lemay contributor 2002-02-28
## 972 10 1 kris ex contributor 2015-11-13
## 1291 10 1 nate patrin contributor 2009-02-13
## 1793 10 0 mark richardson executive editor 2004-04-26
## 1837 10 0 jesse jarnow contributor 2016-10-30
## 1902 10 0 samir khan contributor 1999-09-30
## 2067 10 0 mike powell contributor 2016-09-18
## 2175 10 0 mark richardson executive editor 2005-11-18
## 2349 10 0 douglas wolk contributor 2011-12-09
## 3347 10 1 ryan dombal senior editor 2012-12-12
## 3477 10 0 ryan dombal senior editor 2016-01-22
## 3483 10 0 laura snapes contributor 2016-01-22
## 4028 10 0 chris dahlen 2005-06-09
## 4596 10 0 matt lemay contributor 2002-05-09
## 5087 10 1 jessica hopper contributor 2013-02-08
## 5655 10 0 andy beta contributor 2003-06-19
## 6010 10 1 ian cohen contributor 2012-07-27
## 6811 10 0 dominique leone contributor 2004-03-30
## 6983 10 1 nate patrin contributor 2013-01-16
## 7254 10 0 luke buckman 2001-10-15
## pub_weekday pub_day pub_month pub_year genre
## 606 3 28 2 2002 rock
## 972 4 13 11 2015 rap
## 1291 4 13 2 2009 rap
## 1793 0 26 4 2004 electronic
## 1837 6 30 10 2016 rock
## 1902 3 30 9 1999 folk/country
## 2067 6 18 9 2016 experimental
## 2175 4 18 11 2005 rock
## 2349 4 9 12 2011 rock
## 3347 2 12 12 2012 pop/r&b
## 3477 4 22 1 2016 rock
## 3483 4 22 1 2016 rock
## 4028 3 9 6 2005 electronic
## 4596 3 9 5 2002
## 5087 4 8 2 2013 rock
## 5655 3 19 6 2003 rock
## 6010 4 27 7 2012 rap
## 6811 1 30 3 2004 rock
## 6983 2 16 1 2013 rap
## 7254 0 15 10 2001 jazz
## label diversity length album.number
## 606 interscope 0.4599078 1085 2
## 972 legacy 0.5273133 897 1
## 1291 capitol 0.5477912 1245 5
## 1793 warp 0.4964455 844 4
## 1837 columbia 0.4000923 2167 15
## 1902 palace 0.5804020 398 1
## 2067 island 0.4294508 2112 8
## 2175 columbia 0.5124046 1048 2
## 2349 mute 0.5156069 865 7
## 3347 light in the attic 0.4312401 2516 1
## 3477 rca 0.5216049 1296 15
## 3483 rca 0.5467498 1123 17
## 4028 mo'wax 0.5098952 859 4
## 4596 columbia 0.4714132 927 1
## 5087 warner bros. 0.4280423 1890 1
## 5655 acute 0.5416667 888 1
## 6010 geffen/get on down 0.4520022 1823 3
## 6811 king 0.4504881 1434 2
## 6983 stones throw 0.5144748 1209 7
## 7254 impulse! 0.5238095 1050 1Here are just the first 20 of the 76 albums in the set that have a perfect score of 10. Out of 20 albums, all but one of them were written years after the album (or the music on the album if it’s an anthology/re-release) was actually released. That’s years after the author developed their nostalgia of growing up outside Jersey City listening to the Boss croon, after the album was ushered into the national canon, after the debate over whether this was a Great Work of Art was settled. Ironic—our expedition in data science has taught us that the key lesson is that we don’t need to run a single test or plot a single chart to know that nothing is so great a predictor of who gets a 10 than the difference between the release date of the album and the date of the review. And why is that? What is even the point of assigning a score to Blood on the Tracks in 2016? Maybe it’s partly signaling, to indicate that an author is aware of just how big and important a thing is, that they have the critical eye to know a 10 when they see it. But this is Pitcfork, where, according to our calculations, the higher the score, the more likely the author will unleash a paean to an album’s glory. To dispense a 10 is to make the sign of the cross, which is why Kanye, a pesonality who has somehow convinced hoardes of critics to scramble to see which one can proclaim his genius the loudest, is one of the few active artists who regularly receives them. Tago Mago and Yeezus are now, in the litugry of Pitchfork, sancified with the acknowledgment of perfection. What’s the difference between a 7.3 and a 7.4? Nothing; the decimal points exist only to impress that the evauation of popular music is borne out with scientific rigor and outputs exact, objective, non-artificial results—that the arbitrary order imposed on something more or less orderless is in fact not arbitrary at all. But the difference between 9.9 and 10 is the distinction between measurement and worship.
All of which is to say, I think this dataset would benefit by having more data on the difference between publication date and release date, if someone with the web scraping skills were so inclined.