library(tidyverse)
library(lubridate)
library(tm)
library(tidytext)
library(widyr)
library(wordcloud2)
library(textstem)
library(ggwordcloud)
library(lexicon)
library(textdata)
library(stringr)
library(cld2)
library(caTools)load("./glassDoor.RData")The dataset contains GlassDoor reviews for four organizations. Reviews are separated into pros, cons, and advice.
Problems with text:
organization name mixed in words, eg. ChORGAllenging
Non-english reviews
gdMixed <- glassDoor %>%
filter((str_detect(pros, "[^\\s]+ORG[A-D][^\\s]+"))|
(str_detect(cons, "[^\\s]+ORG[A-D][^\\s]+"))|
(str_detect(advice, "[^\\s]+ORG[A-D][^\\s]+")))
rmarkdown::paged_table(gdMixed)## Error in print.paged_df(x, ...): unused argument (options = options)
gdNonEng <- glassDoor %>%
mutate(language = cld2::detect_language(pros)) %>%
filter(language != "en") %>%
dplyr::select(pros, cons, advice, language)
rmarkdown::paged_table(gdNonEng)## Error in print.paged_df(x, ...): unused argument (options = options)
They are only about 10% of the rows. Deleting them won't affect the result too much. The 36 rows of mixed word errors exhibits no pattern of the replaced letters. We may fuzzy join these words with a english dictionary and replace them. For the non-english reviews, not sure how well the translators perform in R, might just ignore them as well.
gd2 <- glassDoor %>%
filter(!(str_detect(pros, "[^\\s]+ORG[A-D][^\\s]+"))) %>%
filter(!(str_detect(cons, "[^\\s]+ORG[A-D][^\\s]+"))) %>%
filter(!(str_detect(advice, "[^\\s]+ORG[A-D][^\\s]+")) | is.na(advice)) %>%
mutate(language = cld2::detect_language(pros)) %>%
filter(language == "en") %>%
mutate(organization = as.factor(organization),
pros = gsub("([a-z])([A-Z])", "\\1 \\2", pros),
cons = gsub("([a-z])([A-Z])", "\\1 \\2", cons),
advice = gsub("([a-z])([A-Z])", "\\1 \\2", advice),
rating = fct_relevel(rating, "5.0", "4.0", "3.0", "2.0"))
str(gd2)## 'data.frame': 1682 obs. of 11 variables:
## $ pros : chr "Good learning opportunities - a mixture of strategy and banking technology projects Smart people to work with a"| __truncated__ "- Excellent projects- Several growth opportunities" "Like the co-worker I work with." "Exposure to many different sectors" ...
## $ cons : chr "Lack of structure sometimes and unplanned" "- Work life balance is considerably poor as compared to industry standards" "Takes more time than expected" "Often requires you to stay late" ...
## $ advice : chr NA NA "Listen to the workers doing the hands on work." NA ...
## $ rating : Factor w/ 5 levels "5.0","4.0","3.0",..: 2 3 3 2 1 3 4 4 1 2 ...
## $ workLifeRating : Factor w/ 9 levels "1.0","3.0","2.0",..: NA 1 2 NA 3 3 NA 1 4 2 ...
## $ cultureValueRating : Factor w/ 9 levels "2.0","3.0","1.0",..: NA 1 2 NA 2 3 NA 1 4 2 ...
## $ careerOpportunityRating: Factor w/ 8 levels "4.0","2.0","3.0",..: NA 1 2 NA 1 2 NA 2 1 1 ...
## $ compBenefitsRating : Factor w/ 9 levels "2.0","1.0","5.0",..: NA 1 1 NA 1 1 NA 2 3 4 ...
## $ managementRating : Factor w/ 5 levels "2.0","3.0","4.0",..: NA 1 2 NA 3 4 NA 4 5 3 ...
## $ organization : Factor w/ 4 levels "ORGA","ORGB",..: 2 4 3 2 3 4 1 4 2 1 ...
## $ language : chr "en" "en" "en" "en" ...
orga <- gd2 %>%
filter(organization == "ORGA") %>%
dplyr::select(pros, cons, advice) %>%
pivot_longer(., cols = colnames(.), names_to = "type", values_to = "text") %>%
mutate(type = as.factor(type)) %>%
mutate(type = fct_relevel(type, "pros", "cons")) %>%
drop_na(text) %>%
mutate(text = str_replace_all(text, "ORGA", ""),
text = tolower(text),
text = lemmatize_strings(text),
text = stripWhitespace(text),
text = removeNumbers(text))
tokena <- orga %>%
unnest_tokens(word, text) %>%
anti_join(stop_words) %>%
left_join(parts_of_speech) %>%
filter(pos == "Noun" & word != "company") %>%
group_by(type) %>%
count(word) %>%
mutate(freq = n/sum(n)) %>%
top_n(n = 20, wt = freq)
tokena %>%
ggplot(aes(label = word, size = freq,
color = type))+
geom_text_wordcloud_area()+
scale_size_area(max_size = 12)+
theme_minimal()+
scale_color_manual(values = c("pros"="#91cf60", "cons" = "#ef8a62", "advice"="#67a9cf"))+
facet_wrap(~type)+
labs(title = "Organization A")
It seems words such as "people", "management", "employee", "pay", appears in all pros, cons, and advice in all four companies' reviews. How come people love and hate the same aspect about the same company? We need to take a deeper dive in the future.
We are going to do sentiment analyses on Pros, Cons, and Advice separately. Using functions from sentimentr, and applying jocker polarity table (ranges from -1 to 1) and hash_valence_shifters, we can get the sentiment for each reviews. Then we can aggregate the sentiments by rating or company.
library(sentimentr)
library(lexicon)
library(magrittr)
prosDF <- gd2 %>%
dplyr::select(pros, rating, organization) %>%
mutate(pros = str_replace_all(pros, "ORG[A-D]", ""),
pros = str_replace_all(pros, "[\\.\\!\\?]", ""),
pros = tolower(pros),
pros = lemmatize_strings(pros),
pros = stripWhitespace(pros),
pros = removeNumbers(pros))
proSenti = sentiment(get_sentences(prosDF),
polarity_dt = lexicon::hash_sentiment_jockers,
valence_shifters_dt = lexicon::hash_valence_shifters)
## aggregate by company
proSenti %>%
group_by(organization) %>%
summarize(meanSentiment = mean(sentiment))## # A tibble: 4 x 2
## organization meanSentiment
## <fct> <dbl>
## 1 ORGA 0.582
## 2 ORGB 0.577
## 3 ORGC 0.580
## 4 ORGD 0.589
The sentiments in Pros reviews for each company are not too different with ORGD taking a slight lead.
## aggregate sentiment by rating
proSenti %>%
group_by(rating) %>%
summarize(meanSentiment = mean(sentiment))## # A tibble: 5 x 2
## rating meanSentiment
## <fct> <dbl>
## 1 5.0 0.708
## 2 4.0 0.681
## 3 3.0 0.577
## 4 2.0 0.482
## 5 1.0 0.258
It is within expectation. Higher rated reviews have more positive words.
## aggregate by both company and rating
proSenti %>%
group_by(organization, rating) %>%
summarize(meanSentiment = mean(sentiment)) %>%
ggplot(aes(x = rating, y = meanSentiment, fill = organization))+
geom_col()+
facet_wrap(~organization)+
geom_text(aes(label = round(meanSentiment,2)), vjust = -0.2)+
theme_minimal()+
theme(panel.grid.major = element_blank(),
axis.text.y = element_blank(),
legend.position = "none")
Within companies, we can still observe that sentiment goes down as rating decreases.
Using the same method on Cons
## # A tibble: 4 x 2
## organization meanSentiment
## <fct> <dbl>
## 1 ORGA 0.0449
## 2 ORGB 0.0574
## 3 ORGC 0.0310
## 4 ORGD 0.0712
## # A tibble: 5 x 2
## rating meanSentiment
## <fct> <dbl>
## 1 5.0 0.111
## 2 4.0 0.0857
## 3 3.0 0.0228
## 4 2.0 -0.0124
## 5 1.0 -0.0409
Of course the sentiment scores for Cons are much lower, sometime even negative. But they have similar pattern as Pros. And range of the scores is much smaller than Pros. People left more positive words for ORGD again in Cons.
## # A tibble: 4 x 2
## organization meanSentiment
## <fct> <dbl>
## 1 ORGA 0.177
## 2 ORGB 0.154
## 3 ORGC 0.184
## 4 ORGD 0.198
## # A tibble: 5 x 2
## rating meanSentiment
## <fct> <dbl>
## 1 5.0 0.193
## 2 4.0 0.194
## 3 3.0 0.171
## 4 2.0 0.193
## 5 1.0 0.124
Advice is more positive than Cons. However, the sentiment score do not decrease as rating decrease when rating >= 2. Again, ORGD reveives more positive words on Advice. It seems people speak positively in all three kinds of reviews in ORGD.
gd3 <- gd2 %>%
dplyr::select(pros, cons, advice, rating, organization) %>%
drop_na(c(pros, cons, advice)) %>%
mutate(text = str_c(pros, cons, advice, sep = " ")) %>%
dplyr::select(rating, organization, text)library(stm)
set.seed(1234)
holdoutRows = sample(1:nrow(gd3), 100, replace = FALSE)
gdText = textProcessor(documents = gd3$text[-c(holdoutRows)],
metadata = gd3[-c(holdoutRows), ],
stem = FALSE)## Building corpus...
## Converting to Lower Case...
## Removing punctuation...
## Removing stopwords...
## Removing numbers...
## Creating Output...
gdPrep = prepDocuments(documents = gdText$documents,
vocab = gdText$vocab,
meta = gdText$meta)## Removing 3882 of 7056 terms (3882 of 39296 tokens) due to frequency
## Your corpus now has 992 documents, 3174 terms and 35414 tokens.
kTest = searchK(documents = gdPrep$documents,
vocab = gdPrep$vocab,
K = c(3, 4, 5, 10, 20), verbose = FALSE)
# png(file = "./ktest1new.png", width = 800, height = 600)
# plot(kTest)
# dev.off()
Looks like 4 is a reasonable number of topics.
Now let's look at expected topic proportions.
topics4 = stm(documents = gdPrep$documents,
vocab = gdPrep$vocab, seed = 1001,
K = 4, verbose = FALSE)
plot(topics4)
labelTopics(topics4)## Topic 1 Top Words:
## Highest Prob: work, good, great, people, environment, can, benefits
## FREX: lots, students, learning, interns, school, key, open
## Lift: achievement, allocation, alternate, application, archaic, assortment, aug
## Score: good, lots, friendly, learning, great, environment, flexible
## Topic 2 Top Words:
## Highest Prob: company, will, employees, people, dont, get, work
## FREX: please, anything, worst, month, horrible, tell, wrong
## Lift: -called, -hour, -site, absence, absolute, accepting, actions
## Score: worst, half, tell, rep, horrible, property, quit
## Topic 3 Top Words:
## Highest Prob: work, projects, firm, consulting, culture, senior, clients
## FREX: firm, firms, interesting, partners, brand, terms, junior
## Lift: abroad, allowance, amazingly, ambitious, analytical, asia, asian
## Score: firm, projects, consulting, firms, exposure, consultants, partners
## Topic 4 Top Words:
## Highest Prob: company, staff, management, employees, new, employee, training
## FREX: community, values, tasks, amazing, administration, fantastic, loyalty
## Lift: ’ve, abilities, absorb, accounting, accounts, achievements, acquired
## Score: staff, leadership, community, communities, federal, fair, upper
From highest prob and FREX words, we can see the difference between the 5 topics. It seems topic 1 is compliment of the work at the company. Topic 2 is about pay and benefits. Topic 3 is about the cons of the company. Topic 4 is about professional experience.
head(topics4$theta, 15)## [,1] [,2] [,3] [,4]
## [1,] 0.1756340 0.66551173 0.05017709 0.10867716
## [2,] 0.2649366 0.02862035 0.63888994 0.06755310
## [3,] 0.3073273 0.35936017 0.29633838 0.03697412
## [4,] 0.5645380 0.04787455 0.28583712 0.10175032
## [5,] 0.4298070 0.23965697 0.07805118 0.25248484
## [6,] 0.1784487 0.01460559 0.77973885 0.02720691
## [7,] 0.1942635 0.02574412 0.72010407 0.05988827
## [8,] 0.1358200 0.01854973 0.80976235 0.03586792
## [9,] 0.1223953 0.01314036 0.84045773 0.02400658
## [10,] 0.2252693 0.30645612 0.33078665 0.13748797
## [11,] 0.3098211 0.32400444 0.17645013 0.18972433
## [12,] 0.3886456 0.21431376 0.29767919 0.09936142
## [13,] 0.6546003 0.11750467 0.15689051 0.07100456
## [14,] 0.1894023 0.01356781 0.76456404 0.03246589
## [15,] 0.3449836 0.03223478 0.57542942 0.04735223
We can see some documents have a more focused topic, and some documents span different topics.
Let's use the model to classify heldout documents
newgdText = textProcessor(documents = gd3$text[holdoutRows],
metadata = gd3[holdoutRows, ],
stem = FALSE)## Building corpus...
## Converting to Lower Case...
## Removing punctuation...
## Removing stopwords...
## Removing numbers...
## Creating Output...
newgdCorp = alignCorpus(new = newgdText, old.vocab = topics4$vocab)## Your new corpus now has 100 documents, 1447 non-zero terms of 2268 total terms in the original set.
## 821 terms from the new data did not match.
## This means the new data contained 45.6% of the old terms
## and the old data contained 63.8% of the unique terms in the new data.
## You have retained 4812 tokens of the 5714 tokens you started with (84.2%).
newgdFitted = fitNewDocuments(model = topics4, documents = newgdCorp$documents,
newData = newgdCorp$meta, origData = gdPrep$meta)## ....................................................................................................
head(newgdFitted$theta, 15)## [,1] [,2] [,3] [,4]
## [1,] 0.21361860 0.23759019 0.10873902 0.44005218
## [2,] 0.44609139 0.41309024 0.05770793 0.08311044
## [3,] 0.25195379 0.45196056 0.06599594 0.23008971
## [4,] 0.33911532 0.41491338 0.06179325 0.18417805
## [5,] 0.19376840 0.12010528 0.44704999 0.23907632
## [6,] 0.31905012 0.44809372 0.12579906 0.10705710
## [7,] 0.17279768 0.35462032 0.07645427 0.39612773
## [8,] 0.21067426 0.58780508 0.09953192 0.10198875
## [9,] 0.29057213 0.05123786 0.15904898 0.49914104
## [10,] 0.36374866 0.29886380 0.05026104 0.28712650
## [11,] 0.20227397 0.69334832 0.03801047 0.06636724
## [12,] 0.20646682 0.38924165 0.04159133 0.36270020
## [13,] 0.55030704 0.12288921 0.21417721 0.11262655
## [14,] 0.39643243 0.08474309 0.10712628 0.41169821
## [15,] 0.09463731 0.64928810 0.08666505 0.16940954
Let's see if any topic proportion is related to rating.
gd3 <- gd3 %>%
mutate(rating = as.numeric(as.character(rating)))
predictorText = textProcessor(documents = gd3$text,
metadata = gd3,
stem = FALSE)## Building corpus...
## Converting to Lower Case...
## Removing punctuation...
## Removing stopwords...
## Removing numbers...
## Creating Output...
gdPrep = prepDocuments(documents = predictorText$documents,
vocab = predictorText$vocab,
meta = predictorText$meta)## Removing 4202 of 7632 terms (4202 of 44161 tokens) due to frequency
## Your corpus now has 1092 documents, 3430 terms and 39959 tokens.
topicPredictor = stm(documents = gdPrep$documents,
vocab = gdPrep$vocab, prevalence = ~ rating,
data = gdPrep$meta, K = 4, verbose = FALSE)
ratingEffect = estimateEffect(1:4 ~ rating, stmobj = topicPredictor,
metadata = gdPrep$meta)
summary(ratingEffect, topics = c(1:4))##
## Call:
## estimateEffect(formula = 1:4 ~ rating, stmobj = topicPredictor,
## metadata = gdPrep$meta)
##
##
## Topic 1:
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.194394 0.019446 9.996 <2e-16 ***
## rating 0.006213 0.005581 1.113 0.266
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
##
## Topic 2:
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.431551 0.020087 21.484 < 2e-16 ***
## rating -0.037651 0.005587 -6.739 2.57e-11 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
##
## Topic 3:
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.466934 0.018446 25.31 <2e-16 ***
## rating -0.084882 0.004985 -17.03 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
##
## Topic 4:
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -0.092354 0.014687 -6.288 4.64e-10 ***
## rating 0.116199 0.004426 26.252 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
plot.estimateEffect(ratingEffect, "rating", method = "continuous",
model = topicPredictor, topics = 1, labeltype = "frex")
plot.estimateEffect(ratingEffect, "rating", method = "continuous",
model = topicPredictor, topics = 2, labeltype = "frex")
plot.estimateEffect(ratingEffect, "rating", method = "continuous",
model = topicPredictor, topics = 3, labeltype = "frex")
plot.estimateEffect(ratingEffect, "rating", method = "continuous",
model = topicPredictor, topics = 4, labeltype = "frex")
Looks like ratings are associated with topic proportions. Except topic 1 not so significant.
thetas=topicPredictor$thetaIt seems that these topic are related to rating. On the other hand, We may use them to predict ratings.
We are going to use sentiments and topic thetas to predict ratings.
proSenti <- proSenti %>%
rename(c("proSentiment"="sentiment"))
conSenti <- consenti %>%
rename(c("conSentiment"="sentiment"))
adviceSenti <- adviceSenti %>%
rename(c("adviceSentiment"="sentiment"))
gd4 <- gd2 %>%
mutate(proSentiment = proSenti$proSentiment,
conSentiment = conSenti$conSentiment,
adviceSentiment = adviceSenti$adviceSentiment) %>%
drop_na(c(pros, cons, advice)) %>%
mutate(topic1theta = thetas[,1],
topic2theta = thetas[,2],
topic3theta = thetas[,3],
topic4theta = thetas[,4])Split into training and testing sets
gd5 <- gd4 %>%
dplyr::select(rating, proSentiment, conSentiment, adviceSentiment,
topic1theta, topic2theta, topic3theta, topic4theta, organization) %>%
mutate(rating = as.numeric(as.character(rating)))
sample_set <- gd5 %>%
pull(.) %>%
sample.split(SplitRatio = .7)
gdTrain <- subset(gd5, sample_set == TRUE)
gdTest <- subset(gd5, sample_set == FALSE)Let's try linear model first
lmod <- lm(rating~proSentiment+ conSentiment+ adviceSentiment+
topic1theta+ topic2theta+ topic3theta, data=gd5)
summary(lmod)##
## Call:
## lm(formula = rating ~ proSentiment + conSentiment + adviceSentiment +
## topic1theta + topic2theta + topic3theta, data = gd5)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.03007 -0.64274 -0.05801 0.60607 2.84685
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.92017 0.11706 50.576 < 2e-16 ***
## proSentiment 0.31920 0.06770 4.715 2.74e-06 ***
## conSentiment 0.05862 0.08530 0.687 0.492
## adviceSentiment -0.01734 0.07766 -0.223 0.823
## topic1theta -3.45484 0.16038 -21.542 < 2e-16 ***
## topic2theta -3.90417 0.14971 -26.078 < 2e-16 ***
## topic3theta -5.04814 0.15290 -33.016 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8988 on 1085 degrees of freedom
## Multiple R-squared: 0.6218, Adjusted R-squared: 0.6197
## F-statistic: 297.3 on 6 and 1085 DF, p-value: < 2.2e-16
Remove insignificant variables.
lmod2 <- lm(rating~proSentiment+topic1theta+topic2theta+topic3theta, data=gd5)
summary(lmod2)##
## Call:
## lm(formula = rating ~ proSentiment + topic1theta + topic2theta +
## topic3theta, data = gd5)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.02933 -0.64331 -0.05322 0.60442 2.87685
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 5.92554 0.11138 53.202 < 2e-16 ***
## proSentiment 0.32047 0.06748 4.749 2.32e-06 ***
## topic1theta -3.46199 0.15988 -21.654 < 2e-16 ***
## topic2theta -3.91761 0.14631 -26.777 < 2e-16 ***
## topic3theta -5.05671 0.14762 -34.255 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8982 on 1087 degrees of freedom
## Multiple R-squared: 0.6217, Adjusted R-squared: 0.6203
## F-statistic: 446.5 on 4 and 1087 DF, p-value: < 2.2e-16
Adj. R-squared is 0.6203, not too bad.
lmPred <- predict(lmod2, gdTest)
lmPred_round <- ifelse(lmPred <=1.0, 1.0, lmPred)
lmPred_round <- ifelse(lmPred_round >=5.0, 5.0, lmPred_round)
lmPred_round <- round(lmPred_round)
table(lmPred_round, gdTest$rating)##
## lmPred_round 1 2 3 4 5
## 1 18 4 0 1 0
## 2 44 10 12 1 1
## 3 9 20 41 38 13
## 4 1 2 14 25 26
## 5 0 0 2 13 32
mean(as.character(lmPred_round) != as.character(gdTest$rating))## [1] 0.6146789
There are a lot of wrong prediction in the test data. But the predictions are not too far off. Most predictions are within 1.0 points range around the actuals. A 59% of the test data is missclassified.
Multinomial logistic regression
Since the ratings are discrete, let's see how a multinomial logistic regression performs.
gd5_factor <- gd4 %>%
dplyr::select(rating, proSentiment, conSentiment, adviceSentiment,
topic1theta, topic2theta, topic3theta, topic4theta, organization)
set.seed(1234)
sample_set <- gd5_factor %>%
pull(.) %>%
sample.split(SplitRatio = .7)
gdTrain <- subset(gd5_factor, sample_set == TRUE)
gdTest <- subset(gd5_factor, sample_set == FALSE)
library(MASS)## Warning: package 'MASS' was built under R version 3.6.2
##
## Attaching package: 'MASS'
## The following object is masked from 'package:dplyr':
##
## select
ordModel <- polr(rating~proSentiment+topic1theta+topic2theta+topic3theta, data=gd5_factor, Hess = TRUE)
summary (ordModel) ## Call:
## polr(formula = rating ~ proSentiment + topic1theta + topic2theta +
## topic3theta, data = gd5_factor, Hess = TRUE)
##
## Coefficients:
## Value Std. Error t value
## proSentiment -0.8024 0.1590 -5.048
## topic1theta 8.0278 0.4497 17.853
## topic2theta 8.7015 0.4355 19.979
## topic3theta 11.0001 0.4806 22.888
##
## Intercepts:
## Value Std. Error t value
## 5.0|4.0 3.8144 0.2961 12.8839
## 4.0|3.0 5.8562 0.3348 17.4898
## 3.0|2.0 7.5522 0.3657 20.6489
## 2.0|1.0 8.6461 0.3834 22.5500
##
## Residual Deviance: 2401.442
## AIC: 2417.442
predicted_scores <- predict (ordModel, gdTest, "probs")
predicted_class <- predict (ordModel, gdTest)
table(predicted_class, gdTest$rating)##
## predicted_class 5.0 4.0 3.0 2.0 1.0
## 5.0 39 23 3 0 0
## 4.0 31 23 18 4 1
## 3.0 6 25 37 17 9
## 2.0 0 0 0 0 0
## 1.0 1 3 6 10 71
mean(as.character(predicted_class) != as.character(gdTest$rating))## [1] 0.4801223
It improved a litte with a missclassification rate of 48%.
There is still a lot of room to improve the model to predict the rating. Here, we are only using text features such as sentiment and topic thetas. We can add other rating in the model or add some interactions with organization. We can try mixed model, too. Using more complexed model may improve the performance.