Predicting Reputation Score of Users in Stack-overﬂow with Alternate

Data

Sahil Yerawar

, Sagar Jinde

, P. K. Srijith

, Maunendra Sankar Desarkar

, K. M. Annervaz

and Shubhashis Sengupta

Indian Institute of Technology Hyderabad, India

Accenture Technology Labs, India

annervaz.k.m@accenture.com, shubhashis.sengupta@accenture.com

Keywords:

Recommendation System, Cold-start Recommendation, Knowledge Distillation, Transfer Learning.

Abstract:

The community question and answering (CQA) sites such as Stack Overﬂow are used by many users around

the world to obtain answers to technical questions. Here, the reliability of a user is determined using metrics

such as reputation score. It is important for the CQA sites to assess the reputation score of the new users

joining the site. Accurate estimation of reputation scores of these cold start users can help in tasks like

question routing, expert recommendation and ranking etc. However, lack of activity information makes it

quite difﬁcult to assess the reputation score for new users. We propose an approach which makes use of

alternate data associated with the users to predict the reputation score of the new users. We show that the

alternate data obtained using users’ personal websites could improve the reputation score performance. We

develop deep learning models based on feature distillation, such as the student-teacher models, to improve

the reputation score prediction of new users from the alternate data. We demonstrate the effectiveness of the

proposed approaches on the publicly available stack overﬂow data and publicly available alternate data.

1 INTRODUCTION

A large majority of internet users refer to commu-

nity question answering (CQA) sites to gather infor-

mation, knowledge and general consensus views for

their own purposes. Examples of such CQA sites in-

clude Reddit, Stack Exchange, Quora etc. The suc-

cess of these community sites are mostly attributed to

efﬁcient moderation of user questions and answers, as

well as the presence of a large number of users who

consistently provide high quality content. The relia-

bility of these consistent users is quantiﬁed by using

a points based system (Cavusoglu et al., 2015), where

the user is encouraged to interact with the community

by providing comments, posts, questions and answers

to earn their points which is also a measure of trust

and expertise. In Stack Overﬂow, the points earned

by the user are often called the reputation score (Over-

ﬂow, 2022b). The reputation score of the users can be

utilized by the system to ﬁnd expert users to answer

questions, rank the answers by users, and in provid-

ing some privileges. However, when a new user joins

the community QA website, due to lack of any regis-

tered activity, they will be assigned a very low repu-

tation score and consequently their expertise can not

be tapped completely by the system.

In this work, we aim to predict the reputation score

new users could have achieved, so that their expertise

can be fully utilized by the system to get expert an-

swers to the questions, and provide a better user ex-

perience (Slag et al., 2015). We address this by con-

sidering alternate data provided by users such as their

publicly available website information. In addition

to the activity based features like views, upvotes and

downvotes, stack overﬂow often contains user’s pro-

ﬁle information and publicly available website URLs

such as personal websites and GitHub accounts. We

believe that by leveraging the content of the personal

website, we can determine technical expertise, work

and education backgrounds of the user. Consequently,

this can be used to predict an approximate reputation

score, which can be used by the system to provide a

better user experience. This would help in providing a

fair estimate of the users’ expertise in the form of rep-

utation score, which can help in recommending new

users with a certain level of proﬁciency to different

questions in Stack Overﬂow. The proposed approach

addresses the cold start problem of assigning reputa-

Yerawar, S., Jinde, S., Srijith, P., Desarkar, M., Annervaz, K. and Sengupta, S.

Predicting Reputation Score of Users in Stack-overﬂow with Alternate Data.

DOI: 10.5220/0011591900003335

In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 1: KDIR, pages 355-362

ISBN: 978-989-758-614-9; ISSN: 2184-3228

355

tion scores for new users, which would help the down-

stream Recommender Systems for CQA websites.

To predict the reputation score of new users, we

develop deep learning models which could predict

the reputation score from the alternate data. Another

major contribution of our work is to develop effec-

tive deep learning models which learns not only from

the alternate data but also from the activity informa-

tion by treating them as privileged information and

performing feature distillation using a teacher-student

learning technique (Hinton et al., 2015; Lopez-Paz

et al., 2016; Xu et al., 2019). We propose an auto-

encoder based teacher and student model which al-

lows transfer of knowledge from a teacher network

trained on activity information and alternate data, to

a student network trained on alternate data alone. We

propose several ways to transfer the knowledge be-

tween teacher and student for effective learning. We

demonstrate ﬁrst the usefulness of the alternate data

in predicting the reputation score. Following this,

we demonstrate the effectiveness of the proposed ap-

proaches in predicting the reputation score of the new

users.

The contributions of this paper are three-fold:

• We propose to use alternate data such as personal

website information to improve reputation score

prediction in stack overﬂow and use them in par-

ticular to predict reputation score for new users.

• We develop an auto-encoder based teacher and

student model and feature distillation techniques

to predict the reputation score of users from alter-

nate data.

• We conducted experiments on the publicly avail-

able stack overﬂow data and publicly available al-

ternate data, demonstrating usefulness of the al-

ternate data and proposed methodologies for rep-

utation score prediction.

2 RELATED WORK

2.1 Cold Start in Recommendation

Collaborative Filtering (CF) in Recommendation Sys-

tems have given good performances in recommending

items to users who have past user-item interaction his-

tory. Traditionally, Collaborative Filtering involves

matrix factorization to obtain low level embeddings

of user and item vectors. However, most CF algo-

rithms fail to work in the case of discovering inter-

actions between both new users and new items (Lam

et al., 2008; Park and Chu, 2009).

To alleviate the drawbacks of CF-based algo-

rithms in the context of cold start problem, many

researchers also utilize content information of these

users to formulate content based methods. These

methods try to learn an appropriate new user/item rep-

resentation based on content information and learn a

mapping to convert these representations into user-

item interaction vector space. However, these meth-

ods fail to model the complex nature of interaction be-

tween embedding space and the user content embed-

ding space, due to which a new user embedding can-

not associate with the embedding space. More recent

methods to address cold start problem involve a hy-

brid approach of using CF and content based methods

(Sujithra Alias Kanmani et al., 2021; ?), using meta-

learning (Zheng et al., 2021; Chen et al., 2021; Wang

et al., 2021a; Du et al., 2022), using pre-training (Hao

et al., 2021) etc. However, most of these approaches

aim to produce a good ranking of recommended items

for the end users, and do not focus on ﬁnding the true

value of the cold-start item or the true activity-based

user proﬁle of the cold-start user.

2.2 Stack Overﬂow

Stack Overﬂow is one of the largest community

question-answer (CQA) websites, playing a crucial

role in the daily activities of numerous researchers,

developers and scholars all over the world. In Stack

Overﬂow, the point based system is known as rep-

utation score which can be earned by asking good

questions and providing answers on the Stack Over-

ﬂow website. Subsequently, there have been numer-

ous studies on Stack Overﬂow reputation score and

how it behaves for different users. However, to the

best of our knowledge, there is no work which aims

to predict the reputation score of new users based on

the alternate data such as the website information.

Movshovitz-Attias et al. (Movshovitz-Attias

et al., 2013) studied participation patterns of expert

and non-expert users present in StackOverﬂow repu-

tation system. Their study concluded that inﬂuential

users have a good contribution during the ﬁrst month

of activity and that expert users contribute drastically

more as soon as they join the site. They also found

that high reputation users were the primary sources of

high-quality answers. The study in Anderson et al.

(Anderson et al., 2012) looked at the way the com-

munity determines the answers, including how votes

are cast. The authors found that the answerer’s rep-

utation correlates with the speed of answers and that

the arrival time of an answer impacts its chances of

being chosen as the best one. In Slag et al. (Slag

et al., 2015), the authors found that no response to

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

356

a question or unsatisfactory answer is the main issue

a new user faces. Due to this reason, the new users

may get demotivated and stop using the site. Their

analysis showed that half of them have just one con-

tribution on the site, and 60% of users have a reputa-

tion score of one. These points towards the need to

identify expert users early, get satisfactory answers,

provide more privileges, and encourage them to par-

ticipate in the QA site.

In Morrison et al. (Morrison and Murphy-Hill,

2013), the authors found a positive correlation be-

tween reputation score and age. They also examined

users’ familiarity with various skills and technolo-

gies using tags. MacLeod et al. (MacLeod, 2014)

performed an exploratory analysis of stack exchange

data. They found that the reputation score of a user is

positively correlated to the number of tags. Users with

high reputation scores contribute to a diverse number

of topics. Recent approaches to user reputation score

prediction (Woldemariam, 2020; Banati and Seema,

2021) made use of the textual content of the user posts

and its syntactic and semantic information in deter-

mining the reputation score of the users. Unlike all

these previous works, our aim is to develop an ap-

proach to predict the reputation score of cold start

users in the question answering sites.

3 BACKGROUND AND PROBLEM

DEFINITION

3.1 Features from the Dataset

In this work, we want to utilize stack overﬂow dataset

for identifying the user reputation scores. We try to

develop models that can predict the user’s reputation

scores based on the alternate data. This can help the

QA systems to make use of the expertise of new users

in providing relevant answers to questions posted by

users. We divide the features available in the stack

overﬂow dataset (Overﬂow, 2022a) as follows:

• Activity Information: The features reﬂect the ac-

tivity of the Stack Overﬂow user in the system. It

includes the number of views, number of upvotes

received, number of downvotes received and the

number of years the user has been active. We cal-

culate the number of years by subtracting the user

creation date from last accessed date present in

the dataset. We call these features the privileged

features which the QA systems use to compute the

reputation score. The new users joining the sys-

tem lack these features, while only old users will

have these features.

• About Me Text: Stack Overﬂow users can write

about themselves and their interest in a separate

section named ”About Me”. This textual data

could be utilized as our alternate data which pro-

vides an indication about the user’s capabilities.

All the data present in ”About Me” section is pub-

licly available.

• Website Text: In Stack Overﬂow, users can also

share their personal websites containing more in-

formation about their skill sets and achievements.

This textual data can act as another source of al-

ternate data which could be of great help in repu-

tation score prediction. All the data obtained from

personal websites are publicly available.

• Website Category: There are various kinds of

websites provided by the users. We make a broad

classiﬁcation of these websites into three cate-

gories: Social media websites, Academic pages

and Personal blogs. We consider this information

also as part of the alternate data.

3.2 Problem Deﬁnition

We consider the problem predicting the reputation

score of a user y ∈ R , given their alternate data

based features X ∈ R

and privileged activity fea-

tures X

∗

∈ R

. Assume we are given a training

data set D = {X

, X

∗

, y

}

i=1

which consists of users

for whom both the privileged data and alternate data

are available (experienced users), and test data

D =

{

, ˆy

}

i=1

which consists of users for whom only al-

ternate data are available, i.e. new users. We aim to

learn a regression model f : X → y to predict the rep-

utation score of users using alternate data, but also

make use of the availability of the privileged data X

∗

during training.

3.3 Learning using Privileged

Information

Standard machine learning and deep learning tech-

niques assume the same set of features to be avail-

able with all the users. However, it is different in

our case where we have a set of users for whom we

have activity (privileged) data while the rest does not

have. The recent works along the direction of learning

from privileged information (LUPI) (Pechyony and

Vapnik, 2010; Vapnik and Vashist, 2009; Vapnik and

Izmailov, 2015) have shown that making use of the

additional privileged features associated with some

subset of examples could improve the generalization

performance of the learning algorithms. Therefore,

Predicting Reputation Score of Users in Stack-overﬂow with Alternate Data

357

Figure 1: Model Architecture of Teacher and Student model with direct constraints on the latent representations.

we develop our models for reputation score predic-

tion using the framework of learning using privileged

information (Hinton et al., 2015; Lopez-Paz et al.,

2016) and more speciﬁcally the feature distillation ap-

proach (Xu et al., 2019; Wang et al., 2021b).

A more general framework of knowledge distilla-

tion has been used for several problems which involve

transfer of knowledge from one model to another.

In particular, knowledge distillation (Hinton et al.,

2015; Mirzadeh et al., 2020; Zhou et al., 2018) con-

sider two models: teacher f

() with its weights W

and

student f

() with model parameters W

. The teacher

model is more complex than the student model. The

goal of this technique is to improve the training of the

student model with the help of predictions done by

the teacher model. Knowledge distillation is usually

employed in cases where the teacher model is very

complex in architecture and we want to emulate its

prediction distribution within the more simpler stu-

dent model. For the standard model distillation, the

loss is governed by the following objective function:

min

∑

i=1

, f

)

+ λ ∗ L

( f

), f

)

(1)

The loss component L

() is the standard loss (e.g root

mean square error for regression) while L

helps in

distilling the knowledge learned by the teacher and

aids in better training of W

. For instance, through

, the student model tries to match the prediction

(X;W

) generated by the teacher model. Here λ is a

regularization constant that controls the weight asso-

ciated with each loss term.

We consider a different setup of knowledge dis-

tillation known as feature distillation (Vapnik and

Vashist, 2009; Xu et al., 2019) where the feature set

being visible to the student and teacher is different.

It assumes the teacher has access to both X and X

∗

while the student has access to only X. In feature

distillation, the aim of the student is to learn to pre-

dict only using feature X but with the help of the ad-

ditional knowledge available with the teacher model.

The learning of the student is done through feature

distillation where the following loss is used.

min

∑

i=1

, f

)

+λ ∗ (L

( f

, X

∗

), f

))

(2)

Here, L

() is the standard loss (root mean square er-

ror) and L

() is the distillation loss. The teacher is

trained using the standard loss alone but using both

the features X and X ∗. Through feature distillation,

we aim the student to not only predict the output from

X but also learn representations similar to the teacher

through the distillation loss term. The teacher is ex-

pected to have better representation learning capabil-

ity as it learns from both the features and can assist the

student model to learn better representations through

the distillation process.

4 MODELS FOR PRIVILEGED

LEARNING OF REPUTATION

SCORES

We provide a detailed description of the teacher-

student model and the distillation techniques used to

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

358

transfer knowledge from teacher to the student in pre-

dicting the reputation score of users.

We use three models to aid the learning process -

the teacher model f

: (X, X

∗

) → y which has access

to both privileged and common features, the baseline

model f

: X → y which has access to only common

features and no information provided by the teacher,

and student model f

: X → y which can access only

common features and is guided by the teacher. All the

models are based on auto-encoders, which learns a la-

tent representation of the input using an encoder and

then tries to reproduce the input through an decoder.

The auto-encoder allows one to learn a good latent

representation of the features and we intend to trans-

fer knowledge through these latent representations.

The latent representation is used to predict the reputa-

tion score using a neural network. In all these models,

the reputation score prediction from the latent repre-

sentation is done using a two layer neural network.

The teacher will be able to learn to predict reputation

scores from both the alternate data and privileged data

while the student has access to only alternate data,

but we intend to improve student learning using the

teacher.

We transfer the knowledge from teacher to stu-

dent by constraining the latent space of the teacher

and student to be similar. This is achieved by putting

an RMSE loss directly on the latent representations

of the student z

and teacher z

. We follow two varia-

tions of training this student-teacher model: 1) train-

ing student and teacher simultaneously (co-train), 2)

training the teacher model ﬁrst and then the student

model with a constraint on the latent representations

(pre-train). Figure 1 represents the overall architec-

ture of the proposed model.

4.1 Loss Function

In the ﬁrst case (co-train), we use joint loss functions

over both teacher and student models along with the

RMSE loss over the latent space (L

constr

()) which en-

sures similarity.

min

∑

i=1

, f

, X

∗

))

+ λ

rec

((X

, X

∗

), (X

, X

∗

)) + L

, f

))

+ λ

rec

, X

) + λ

constr

, z

))

(3)

where (X

, X

∗

) and X

are reconstructed input by

the decoder of teacher and student networks respec-

tively. Here, L

rec

() is the auto-encoder reconstruction

loss (RMSE) between the output of the decoder and

input of the encoder. The hyper-parameters λ

, λ

control the degree of regularization through the recon-

struction loss and strength of transfer.

In the second case (pre-train), we train the teacher

model ﬁrst using the following loss

min

∑

i=1

, f

, X

∗

))

+ λL

rec

((X

, X

∗

), (X

, X

∗

))

(4)

Then, we train the student model with the following

loss which includes the constraint loss component.

min

∑

i=1

, f

))

+λ

rec

, X

) + λ

constr

, z

)

(5)

4.2 Experimental Details

We perform the experiments on a subset of the Stack

Overﬂow dataset. From 13 million users, we select

65957 users who have valid about me, website data

and have their reputation scores to be between 2 and

10000, both inclusive. We exclude users with repu-

tation score value of 1 as every new user is assigned

that score at the time of creation and signiﬁes no user

activity for a lot of users. We use LDA (Blei et al.,

2003) and GLoVE (Pennington et al., 2014) to con-

vert about me and textual data into embedding vec-

tors. As presented in section 3.1, we also use the

website category feature, which is a vector of size 3.

We split our dataset into train split and test split in

about 80:20 ratio and use the results as reported on the

test data. We use the root mean square error (RMSE)

metric to evaluate the reputation score prediction ca-

pability of the model. The teacher model will have

access to both the privileged and alternate data dur-

ing evaluation while the student will have access to

only alternate data. The experiments use a 5 fold

cross validation technique on the train split and the

hyper-parameters are selected using grid search. We

also consider Bayesian optimization (Shahriari et al.,

2016) to get the optimal hyper-parameters for the co-

train model. The batch size for all the experiments

is 64 and the learning rate is 0.001 which is used in

Adam Learning Optimizer (Kingma and Ba, 2014). In

Figure 1, while using LDA we consider input vectors

of size 20 for both about me and web data features

(and 100 for Glove), a 3 dimensional web category

and a 4 dimensional vector of activity features. The

teacher encoder will have 4 fully connected layers of

size 30, 20, 15 and 10. We provide the activity feature

vector in the second layer of the encoder. The teacher

decoder consists of fully connected layers of size 15,

24, 30 and 43. The reputation prediction module con-

sists of two fully connected layers of size 5 and 1.

Predicting Reputation Score of Users in Stack-overﬂow with Alternate Data

359

Table 1: Prediction errors (RMSE) of the models on the stack overﬂow data.

Model Training RMSE (LDA) RMSE (GLoVE)

Feed Forward Network only activity features 943.6

Baseline Auto-encoder only alternate data 2181.4 2266.5

Teacher Pre-train and 608.5 756.5

Student Grid search 1960.0 1968.8

Teacher Co-train and 626.6 757.5

Student Grid search 1937.6 2173.5

Teacher Co-train and 589.7 649.1

Student Bayesian optimization 1949.3 2130.2

The student component of the model is very similar to

that of the teacher, except that the second layer of the

encoder and third last layer of decoder is of size 20,

since we are not passing any activity features directly

to the student. For reference, we have developed a

simple autoencoder network which uses only activity

features to predict the reputation score. The layers in

the encoder are fully connected layers of size 4,3,3,2

and that of decoder is 3,3,4,4. The reputation predic-

tion module consists of two fully connected layers of

size 2 and 1.

The hyper-parameters selected are as follows:

• For the Teacher model, the learning rate is 0.01,

batch size is 128, number of epochs is 100 and the

λ used in reconstruction loss is 0.95.

• For the Teacher assisted student model, the learn-

ing rate is 0.001, batch size is 128, number of

epochs is 100 and the λ used in reconstruction loss

is 0.8.

4.3 Results

Table 1 shows the results of the various models used

for reputation prediction. Note that all reported values

correspond to the RMSE errors between reputation

prediction and actual score. When comparing the re-

sults of the feed forward model with the teacher base

model, it is clear that a combination of activity and

alternate features gives better results than considering

only activity features. We can also see that for all the

proposed student models, the RMSE is less than the

baseline auto-encoder without any distillation. This

shows the effectiveness of our distillation approach

and the student auto-encoder. Although feature dis-

tillation reduces RMSE, the gap between teacher and

student models still needs some improvement.

Comparing the two representations of the text,

we observe that the LDA based representation gives

a better performance than the Glove representation.

This is possibly due to two reasons. The dimension-

ality of GLoVE embeddings is higher than the LDA

based embeddings. This may distract the model from

the activity based features which are lesser in num-

ber. The student and the distilled model learn better

for LDA as it has a lesser number of dimensions. Sec-

ondly, LDA tries to identify the topic distribution in

the text holistically by considering the entire text to-

gether. On the other hand, the GLoVE based method

aggregates the information from individual words to

get the ﬁnal vector. Thus, LDA is more useful in cap-

turing the expertise of the user and consequently pre-

dict the reputation score.

Table 2 shows a case study involving two users

from the dataset. Both the users have a reasonable

amount of alternate data available on their website.

Our student model could effectively learn from the al-

ternate data and has assigned reasonably good scores

to both of them, higher scores to one with higher

ground truth reputation score and vice-versa. We

see many more such examples from our experiments

which support the applicability of the approach.

5 CONCLUSIONS

We have developed an auto-encoder based regression

model and utilized privileged distillation to transfer

the knowledge of teachers, to assist the training of

the student model. We consider two different train-

ing procedures for the proposed model. The proposed

approaches were found to be effective in predicting

reputations score of users in the stack overﬂow data.

We found the alternate data to provide additional in-

formation which improved the reputation score pre-

diction performance. Further, we found the student

model learnt using feature distillation to also improve

the performance compared to the baseline model. As

a future work, we would like to further reduce the per-

formance gap between student and the teacher in pre-

dicting the reputation score.

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

360

Table 2: Actual and predicted scores for two users from the test split of dataset using student model with LDA input vectors.

The examples are tken from public data dump of the Stack Exchange forum.

User

name

About Me Web Data Prediction Target

****** https://www.*****.co

How To Watch For Changes Tuesday March

30 2021 5 minute read I needed to customize

the reviews UI on a product detail page for a

Shopify site whenever there were no reviews.

However, I ran into a series of unexpected

problems ...

1227.6 3114

******* Computer ***** Stu-

dent at Georgia Tech

A 60 GHz phased arrayfor $10 A 60 GHz

phased array for $10 In 2018, I gave an talk

that was basically Phased Arrays 101...

628.7 1223

REFERENCES

Anderson, A., Huttenlocher, D., Kleinberg, J., and

Leskovec, J. (2012). Discovering value from com-

munity activity on focused question answering sites:

A case study of stack overﬂow. In Proceedings of

the 18th ACM SIGKDD International Conference on

Knowledge Discovery and Data Mining, KDD ’12,

page 850–858, New York, NY, USA. Association for

Computing Machinery.

Banati, H. and Seema (2021). Proﬁciency assessment of

experts in online social network q/a communities. In

2021 9th International Conference on Reliability, In-

focom Technologies and Optimization (Trends and Fu-

ture Directions) (ICRITO), pages 1–5.

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent

dirichlet allocation. volume 3. JMLR.org.

Cavusoglu, H., Li, Z., and Huang, K.-W. (2015). Can gam-

iﬁcation motivate voluntary contributions? the case

of stackoverﬂow q&a community. In Proceedings

of the 18th ACM Conference Companion on Com-

puter Supported Cooperative Work & Social Comput-

ing, CSCW’15 Companion, page 171–174, New York,

NY, USA. Association for Computing Machinery.

Chen, Z., Wang, D., and Yin, S. (2021). Improving cold-

start recommendation via multi-prior meta-learning.

In European Conference on Information Retrieval,

pages 249–256. Springer.

Du, Y., Zhu, X., Chen, L., Fang, Z., and Gao, Y. (2022).

Metakg: Meta-learning on knowledge graph for cold-

start recommendation. IEEE Transactions on Knowl-

edge and Data Engineering.

Hao, B., Zhang, J., Yin, H., Li, C., and Chen, H. (2021).

Pre-training graph neural networks for cold-start users

and items representation. In Proceedings of the 14th

ACM International Conference on Web Search and

Data Mining, pages 265–273.

Hinton, G., Vinyals, O., and Dean, J. (2015). Distilling

the knowledge in a neural network. In NIPS Deep

Learning and Representation Learning Workshop.

Kingma, D. and Ba, J. (2014). Adam: A method for

stochastic optimization. International Conference on

Learning Representations.

Lam, X. N., Vu, T., Le, T. D., and Duong, A. D. (2008). Ad-

dressing cold-start problem in recommendation sys-

tems. In Proceedings of the 2nd international con-

ference on Ubiquitous information management and

communication, pages 208–211.

Lopez-Paz, D., Sch

olkopf, B., Bottou, L., and Vapnik, V.

(2016). Unifying distillation and privileged informa-

tion. In International Conference on Learning Repre-

sentations (ICLR).

MacLeod, L. (2014). Reputation on stack exchange: Tag,

you’re it! In 2014 28th International Conference on

Advanced Information Networking and Applications

Workshops, pages 670–674.

Mirzadeh, S. I., Farajtabar, M., Li, A., Levine, N., Mat-

sukawa, A., and Ghasemzadeh, H. (2020). Improved

knowledge distillation via teacher assistant. Proceed-

ings of the AAAI Conference on Artiﬁcial Intelligence,

34(04):5191–5198.

Morrison, P. and Murphy-Hill, E. (2013). Is programming

knowledge related to age? an exploration of stack

overﬂow. In 2013 10th Working Conference on Min-

ing Software Repositories (MSR), pages 69–72.

Movshovitz-Attias, D., Movshovitz-Attias, Y., Steenkiste,

P., and Faloutsos, C. (2013). Analysis of the reputa-

tion system and user contributions on a question an-

swering website: Stackoverﬂow. In 2013 IEEE/ACM

International Conference on Advances in Social Net-

works Analysis and Mining (ASONAM 2013), pages

886–893.

Overﬂow, S. (2022a). Stack Overﬂow Datadump.

Overﬂow, S. (2022b). What is reputation? How do I earn

(and lose) it? - Help Center - Stack Overﬂow.

Park, S.-T. and Chu, W. (2009). Pairwise preference regres-

sion for cold-start recommendation. In Proceedings of

the third ACM conference on Recommender systems,

pages 21–28.

Pechyony, D. and Vapnik, V. N. (2010). On the theory of

learning with privileged information. In NIPS 2010.

Pennington, J., Socher, R., and Manning, C. (2014). GloVe:

Global vectors for word representation. In Proceed-

Predicting Reputation Score of Users in Stack-overﬂow with Alternate Data

361

ings of the 2014 Conference on Empirical Methods in

Natural Language Processing (EMNLP), pages 1532–

1543, Doha, Qatar. Association for Computational

Linguistics.

Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., and

de Freitas, N. (2016). Taking the human out of the

loop: A review of bayesian optimization. Proceedings

of the IEEE, 104(1):148–175.

Slag, R., de Waard, M., and Bacchelli, A. (2015). One-day

ﬂies on stackoverﬂow - why the vast majority of stack-

overﬂow users only posts once. In 2015 IEEE/ACM

12th Working Conference on Mining Software Repos-

itories, pages 458–461.

Sujithra Alias Kanmani, R., Surendiran, B., and Ibrahim,

S. (2021). Recency augmented hybrid collaborative

movie recommendation system. International Journal

of Information Technology, 13(5):1829–1836.

Vapnik, V. and Izmailov, R. (2015). Learning using priv-

ileged information: similarity control and knowledge

transfer. J. Mach. Learn. Res., 16:2023–2049.

Vapnik, V. and Vashist, A. (2009). 2009 special issue: A

new learning paradigm: Learning using privileged in-

formation. Neural Netw., 22(5–6):544–557.

Wang, J., Ding, K., and Caverlee, J. (2021a). Sequential

recommendation for cold-start users with meta tran-

sitional learning. In Proceedings of the 44th Inter-

national ACM SIGIR Conference on Research and

Development in Information Retrieval, pages 1783–

1787.

Wang, S., Zhang, K., Wu, L., Ma, H., Hong, R., and Wang,

M. (2021b). Privileged graph distillation for cold start

recommendation. Proceedings of the 44th Interna-

tional ACM SIGIR Conference on Research and De-

velopment in Information Retrieval.

Woldemariam, Y. (2020). Assessing users’ reputation from

syntactic and semantic information in community

question answering. In Proceedings of the 12th Lan-

guage Resources and Evaluation Conference, pages

5383–5391, Marseille, France. European Language

Resources Association.

Xu, C., Li, Q., Ge, J., Gao, J., Yang, X., Pei, C., Sun, F., Wu,

J., Sun, H., and Ou, W. (2019). Privileged features

distillation at taobao recommendations.

Zheng, Y., Liu, S., Li, Z., and Wu, S. (2021). Cold-start

sequential recommendation via meta learner. In Pro-

ceedings of the AAAI Conference on Artiﬁcial Intelli-

gence, volume 35, pages 4706–4713.

Zhou, G., Fan, Y., Cui, R., Bian, W., Zhu, X., and Gai, K.

(2018). Rocket launching: A universal and efﬁcient

framework for training well-performing light net. Pro-

ceedings of the AAAI Conference on Artiﬁcial Intelli-

gence, 32(1).

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

362