What Contributes to a Crowdfunding Campaign's Success? Evidence and Analyses from GoFundMe Data

Abstract

Researchers have attempted to measure the success of crowdfunding campaigns using a variety of determinants, such as the descriptions of the crowdfunding campaigns, the amount of funding goals, and crowdfunding project characteristics. Although many success determinants have been reported in the literature, it remains unclear whether the cover photo and the text in the title and description could be combined in a fusion classifier to better predict the crowdfunding campaign's success. In this work, we focus on the performance of the crowdfunding campaigns on GoFundMe over a wide variety of funding categories. We analyze the attributes available at the launch of the campaign and identify attributes that are important for each category of the campaigns. Furthermore, we develop a fusion classifier based on random forest that significantly improves the prediction result, thus suggesting effective ways to make a campaign successful.

Full text

What Contributes to a Crowdfunding Campaigns Success?
Evidences and Analyses from GoFundMe Data
Xupin Zhang, 1Hanjia Lyu, 2Jiebo Luo3
University of Rochester
xzhang72@u.rochester.edu, 1hlyu5@ur.rochester.edu, 2jluo@cs.rochester.edu 3
Abstract
Researchers have attempted to measure the success of crowd-
funding campaigns using a variety of determinants, such as
the descriptions of the crowdfunding campaigns, the amount
of funding goals, and crowdfunding project characteristics.
Although many success determinants have been reported in
the literature, it remains unclear whether the cover photo and
the text in the title and description could be combined in a fu-
sion classifier to better predict the crowdfunding campaigns
success. In this work, we focus on the performance of the
crowdfunding campaigns on GoFundMe over a wide variety
of funding categories. We analyze the attributes available at
the launch of the campaign and identify attributes that are
important for each category of the campaigns. Furthermore,
we develop a fusion classifier based on random forest that
significantly improves the prediction result, thus suggesting
effective ways to make a campaign successful.
Introduction
In recent years, the rise of charitable crowdfunding plat-
forms such as GoFundMe makes it possible for Internet
users to offer direct help to those who need emergency finan-
cial assistance. However, the success rate of the campaigns
is found to be less than 50% (success is defined as a cam-
paign that reaches its funding goal).
In this study, we analyze GoFundMe, which is currently
the biggest crowdfunding platform. It has helped to raise
over $5 billion since its debut in 2010. This site allows peo-
ple to raise money for various events, from life events like
a wedding to challenging situations such as accidents or ill-
ness (Monroe 2019). We crawled the pages of all 10,974
available crowdfunding campaigns on the site. This research
investigates success determinants for a crowdfunding cam-
paign. Our research questions are: can we quantify the eco-
nomic returns of the image and text features? If so, can we
reliably predict the fundraisings performance using the at-
tributes available at the launch of the crowdfunding cam-
paign?
Using the variables extracted from our dataset, we de-
fine the measure of the crowdfundings success as the ratio
of the current amount of money that has been raised to the
fundraisers goal amount. To further understand the determi-
nants of successful campaigns, we separately analyze image
and text features to understand how much each of these fac-
tors contributes to the success of a crowdfunding campaign.
So far, almost all research concerning crowdfunding has
analyzed Kickstarter to predict crowdfunding performance.
Kickstarter is a fundraising platform that differs from Go-
FundMe in the usership demographics and management of
the raised money. We investigate whether the effects of text
and image features are consistent across Kickstarter and
GoFundMe and how much the difference between success-
ful and unsuccessful campaigns can be explained by such
factors. We predict crowdfunding outcomes by combining
both textual and pictorial descriptions of the crowdfunding
projects. This combination provides a more comprehensive
view of the factors in successful crowdfunding projects and
helps to better take into account possible interrelations.
Because GoFundMe categorizes fundraising by its pur-
poses (e.g. medical, education, wedding, etc.), we analyze
the important textual and pictorial features that contribute
to the success of fundraising for specific purposes. For in-
stance, volunteer and service campaigns with more people
in the cover photo might have a higher chance to succeed
than those with a cover image of a single person.
The main contributions of our research are:
•We analyze image and text features that are important to
specific categories of crowdfunding campaigns.
•We analyze facial attributes in the cover image and exam-
ine their impact on crowdfunding performance.
•We design a fusion analytic framework that can combine
both textual and pictorial descriptions of crowdfunding
projects to reliably predict crowdfunding outcomes.
To the best of our knowledge, this is the first success-
ful research that applies both language topic modeling and
computer vision methods to extract features from project de-
scriptions and cover images, in order to analyze and predict
crowdfunding project success. We analyze GoFundMe be-
cause campaigns launched here are charity-minded and rich
in categories, while campaigns launched on Kickstarter are
exclusively entrepreneurial. This project provides a more
comprehensive view of the factors in successful project
funding of projects and better takes into account possible
interrelations. The managerial implication of our research is
arXiv:2001.05446v2 [cs.SI] 16 Jan 2020

that the crowdfunding platforms can better identify the most
influential image and text features. They can offer strate-
gic suggestions to help their users (fundraisers) raise more
money and also attract more donors to their websites.
Related Work
Many studies have been conducted to explore the deter-
minants of campaign success on the crowdfunding plat-
form Kickstarter. It has been found that active commu-
nications with the platform members (Xiao et al. 2014),
project description and image (Greenberg et al. 2013),
project characteristics (Mitra and Gilbert 2014; Etter et al.
2014), geographical factors (Mollick 2014), linguistic style
(Parhankangas and Renko 2017), the amount of the funding
goal (Miller et al. 2013), the number of first backers, and
the content of the project updates (Kuppuswamy and Bayus
2015) have a significant impact on the success of crowdfund-
ing projects.
Several researchers have attempted to predict crowdfund-
ing success using various machine learning techniques. In
a study conducted by Greenberg et al. (2013), they found
that the decision tree classifier predicted the crowdfunding
success with an accuracy of 68% at best. Instead of using
static attributes (i.e., attributes available at the launch of the
campaign), Etter, Grossglauser, and Thiran (2013) combined
both direct features and social features to predict the cam-
paign outcome, and their model achieved a 76% accuracy.
Yuan, Lau, and Xu (2016) proposed a semantic text analytic
approach to predicting crowdfunding success; they found
that topic models mined from topic descriptions are useful
for prediction. In addition, they found that an ensemble of
weak classifiers - random forest performed better than a sin-
gle strong classifier - support vector machine.
Moreover, facial expression such as smile is shown to be
helpful in predicting crowdfunding success. Smile is collec-
tively understood as a sign of friendliness, generosity, and
other altruistic behaviors (Brown et al. 2003; Gabriel et al.
2015; Grandey and Gabriel 2015; Lau 1982; Mehu et al.
2008). An employee-displayed smile brings three primary
positive outcomes, including immediate gains (e.g., sales
and tipping), encore gains (e.g., customer loyalty), and con-
tagion gains (e.g., positive word-of-mouth). Smiling is also
widely used as a useful marketing tool to improve impres-
sions among consumers (customer relationship) and to en-
hance consumers consumption experiences (Lee and Lim
2010; Barger and Grandey 2006). Kim and Park (2017) con-
ducted an empirical analysis to examine the relationship
between facial expression and crowdfunding success. They
found out the inclusion of a smiling face is associated with
5% increase in the funding amount.
The literature suggests many factors for crowdfunding
success using quantitative methods, and some articles uti-
lize qualitative methods to examine the crowdfunding per-
formance (e.g. Johansen 2019) by interviewing the back-
ers of the campaigns. Their results showed that backers are
chiefly concerned with the legitimacy of the projects. Build-
ing on this body of research, we select success determinants
and evaluate them against the success metric.
Data and Extracted Features
Data Collection. We first crawled all the available crowd-
funding campaigns on GofundMe.com. As the example
shown in Figure 1, we were able to collect 10,974 crowd-
funding campaigns around the world. Since it is not possible
to differentiate the donation currency when the fundraiser is
from another country other than US, we decided to use only
U.S. data in our analysis, which amounts to 8,355 U.S. cam-
paigns on the GoFundMe.com.
Figure 1: A Campaign on GoFundMe.com (recognizable
faces are masked to preserve privacy).
Crawled Features. Table 1 shows the extracted features
directly crawled from the website. Dynamic features such
as the number of followers, the number of shares, and the
number of donors are also included.
Table 1: Crawled Features
Launch Date, Location, Title Cover image, Description,
Category, Current Amount, Goal Amount,
# of Followers, # of Shares, # of Donors
Inferred Features. Table 2 shows the inferred features.
•Population: Since the available attributes do not directly
list the population, we infer that from the fundraisers lo-
cation (e.g. Los Angeles) using the US Census Bureau
data (2018).
•Image Quality Assessment: We use a pre-trained model
called NIMA (Lennan, Nguyen, and Tran 2018) to predict
aesthetic and technical quality scores for each cover im-
age. The models were trained via transfer learning, where
ImageNet pre-trained CNNs were used and fine-tuned for
the classification task.
•We use Face++ (faceplusplus.com), which is a face
recognition platform based on deep learning (Yin et

al. 2019). As can be seen from Table 3, for each cover
image, Face++ returns the following values:
Number of faces in the cover image: Numeric, the
number of faces in the cover image
Gender: String, Male or Female
Beauty: Object, attractiveness score given by male and
female evaluators, individually
Smile: Integer, a value between [0,100]
Emotion: Object, contains the values for anger, disgust,
fear, happiness, neural, sadness, surprise
Age: Integer, a value between [0,100]
#IS child: String, Yes or No, a variable shows whether
a child’s face is in the cover image. If a person’s face is
detected and the age is under 10
Table 2: Inferred Features
Population of the Fundraisers Location; # of People in the
Cover Image; Peoples Facial Attributes on the Cover
Image; Technical and Aesthetic Scores of the Cover Images
Methodology
Crowdfunding Success Metrics. We bin the data into
four groups (Figure 2) according to the smoothed goal
amount shown in the distribution in Figure 2, which
reveals four distinctive groups: (0,8000],(8000,40000],
(40000,68000], and (68000,100000]. In each group, we de-
fine the success of a crowdfunding campaign using the ra-
tio of the amount of money that has been raised so far to
the fundraisers goal amount. The ratio is empirically binned
into 4 groups: (0,0.5],(0.5,1],(1,1.25], and (1.25,2.5].
Fundraisers with ratios from 0 to 0.5 are defined as unsuc-
cessful, while those larger than 1.25 but not greater than 2.5
are regarded successful.
Ratio =M oney T hat H as Been Raised S o F ar
Goal Amount
Figure 2: Histogram of the goal amount.
Table 3: Examples of the face attributes.
Gender Female Male
Age 22 18
Emotion
(highest score)
Happiness: 99.99 Neutral: 75.34
Beauty Female: 59.21
Male: 59.03
Female: 75.22
Male: 73.9
Smile Yes No
Image Features. We used a pre-trained model called
NIMA (Lennan et al. 2018) to predict the aesthetic qual-
ity and technical quality of cover images, respectively. The
models were trained via transfer learning, where ImageNet
pretrained CNNs are used and fine-tuned for the image qual-
ity classification task. The predictions (Figure 3) show that
the aesthetic classifier correctly ranks the cover images from
very aesthetic (the rightmost creative art image) to the least
aesthetic (the leftmost image with two boring cars). Sim-
ilarly, the technical quality classifier predicts (Figure 4)
higher scores for visually pleasing images (third and fourth
from the left) versus the images with JPEG compression ar-
tifacts (second) or blur (first).
In order to a better understand the influence of facial at-
tributes on crowdfunding success, we use the prediction out-
comes from Face++. Figure 5 shows the summary statistics
of the Face++ results.
Text Features. We computed 92 LIWC features (e.g. word
categories such as social and affect) to model the text Data
(Pennebaker, Francis, and Booth 2001). These features can
potentially reflect the distribution of the text data.
Fusion Methods. We apply both early fusion and late fu-
sion (You et al. 2016) to predict the crowdfunding suc-
cess using pictorial and textual features. Figure 6 shows the
flowchart of the multimodal data fusion.
Experiments and Discussions
In this section, we investigate the relationship between
fundraiser success and the attributes of the fundraisers. First,
we analyze the category proportion of each goal amount
group. After controlling for category, we dig deep into the
city population, the LIWC features and the image quality.
Campaign Category. Each campaign on GoFundMe be-
longs to one of 19 unique categories (e.g. Weddings & Hon-
eymoons). The number of campaigns in each category is
evenly distributed with the exception of Other and Non-
Profits & Charities. However, the proportion of successful
to unsuccessful campaigns in each category is not uniform.
Weddings & Honeymoons, Competitions & Pageants and
Travel & Adventure are the top three categories that are
most likely to fail in a fundraising campaign that has a goal
amount between $0 and $8000. The top three categories that

Figure 3: Examples of aesthetic score prediction by the MobileNet.
Figure 4: Examples of technical score prediction by the MobileNet.
Figure 5: Face attributes.
are most likely to succeed in a fundraiser from that goal
amount group are Volunteer & Service, Dreams, Hopes &
Wishes and Celebrations & Events. The distribution of the
categories is even in the successful groups, but the Wed-
dings & Honeymoons, Competitions & Pageants and Travel
& Adventure seem to fail more often compared with other
categories. In the goal amount between $8000 and $40000
group, the top three unsuccessful categories become Busi-
ness & Entrepreneurs, Missions, Faith & Church and Sports,
Teams & Clubs. The top three successful categories are Ba-
bies, Kids & Family, Accidents & Emergencies and Funer-
als & Memorials. Regardless of the goal amount, the cate-
gories that are most likely to succeed are health related. As
the goal amount increases, Medical, Illness & Healing and
Funerals & Memorials remain the two categories with the
highest likelihood of success. More donations are made to
the events related to health.
Population. We analyze the fundraisers city population
for each category. We find that only Babies, Kids & Fam-
ily is influenced by the population of the fundraisers city
population, and it is only significant in the $8000 to $40000
goal amount group (p < 0.05). The success ratio in small
towns is significant higher (p < 0.05) than that in big cities
for this category. This suggests that it is easier for fundrais-
ers to achieve their fundraising goals if they are from a
smaller city. We dig deeper by taking a look at the number
of the times a fundraiser gets shared via social media. The
fundraiser from a small city raising money for Babies, Kids
& Family has significantly more shares than the one from a
big city (p < 0.05). A possible explanation is that the peo-
ple from a small city have a stronger sense of belonging or
community than the people from a large city.
Campaign Description. Since category is a main variable
to analyze, we examine the LIWC features in each category
so as to control the influence of a category. As we expected,
some categories of the fundraising campaigns are influenced
by the LIWC features.
Table 4 shows the LIWC features that have a significant
influence on the crowdfundings performance. In the low-
est goal amount group, the results of the Pearson correla-

Figure 6: Flowchart of the data fusion.
Table 4: LIWC features that have significant influences
Goal Category LIWC Examples Mean SD r p-value
$0-$8000
Volunteer & Service time end, until, season 4.08 1.90 -0.205 0.0059
Competitions & Pageants
Clout High: confident 73.02 21.84 -0.181 0.0002
Low: humble
i I, me, mine 3.10 3.50 0.164 0.0007
social mate, talk, they 9.96 4.02 -0.168 0.0005
Animals & Pets insight think, know 1.51 0.87 -0.304 2.8E-05
Sports, Teams & Clubs bio eat, blood, pain 0.86 0.95 0.275 0.0002
health clinic, flu, pill 0.27 0.62 0.251 0.0006
$8000-$40000
Community & Neighbors
shehe she, her, him 1.46 2.23 0.220 4.7E-05
social mate, talk, they 12.75 4.68 0.200 0.0002
affect happy, cried 5.13 2.41 0.180 0.0009
Travel & Adventure insight think, know 1.60 1.01 0.363 0.0007
Dreams, Hopes & Wishes anger hate, kill, annoyed 0.26 0.51 -0.219 0.0008
focusfuture may, will, soon 1.39 0.95 0.256 0.0001
Missions, Faith & Church anx worried, fearful 0.11 0.24 0.305 1.2E-06
Weddings & Honeymoons Clout High: confident 93.63 10.50 -0.390 0.0004
Low: humble
i I, me, mine 0.91 1.56 0.399 0.0007
Creative Arts, Music & Film tentat maybe, perhaps 1.40 0.86 0.195 0.0007
Sports, Teams & Clubs achieve win, success, better 4.26 2.34 -0.196 0.0003
tion indicate that there is a significant positive association
between the crowdfunding success and the projects descrip-
tions (p < 0.0001). In the Volunteer & Service category, a
higher ratio of “time” related words (e.g. end, until, season)
contributes to a lower chance of success in the crowdfund-
ing campaigns. This suggests that people are less likely to
donate to a volunteer activity that has a specific period.
Animals & Pets and Competitions & Pageants are cor-
related to the way the project description is written (p <
0.0001). For the Animals & Pets category, insight is neg-
atively correlated with the success of a crowdfunding cam-
paign, which suggests that people are more willing to donate
for animals and pets if the description includes a direct ex-
pression of the fundraisers feelings.
GoFundMe also allows people to raise money for some-
one else. That is why the description is not always written
by the people who actually need help and is also why the
description is not always written using the first-person pro-
noun. For the Competitions & Pageants category, I is posi-
tively correlated with success while Clout and they are neg-
atively correlated. This suggests that if someone wants to
raise some money for their competitions or pageants, they
should write the description from his/her perspective and try
to avoid asking someone else to write the fund description
for them. They should also try to write the description in a
more humble way.
The reason that bio and health are positively correlated
with the chance of success of Sports, Teams & Clubs is that
these words are more related to health issues. This suggests
that probably the person behind the fundraiser is likely in
need of a medical treatment. As we saw before, the fundrais-
ers about medical treatments are always more likely to re-
ceive donations.
In the second goal amount group, even more categories
are influenced by the LIWC features. shehe, affect and social
are positively correlated with the success of a fundraiser in
the Community & Neighbors category. A higher number of
tentat suggests a higher probability of success in the Creative

Table 5: Image quality features that have significant influences
Goal Category Image Info Mean SD r p-value
$8000-$40000
Competitions & Pageants aesthetic score 5.00 0.40 -0.355 0.0002
technical score 5.72 0.53 -0.295 0.0023
Community & Neighbors technical score 5.28 0.78 0.174 0.0013
Weddings & Honeymoons aesthetic score 4.65 0.56 0.320 0.0042
Figure 7: Randomly selected photos from the successful (left) and unsuccessful (right) groups.
Arts, Music & Film category. tentat is related to words like
perhaps and maybe. If someone wants to raise some money
for his/her artwork and he/she is describing it in a tentative
way, probably that means he/she needs more help. We guess
that is why people are more willing to donate money. In ad-
dition, a description that seems more anxious or more wor-
ried is more likely to help the fundraiser raise more money
if it is about Missions, Faith & Church.
If people want to raise some money for their trips, their
plan is more likely to be successful if they can convince the
donors that their project has a far-reaching purpose.
Anger is negatively correlated with a fundraiser’s suc-
cess in the Dreams, Hopes & Wishes category. However,
focusfuture is positively related. We think this is because
that dreams are positive things and they make people feel
warm, energetic, and excited. When people write a descrip-
tion about dreams, they should not involve words that have
negative energy. In addition, they should focus more on the
future.
A higher Clout value indicates a more confident descrip-
tion and a lower value indicates a more humble description.
As the proposed analysis shows, categories like weddings
and honeymoons are not easy to get donation. People proba-
bly favor humble couples. It seems reasonable that a higher
value of I increases the probability of success. For example,
the descriptions that have a high value of i mainly describe
how one of the couple wants to prepare something nice for
his/her partner, therefore we can find many first-person pro-
nouns in that description. When sweet language is used, we
find that campaigns are more likely to receive donations.
For the Sports, Teams & Clubs category, it is counter-
intuitive that the descriptions with more achieve words have
a negative influence on donation. Normally, people would
love to see someone with the ambition and desire to win
when it comes to sports. It is really interesting that based on
our data, those people actually receive less donation. For this
part, we still do not have a convincing explanation other than
the suspicion that overstating may actually turn third-party
people off.
We have not found enough evidence to conclude that there
is relationship between the description and the success of a
fundraiser in the third or the fourth group.
Image Quality. Table 5 shows the image quality features
that have a significant influence on the fundraisers perfor-
mance. We found that Competitions & Pageants, Commu-
nity & Neighbors, and Weddings & Honeymoons are the
only three categories that are influenced by image qual-
ity. This suggests that the success of fundraisers is related
to their cover image quality if their fundraising purpose
falls into one of these three categories. Take the Weddings
& Honeymoon category as an example. Figure 7 shows
several randomly selected photos from the successful and
unsuccessful groups. The ones from the successful group
are on the left, and the ones from the unsuccessful group
are on the right. By looking at them, we find that cam-
paigns with higher success rates have fancy cover images
and the fundraisers apparently made conscious efforts in tak-
ing those photos. In contrast, campaigns with lower success
rates have casual selfie cover images. However, it is still un-
clear why the image quality is negatively correlated to the
success of fundraisers in the Competitions & Pageants cate-
gory. Perhaps overdoing the cover photos make people think
that the activity should already be well funded.
Face Attributes. Table 6 shows face attribute features that
have significant influences on the crowdfundings perfor-
mance. In the ($0,$8000] goal amount group, a smaller
number of faces corresponds to a higher chance to succeed
in a crowdfunding related to competitions and pageants. In
fact, the mean number of faces of the most unsuccessful

Table 6: Facial attributes that have significant influences
Goal Category Face++ Mean SD r p-value
$0-$8000 Competitions & Pageants Num face 3.39 5.48 -0.141 0.0037
$8000-$40000
Medical, Illness & Healing Num face 1.36 1.31 0.344 0.0082
Age 27.17 10.39 0.349 0.0073
Animals & Pets is child 0.01 0.09 0.474 2.3E-16
Travel & Adventure is child 0.02 0.15 0.458 1.2E-05
Missions, Faith & Church Age 24.18 18.97 0.168 0.0085
Table 7: Comparison of accuracy, precision, recall, and F-score
Goal Amount Accuracy Precision Recall F-score
$0-$8000
Basic 0.43 0.40 0.43 0.40
LIWC 0.45 0.47 0.45 0.43
Face++ 0.44 0.34 0.44 0.37
Basic+LIWC+Face++ 0.51 0.49 0.51 0.49
Late Fusion 0.49 0.46 0.49 0.45
$8000-$40000
Basic 0.66 0.66 0.66 0.66
LIWC 0.73 0.72 0.73 0.71
Face++ 0.62 0.43 0.62 0.50
Population 0.30 0.21 0.30 0.24
Image Quality 0.83 0.88 0.83 0.81
B+LIWC+F+P+I 0.76 0.71 0.76 0.72
Late Fusion 0.74 0.73 0.74 0.71
$40000-$68000 Basic 0.72 0.75 0.72 0.72
$68000-$100000 Basic 0.6 0.55 0.6 0.57
Total(Weighted)
Basic 0.58 0.57 0.58 0.57
Early Fusion 0.65 0.61 0.65 0.62
Late Fusion 0.63 0.61 0.63 0.60
group where the ratio is between 0 and 0.5 is 4.07, and that
of the most successful group where the ratio is between 1.25
and 2.5 is 2.41. In the ($8000,$40000] goal amount group,
we find that the number of faces and age are positively corre-
lated with the crowdfunding performance if it is about med-
ical, illness and healing. In the most successful group, the
mean number of faces and the age is 1.81 and 34.31, respec-
tively. For the most unsuccessful group, the mean is 1.00 and
21.50, respectively. We analyze some cover images and find
that donors respond positively to the family photos which
were taken before the accidents happened. More people and
faces of the elders in the cover image might make the donors
feel more sympathy. However, what is interesting here is that
the number of faces does not work the same way in the com-
petitions and pageants category as it does in the medical and
healing category. We think the reason is that they are dif-
ferent categories. Recall in the previous sections, we found
that medical fundraisers are almost always the most success-
ful. People do care about others and are willing to donate, if
it is urgent or about life and death. People are less likely
to donate to less urgent events like weddings and competi-
tions. For ”Animals & Pets” and ”Travel & Adventure”, we
find that the appearance of a child boosts donations. Prob-
ably that is because that the thought of children and those
activities make people feel the need to support.
Classification Evaluation. In this section, we conduct
classification experiments to evaluate the effectiveness of
our proposed features using the Random Forest model. Con-
sidering that we take the ratio group as our output, we choose
different features as the input to perform the classification:
•Basic Information with Category. The input of the
model only includes basic information like launch date,
city, state and category information. This is the baseline
model.
•Single LIWC. The input of the model only includes
LIWC features of each category. Specifically, we choose
different LIWC features in each category based on our
proposed analysis.
•Single City Population. The input of the model only in-
cludes the city population of each category. Specifically,
we choose the data of category based on our proposed
analysis.
•Single Face++. The input of the model only includes the
Face++ features of each category. Specifically, we choose
different Face++ features in each category based on our
proposed analysis.
•Single Image Quality. The input of the model only in-
cludes the image quality features of each category. Specif-
ically, we choose different image quality features in each
category based on our proposed analysis.

•Early Fusion. The text description and image informa-
tion are combined as the input.
•Late Fusion. The text description and image information
are used to construct separate models, respectively, before
a final decision is combined.
For all the above settings, we employ Random Forest as
the classifier. We show the quantitative results of our exper-
iments in Table 8. The performance of different models is
evaluated by four metrics: accuracy, precision, recall, and
F-measure. During the experiments, the number of estima-
tors is set to 1000 for every setting, and each setting is run 5
times and the results are averaged. We conduct experiments
in each goal amount group, and calculate the weighted met-
rics.
As the table shows, the baseline models with the basic
information are the worst at prediction in each group. In
contrast, adding extra information can always increase the
classification performance.
In the $0-$8000 group, the performance of the models us-
ing the basic information and the models using LIWC or
Face++ is quite comparable. The one that combines all of
the features is the best according to all metrics.
In the $8000-$40000 group, the models using LIWC are
relatively better than the models using the basic information
or Face++. City population is not really useful during the
classification. The aesthetic score and technical score of the
cover image are really helpful for making a classification. It
is even a surprise that the Single Image Quality setting is the
best at classification.
Since there is no sufficient evidence to conclude the re-
lationship between features and success within the last two
groups, we use the basic information as the input. As we can
see from the table, it is easier to make a reliable classification
within a higher goal amount group.
The weighted metrics are shown in the table. Early fusion
and late fusion are both better than the baseline, and early
fusion is the best choice.
Conclusion
In this study, we focus on understanding and predicting
the performance of the crowdfunding campaigns on Go-
FundMe, which is diverse in funding categories and charity-
minded. We analyze the attributes available at the launch
of the campaign and identify attributes that are important
for the major campaign categories. Furthermore, we have
bench-marked several computational models and identified
a multi-modal fusion classifier that significantly improves
the prediction result. We believe that the findings and mod-
els from this study provide effective mechanisms to make a
crowdfunding campaign successful in different categories.
References
Cordova, A.; Dolci, J.; and Gianfrate, G. 2015. The Deter-
minants of Crowdfunding Success: Evidence from Technol-
ogy Projects. Procedia-Social and Behavioral Sciences 181,
115-124.
Etter, V.; Grossglauser, M.; and Thiran, P. 2013. Launch
Hard or Go Home! Predicting the Success of Kickstarter
Campaigns. In Proceedings of the first ACM conference on
Online Social Networks, 177-182. ACM.
Greenberg, M. D.; Pardo, B.; Hariharan, K.; and Gerber,
E. 2013. Crowdfunding support tools: Predicting success
& failure. CHI13 Extended Abstracts on Human Factors in
Computing Systems, 18151820. ACM.
Brown, W.M.; Palameta, B.; and Moore, C. 2003. Are there
nonverbal cues to commitment? An exploratory study using
the zero-acquaintance video presentation paradigm. Evolu-
tionary Psychology, Vol. 1 No. 1, pp. 4269.
Gabriel, A.S.; Acosta, J.D.; and Grandey, A.A. 2015. The
value of a smile: Does emotional performance matter more
in familiar or unfamiliar exchanges?. Journal of Business
and Psychology, Vol. 30 No. 1, pp. 3750.
Grandey, A.A.; and Gabriel, A.S. 2015. Emotional labor at
a crossroads: Where do we go from here? Annual Review
of Organizational Psychology and Organizational Behavior,
Vol. 2 No. April, pp. 323349.
Lau, S. 1982. The effect of smiling on person perception.
The Journal of Social Psychology, Vol. 117 No. 1, pp. 6367.
Mehu, M.; Little, A.C.; and Dunbar, R.I. 2008. Sex differ-
ences in the effect of smiling on social judgments: An evolu-
tionary approach. Journal of Social, Evolutionary, and Cul-
tural Psychology, Vol. 2 No. 3, pp. 103121.
Lee, Y. H.; and Lim, E. A. C. 2010. When good cheer goes
unrequited: How emotional receptivity affects evaluation of
expressed emotion. Journal of Marketing Research, 47(6),
1151-1161.
Barger, P.B.; and Grandey, A.A. 2006. Service with a Smile
and Encounter Satisfaction: Emotional Contagion and Ap-
praisal Mechanism. Academy of Management Journal, 49,
1229-1238.
Kim, J.; and Park, J. 2017. Does Facial Expression Matter
Even Online? An Empirical Analysis of Facial Expression
of Emotion and Crowdfunding Success.
Johansen, C. A. 2019. Variables that Influence Crowdfund-
ing Success.
Kuppuswamy, V.; and Bayus, B. L. 2015. Crowdfunding
Creative Ideas: The Dynamics of Project Backers in Kick-
starter (SSRN Scholarly Paper No. ID 2234765).
Mitra, T.; and Gilbert, E. 2014. The Language That Gets
People to Give: Phrases That Predict Success on Kickstarter.
In Proceedings of the 17th ACM conference on Computer
supported cooperative work & social computing, 49-61.
ACM.
Mollick, E. 2014. The Dynamics of Crowdfunding: An Ex-
ploratory Study. Journal of Business Venturing 29(1), 1-16.
Lennan, C.; Nguyen, H.; and Tran, D. 2018. Image quality
assessment.
Parhankangas, A.; and Renko, M. 2017. Linguistic Style and
Crowdfunding Success among Social and Commercial En-
trepreneurs. Journal of Business Venturing 32(2), 215236.
Pennebaker, J.; Francis, M.; and Booth, R. 2001. Linguistic
Inquiry and Word Count (LIWC): LIWC2001. 71.
Xiao, S.; Tan, X.; Dong, M.; and Qi, J. 2014. How to Design
Your Project in the Online Crowdfunding Market? Evidence
from Kickstarter. ICIS.
Xu, A.; Yang, X.; Rao, H.; Fu, W. T.; Huang, S. W.; and Bai-
ley, B. P. 2014. Show Me the Money: An Analysis of Project

Updates during Crowdfunding Campaigns. In Proceedings
of the SIGCHI conference on human factors in computing
systems, 591-600. ACM.
Yuan, H.; Lau, R. Y.; and Xu, W. 2016. The Determinants
of Crowdfunding Success: A Semantic Text Analytics Ap-
proach. Decision Support Systems 91, 67-76.
You Q.; Jin, H.; Yang, J.; and Luo, J. 2016. Cross-modality
Consistent Regression for Joint Visual-Textual Sentiment
Analysis. In Proceedings of the International Conference on
Web Search and Data Ming (WSDM). ACM.