I’ve been working through the videos that accompany the Introduction to Statistical Learning with Applications in R book and thought it’d be interesting to try out the linear regression algorithm against my meetup data set.
I wanted to see how well a linear regression algorithm could predict how many people were likely to RSVP to a particular event. I started with the following code to build a data frame containing some potential predictors:
library(RNeo4j)
officeEventsQuery = "MATCH (g:Group {name: \"Neo4j - London User Group\"})-[:HOSTED_EVENT]->(event)<-[:TO]-({response: 'yes'})<-[:RSVPD]-(),
(event)-[:HELD_AT]->(venue)
WHERE (event.time + event.utc_offset) < timestamp() AND venue.name IN [\"Neo Technology\", \"OpenCredo\"]
RETURN event.time + event.utc_offset AS eventTime,event.announced_at AS announcedAt, event.name, COUNT(*) AS rsvps"
events = subset(cypher(graph, officeEventsQuery), !is.na(announcedAt))
events$eventTime <- timestampToDate(events$eventTime)
events$day <- format(events$eventTime, "%A")
events$monthYear <- format(events$eventTime, "%m-%Y")
events$month <- format(events$eventTime, "%m")
events$year <- format(events$eventTime, "%Y")
events$announcedAt<- timestampToDate(events$announcedAt)
events$timeDiff = as.numeric(events$eventTime - events$announcedAt, units = "days")If we preview ‘events’ it contains the following columns:
> head(events)
eventTime announcedAt event.name rsvps day monthYear month year timeDiff
1 2013-01-29 18:00:00 2012-11-30 11:30:57 Intro to Graphs 24 Tuesday 01-2013 01 2013 60.270174
2 2014-06-24 18:30:00 2014-06-18 19:11:19 Intro to Graphs 43 Tuesday 06-2014 06 2014 5.971308
3 2014-06-18 18:30:00 2014-06-08 07:03:13 Neo4j World Cup Hackathon 24 Wednesday 06-2014 06 2014 10.476933
4 2014-05-20 18:30:00 2014-05-14 18:56:06 Intro to Graphs 53 Tuesday 05-2014 05 2014 5.981875
5 2014-02-11 18:00:00 2014-02-05 19:11:03 Intro to Graphs 35 Tuesday 02-2014 02 2014 5.950660
6 2014-09-04 18:30:00 2014-08-26 06:34:01 Hands On Intro to Cypher - Neo4j's Query Language 20 Thursday 09-2014 09 2014 9.497211We want to predict ‘rsvps’ from the other columns so I started off by creating a linear model which took all the other columns into account:
> summary(lm(rsvps ~., data = events))
Call:
lm(formula = rsvps ~ ., data = events)
Residuals:
Min 1Q Median 3Q Max
-8.2582 -1.1538 0.0000 0.4158 10.5803
Coefficients: (14 not defined because of singularities)
Estimate Std. Error t value Pr(>|t|)
(Intercept) -9.365e+03 3.009e+03 -3.113 0.00897 **
eventTime 3.609e-06 2.951e-06 1.223 0.24479
announcedAt 3.278e-06 2.553e-06 1.284 0.22339
event.nameGraph Modelling - Do's and Don'ts 4.884e+01 1.140e+01 4.286 0.00106 **
event.nameHands on build your first Neo4j app for Java developers 3.735e+01 1.048e+01 3.562 0.00391 **
event.nameHands On Intro to Cypher - Neo4j's Query Language 2.560e+01 9.713e+00 2.635 0.02177 *
event.nameIntro to Graphs 2.238e+01 8.726e+00 2.564 0.02480 *
event.nameIntroduction to Graph Database Modeling -1.304e+02 4.835e+01 -2.696 0.01946 *
event.nameLunch with Neo4j's CEO, Emil Eifrem 3.920e+01 1.113e+01 3.523 0.00420 **
event.nameNeo4j Clojure Hackathon -3.063e+00 1.195e+01 -0.256 0.80203
event.nameNeo4j Python Hackathon with py2neo's Nigel Small 2.128e+01 1.070e+01 1.989 0.06998 .
event.nameNeo4j World Cup Hackathon 5.004e+00 9.622e+00 0.520 0.61251
dayTuesday 2.068e+01 5.626e+00 3.676 0.00317 **
dayWednesday 2.300e+01 5.522e+00 4.165 0.00131 **
monthYear01-2014 -2.350e+02 7.377e+01 -3.185 0.00784 **
monthYear02-2013 -2.526e+01 1.376e+01 -1.836 0.09130 .
monthYear02-2014 -2.325e+02 7.763e+01 -2.995 0.01118 *
monthYear03-2013 -4.605e+01 1.683e+01 -2.736 0.01805 *
monthYear03-2014 -2.371e+02 8.324e+01 -2.848 0.01468 *
monthYear04-2013 -6.570e+01 2.309e+01 -2.845 0.01477 *
monthYear04-2014 -2.535e+02 8.746e+01 -2.899 0.01336 *
monthYear05-2013 -8.672e+01 2.845e+01 -3.049 0.01011 *
monthYear05-2014 -2.802e+02 9.420e+01 -2.975 0.01160 *
monthYear06-2013 -1.022e+02 3.283e+01 -3.113 0.00897 **
monthYear06-2014 -2.996e+02 1.003e+02 -2.988 0.01132 *
monthYear07-2014 -3.123e+02 1.054e+02 -2.965 0.01182 *
monthYear08-2013 -1.326e+02 4.323e+01 -3.067 0.00976 **
monthYear08-2014 -3.060e+02 1.107e+02 -2.763 0.01718 *
monthYear09-2013 NA NA NA NA
monthYear09-2014 -3.465e+02 1.164e+02 -2.976 0.01158 *
monthYear10-2012 2.602e+01 1.959e+01 1.328 0.20886
monthYear10-2013 -1.728e+02 5.678e+01 -3.044 0.01020 *
monthYear11-2012 2.717e+01 1.509e+01 1.800 0.09704 .
month02 NA NA NA NA
month03 NA NA NA NA
month04 NA NA NA NA
month05 NA NA NA NA
month06 NA NA NA NA
month07 NA NA NA NA
month08 NA NA NA NA
month09 NA NA NA NA
month10 NA NA NA NA
month11 NA NA NA NA
year2013 NA NA NA NA
year2014 NA NA NA NA
timeDiff NA NA NA NA
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 5.287 on 12 degrees of freedom
Multiple R-squared: 0.9585, Adjusted R-squared: 0.8512
F-statistic: 8.934 on 31 and 12 DF, p-value: 0.0001399As I understand it we can look at the R-squared value to understand how much of the variance in the data has been explained by the model – in this case it’s 85%.
A lot of the coefficients seem to be based around specific event names which seems a bit too specific to me so I wanted to see what would happen if I derived a feature which indicated whether a session was practical:
events$practical = grepl("Hackathon|Hands on|Hands On", events$event.name)We can now run the model again with the new column having excluded ‘event.name’ field:
> summary(lm(rsvps ~., data = subset(events, select = -c(event.name))))
Call:
lm(formula = rsvps ~ ., data = subset(events, select = -c(event.name)))
Residuals:
Min 1Q Median 3Q Max
-18.647 -2.311 0.000 2.908 23.218
Coefficients: (13 not defined because of singularities)
Estimate Std. Error t value Pr(>|t|)
(Intercept) -3.980e+03 4.752e+03 -0.838 0.4127
eventTime 2.907e-06 3.873e-06 0.751 0.4621
announcedAt 3.336e-08 3.559e-06 0.009 0.9926
dayTuesday 7.547e+00 6.080e+00 1.241 0.2296
dayWednesday 2.442e+00 7.046e+00 0.347 0.7327
monthYear01-2014 -9.562e+01 1.187e+02 -0.806 0.4303
monthYear02-2013 -4.230e+00 2.289e+01 -0.185 0.8553
monthYear02-2014 -9.156e+01 1.254e+02 -0.730 0.4742
monthYear03-2013 -1.633e+01 2.808e+01 -0.582 0.5676
monthYear03-2014 -8.094e+01 1.329e+02 -0.609 0.5496
monthYear04-2013 -2.249e+01 3.785e+01 -0.594 0.5595
monthYear04-2014 -9.230e+01 1.401e+02 -0.659 0.5180
monthYear05-2013 -3.237e+01 4.654e+01 -0.696 0.4952
monthYear05-2014 -1.015e+02 1.509e+02 -0.673 0.5092
monthYear06-2013 -3.947e+01 5.355e+01 -0.737 0.4701
monthYear06-2014 -1.081e+02 1.604e+02 -0.674 0.5084
monthYear07-2014 -1.110e+02 1.678e+02 -0.661 0.5163
monthYear08-2013 -5.144e+01 6.988e+01 -0.736 0.4706
monthYear08-2014 -1.023e+02 1.784e+02 -0.573 0.5731
monthYear09-2013 -6.057e+01 7.893e+01 -0.767 0.4523
monthYear09-2014 -1.260e+02 1.874e+02 -0.672 0.5094
monthYear10-2012 9.557e+00 2.873e+01 0.333 0.7430
monthYear10-2013 -6.450e+01 9.169e+01 -0.703 0.4903
monthYear11-2012 1.689e+01 2.316e+01 0.729 0.4748
month02 NA NA NA NA
month03 NA NA NA NA
month04 NA NA NA NA
month05 NA NA NA NA
month06 NA NA NA NA
month07 NA NA NA NA
month08 NA NA NA NA
month09 NA NA NA NA
month10 NA NA NA NA
month11 NA NA NA NA
year2013 NA NA NA NA
year2014 NA NA NA NA
timeDiff NA NA NA NA
practicalTRUE -9.388e+00 5.289e+00 -1.775 0.0919 .
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 10.21 on 19 degrees of freedom
Multiple R-squared: 0.7546, Adjusted R-squared: 0.4446
F-statistic: 2.434 on 24 and 19 DF, p-value: 0.02592Now we’re only accounting for 44% of the variance and none of our coefficients are significant so this wasn’t such a good change.
I also noticed that we’ve got a bit of overlap in the date related features – we’ve got one column for monthYear and then separate ones for month and year. Let’s strip out the combined one:
> summary(lm(rsvps ~., data = subset(events, select = -c(event.name, monthYear))))
Call:
lm(formula = rsvps ~ ., data = subset(events, select = -c(event.name,
monthYear)))
Residuals:
Min 1Q Median 3Q Max
-16.5745 -4.0507 -0.1042 3.6586 24.4715
Coefficients: (1 not defined because of singularities)
Estimate Std. Error t value Pr(>|t|)
(Intercept) -1.573e+03 4.315e+03 -0.364 0.7185
eventTime 3.320e-06 3.434e-06 0.967 0.3425
announcedAt -2.149e-06 2.201e-06 -0.976 0.3379
dayTuesday 4.713e+00 5.871e+00 0.803 0.4294
dayWednesday -2.253e-01 6.685e+00 -0.034 0.9734
month02 3.164e+00 1.285e+01 0.246 0.8075
month03 1.127e+01 1.858e+01 0.607 0.5494
month04 4.148e+00 2.581e+01 0.161 0.8736
month05 1.979e+00 3.425e+01 0.058 0.9544
month06 -1.220e-01 4.271e+01 -0.003 0.9977
month07 1.671e+00 4.955e+01 0.034 0.9734
month08 8.849e+00 5.940e+01 0.149 0.8827
month09 -5.496e+00 6.782e+01 -0.081 0.9360
month10 -5.066e+00 7.893e+01 -0.064 0.9493
month11 4.255e+00 8.697e+01 0.049 0.9614
year2013 -1.799e+01 1.032e+02 -0.174 0.8629
year2014 -3.281e+01 2.045e+02 -0.160 0.8738
timeDiff NA NA NA NA
practicalTRUE -9.816e+00 5.084e+00 -1.931 0.0645 .
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 10.19 on 26 degrees of freedom
Multiple R-squared: 0.666, Adjusted R-squared: 0.4476
F-statistic: 3.049 on 17 and 26 DF, p-value: 0.005187Again none of the coefficients are statistically significant which is disappointing. I think the main problem may be that I have very few data points (only 42) making it difficult to come up with a general model.
I think my next step is to look for some other features that could impact the number of RSVPs e.g. other events on that day, the weather.
I’m a novice at this but trying to learn more so if you have any ideas of what I should do next please let me know.
| Reference: | R: A first attempt at linear regression from our JCG partner Mark Needham at the Mark Needham Blog blog. |
