Session Similarity based Approach for Alleviating Cold-start Session

Problem in e-Commerce for Top-N Recommendations

Ramazan Esmeli

1 a

, Mohamed Bader-El-Den

1 b

and Hassana Abdullahi

2 c

School of Computing, University of Portsmouth, Lion Terrace, Portsmouth, U.K.

School of Mathematics and Physics, University of Portsmouth, Lion Terrace, Portsmouth, U.K.

Keywords:

Cold-start Sessions, Recommender Systems, Session-based Recommender Systems.

Abstract:

Cold-Start problem is one of the main challenges for the recommender systems. There are many methods

developed for traditional recommender systems to alleviate the drawback of cold-start user and item problems.

However, to the best of our knowledge, in session based recommender systems cold-start session problem still

needs to be investigated. In this paper, we propose a session similarity-based method to alleviate drawback of

cold-start sessions in e-commerce domain, in which there are no interacted items in the sessions that can help

to identify users’ preferences. In the proposed method, product recommendations are given based on the most

similar sessions that are found using session features such as session start time, location, etc. Computational

experiments on two real-world datasets show that when the proposed method applied, there is a signiﬁcant

improvement on the performance of recommender systems in terms of recall and precision metrics comparing

to random recommendations for cold-start sessions.

1 INTRODUCTION

Recommender Systems (RS) are widely used in many

digital platforms including e-commerce, video and

music since RS can help retrieve and recommend

most relevant items from massive item options. Al-

though RS are important, they suffer several setbacks.

For example, in the e-commerce domain, when a

user visits anonymously or visits without browsing

any item, providing product recommendations will

be challenging since there is little or no information

about the user. According to a dataset released by

DIGINETICA

, it is found that 62.3% of the users

are not registered in the system (Wu and Yan, 2017;

Esmeli et al., 2019a). This is known as cold-start

problem (Son, 2016). There are two types of cold-

start problems in RS (Son, 2016; Herce-Zelaya et al.,

2020), namely, cold-start item problem and cold-start

user problem. Cold-start item problem is when a new

item added to a system or database has few or no user-

new item interactions, while in cold-start user prob-

lem, a new user who is not registered to the system

https://orcid.org/0000-0002-2634-6224

https://orcid.org/0000-0002-1123-6823

https://orcid.org/0000-0003-2347-6848

http://diginetica.com/

starts browsing items or is a newly registered user

who has no previous history, making it difﬁcult for

RS algorithms to identify this user’s preferences.

Generally, in cold-start item problems, content-

based approaches are preferred (Lika et al., 2014; An-

waar et al., 2018; Esmeli et al., 2019b). For example,

using items’ content features, similar items are ranked

and recommended. On the other hand, since it is not

easy to give speciﬁc product recommendations for the

cold-start user, random or most popular items are rec-

ommended(Son, 2016; Esmeli et al., 2020). There-

fore, in both cases, the accuracy of the RS can be low.

There are several approaches that have been proposed

to alleviate new user and item problems (Son, 2016;

Herce-Zelaya et al., 2020; Silva et al., 2019). Users’

demographic data (Son, 2016; Silva et al., 2019) is

the most commonly used data where similar users

are found using their demographic information and

items these users interacted are recommended for the

new user. Also, in some works (Safoury and Salah,

2013; Bouadjenek et al., 2016), user demographic la-

bels are matched with product features. For instance,

the match between Movie’s age group and user’s age

group products were recommended. Furthermore,

users’ account are associated with their social media

proﬁles (Carrer-Neto et al., 2012; Rosli et al., 2015;

Bouadjenek et al., 2016). In this case, their social

Esmeli, R., Bader-El-Den, M. and Abdullahi, H.

Session Similarity based Approach for Alleviating Cold-start Session Problem in e-Commerce for Top-N Recommendations.

DOI: 10.5220/0010107001790186

In Proceedings of the 12th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2020) - Volume 1: KDIR, pages 179-186

ISBN: 978-989-758-474-9

179

media accounts can provide useful information about

users preferences such as the pages, movies and books

they like, and these details can be used to create better

recommendations.

In this paper, we propose a method for solving

cold-start user problem for Session based RS(SBRS)

in e-commerce domain where no user prior informa-

tion is available. Our proposed method uses session

features such as the time session started, location de-

tails where users are connected to the website, the

platform they use (mobile or desktop platform), etc.

to ﬁnd similar sessions. From these sessions, using

their last interacted item, we create product recom-

mendations. A ﬁnal recommendation list is created

applying ”majority voting” (Aledo et al., 2017) ap-

proach from recommendation pool created from sim-

ilar sessions. Computational experiments show that

our proposed approach has improved performance of

RS when compared to random item RS. The main

contributions of this paper are as follows:

1. a framework for cold-start session is introduced.

2. majority voting based recommendation creation

process is designed.

3. proposed framework has been validated and com-

pared with random item recommendation using

two real-world datasets.

The remainder of the paper is organised as follows. In

Section 2, we present a literature review of the rele-

vant studies in e-commerce domain for solving cold-

start problems. Section 3 introduces the proposed

framework. Section 4 presents the comparative ex-

periments of the proposed framework involving two

real-world RS datasets. Finally, Section 5 draws con-

clusions and provides future research directions.

2 RELATED WORKS

In this section, we provide the most relevant works

related to cold-start problem in e-commerce domain.

As cold-start problems are the major problems of the

RS, they have gained a lot of attention from the sci-

entiﬁc community. Approaches for solving cold-start

problems can be grouped into three categories based

on the solution approach followed. These are: (i)

content-based approaches such as demographic infor-

mation or individual labels for items, (ii) approaches

that extract and use user’s external data (for example,

information from social media) and (iii) approaches

that use initial information provided by users.

Regarding content-based approach (Safoury and

Salah, 2013) designed a method that takes users’ de-

mographic information and movies’ attributes into

consideration to create a recommendation list based

on the similarity of these features. They aimed to re-

duce the impact of cold-start movie problem that have

little user interactions. Also, in (Lika et al., 2014), au-

thors used classiﬁcation algorithms using users’ de-

mographic information in order to identify similar

users. Their proposed classiﬁcation algorithms were

integrated into pure Collaborative Filtering (CF) algo-

rithms.

Moreover, in another work (Ralph et al., 2020)

user and item descriptions are mined to ﬁnd relation-

ships between items and users. Relationships were

created by text mining (item, user descriptions) and

vectors for each document were created. These vec-

tors were used to ﬁnd similar products for the new

user based on cosine similarity. Clustering is an-

other method applied for alleviating cold-start user

problem in CF (Pereira and Hruschka, 2015; Zhang

et al., 2013). Zhang et al. (2013) applied bi-clustering

that combines user and item features to associate like

minded users with items. For the association, they

predicted ratings for unrated items from like-minded

user clusters and similar item clusters. Pereira and Hr-

uschka (2015) designed a framework that combines

simultaneous co-clustering learning with CF in order

to alleviate new user problem. In their method, users’

demographic information was used for clustering pur-

pose.

Users’ external data such as social media data is

used to reduce the effect of cold-start problems (Sa-

hebi and Cohen, 2011; Castillejo et al., 2012; Tian and

Liang, 2017; Sedhain et al., 2014; Guy et al., 2010).

For example, Guy et al. (2010) trained a model using

Matrix Factorisation (MF) that learn latent factors for

social tags. These factors were used to ﬁnd relevant

items for new users. Moreover, in (Tian and Liang,

2017) trust relationship in the social media is utilised,

in which users’ friends were categorised as trusted

users and based on their preferences, new recommen-

dations were created for the new user. Similarly, He

and Chu (2010) presented a method that takes a user’s

social media preferences and inﬂuence from friends

into account for alleviating cold-start user problem.

Lastly, using randomly collected user preferences

for new items is another approach for reducing the

effect of cold-start problem. An example of this ap-

proach can be seen in (Aharon et al., 2015). The au-

thors designed an online method that randomly asks

a group of user for their preferences for newly added

items. Based on the collected preferences, new items

are recommended to other like-minded users. Their

method showed a signiﬁcant performance improve-

ment in terms of Root Mean Square Error (RMSE)

evaluation metric.

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

180

2.1 Limitation of Previous Works and

Proposed Contributions

Previous works showed improvement on alleviating

cold-start problem when users and items have fea-

tures (tags, preferences, social media). However,

in e-commerce domain users can visit the website

anonymously or without any data description. Con-

sequently, SBRS suffers while providing product rec-

ommendations. In order to decrease the effect of cold-

start session problem in the e-commerce domain, we

propose a framework that can ﬁnd most similar ses-

sions based on the session attributes (user location,

time user visited, device type, etc.) which does not

require any user input.

The framework provides recommendations for the

most similar sessions, and using majority voting strat-

egy, ﬁnal recommendation list is created for the cold-

start session.

3 FRAMEWORK

In order to get more accurate product recommenda-

tions, a framework that utilises session features is de-

signed (Figure 1). Since in cold-start sessions, there

are no item interactions, identifying users’ exact in-

tention in the session is very challenging. However,

product recommendations can be estimated from ses-

sions that have similar features with cold-start ses-

sion. This framework helps to ﬁnd the most similar

sessions and create a product recommendation pool

using the item interactions of these sessions. Based on

the majority voting method ﬁnal recommendation list

is created. The reason for choosing majority voting

is to choose top n n ∈ {10, 20, 30}) high conﬁdence

(high frequency) recommended items from similar

sessions.

The designed framework has three main steps.

These are;

1. Data preparation, which includes session log col-

lection and pre-processing. In pre-processing, the

sessions have less than two item interactions are

eliminated in order to have a meaningful RS mod-

els in the sense of creating item relation.

2. Second phase is designed for training the nearest

neighbour model that helps to ﬁnd the most 20

similar sessions from train dataset when a set of

features is provided as input. Test dataset con-

sists of session logs created in the last week of

the whole duration of session logs. Also, in test

dataset interacted items are hidden where these

items are used for validation purpose. Also, the

features of each test session are used to ﬁnd the

most similar sessions in train dataset.

3. Final phase of the framework is necessary for cre-

ating recommendation list using majority voting

method, in which top n (n ∈ {10, 20, 30}) high-

est number of appeared items in the recommen-

dations from each session are selected. In this

phase recommendations from similar sessions are

derived using different RS models. These are

Item-Item KNN, Random and Session Popular-

ity based recommendations. We use same pa-

rameter settings for these models, as explained

in (Hidasi et al., 2015). In these models, next

item recommendations are given based on the

last clicked item. After collecting recommenda-

tions from similar sessions and selecting most fre-

quented items, a ﬁnal recommendation list is cre-

ated. Then the ﬁnal created recommendation list

is evaluated with ground truth items in the test ses-

sion. Evaluation metrics used in this paper are re-

call and precision, which are very common in the

RS domain (Silva et al., 2019).

The details of the second and third phases of the

framework are presented in following subsections.

3.1 Training Nearest Neighbour Model

This step consists of feature selection to train the near-

est neighbour model and to ﬁnd top n

nearest ses-

sions.

For the model training, we set n = 100 as neigh-

bours, and other parameters have remained as default

for the nearest neighbour algorithm as shown in

Selected features from session speciﬁcations are:

1. day of the week: This feature indicates which day

of the week the session has started.

2. hour of the day: Show the hour of the day the

session has started.

3. day of the year: This feature represents which day

of the year session started. This is important since

users in a speciﬁc season can look for a particular

group of items.

4. device type(mobile, desktop): Shows what kind

of platform user uses in order to visit the e-

commerce platform.

5. operator system: Shows the operating system the

device used to connect to e-commerce platform

uses such as Windows 10, IOS, Android, etc.

6. longitude: Shows longitude information of the lo-

cation that the user is connected from.

https://scikit-learn.org/stable/modules/neighbors.html

Session Similarity based Approach for Alleviating Cold-start Session Problem in e-Commerce for Top-N Recommendations

181

Figure 1: Framework for alleviating cold-start session problem.

7. latitude: Indicates latitude information of the lo-

cation that the user is connected from.

Except for longitude, latitude, day of the year and

hour of the day, other features are categorical. There-

fore in order to represent these features, one-hot en-

coding is applied. After applying one-hot encoding,

we have 34 features to represent each session. As

Algorithm 1: Hybrid approach for cold-start user session

problem.

dataset for training D.

dataset for testing T .

= TrainModel(D) // Item-Item KNN and

Session Pop based models

NN = TrainNNModel(D) // Nearest Neighbour

model to ﬁnd similar sessions

for each n

in N cross validation do

t = Selected Test Sessions // select 1000

sessions in each cross-validation iteration

for each s

in t do

// ground truth items of s

S = NN(s

) // get N

∈ {10, 20, 30, 40}

similar sessions S based on feature of s

r = {}

for each s in S do

= RS

(s)

r = r ∪ r

// Collect recommended items

from each similar session s based on the

last clicked item of s

end for

r = Rank(r) // Rank recommendation List

based on Majority voting

f inal

// create ﬁnal recommendation list by

top N ∈ {10, 20, 30} highest voted items.

Evaluate(r

f inal

, GT

)

evaluate using recall and precision metrics.

end for

seen in Algorithm 1, after creating nearest neighbour

model, for each session we ﬁnd n

∈ {10, 20, 30, 40}

most similar sessions S. These sessions are used as

seed sessions in order to retrieve product recommen-

dations r

with the last interacted item of each session.

This step is explained in detail in the next subsection.

3.2 Training Recommender System

Models and Creating Final

Recommendation List

In this paper, three different RS models are imple-

mented. These are Item-Item KNN, Session-Based

Popularity and Random RS models. In order to design

these models, we apply a similar approach to (Hidasi

et al., 2015). Also, the parameters for the models are

kept the same as shown in (Hidasi et al., 2015). We

give details of the parameters used in each RS model

as follows;

1. Random Based RS: This recommender model

does not get any special parameters.

2. Session Popularity Based RS: Session Popularity

predictor that gives higher scores to items with

higher number of occurrences in the session. Ties

are broken up by adding the popularity score of

the itemHidasi et al. (2015). We set top

= 100

for items in each session with top 100 highest

score.

3. Item-Item KNN based RS: This recommender

model gives prediction scores for a selected set of

items on how likely they are the next items in the

session. Parameters for this recommender model

are: n

sims

= 100 which indicates the number of

recommended items that only give back non-zero

scores to the n

sims

most similar items. lmbd = 20

is a regularisation parameter that exempts simi-

lar items with incidental co-occurrences. al pha =

0.5 is the cosine normalisation value that balances

the similarity between two items.

Train dataset is used in order to create RS mod-

els. For each session in the test dataset, most sim-

ilar sessions are found. One hundred recommended

items are retrieved and added to the recommendation

list pool by using last interacted item in each simi-

lar session. Final recommendation is created using

majority voting method, in which highest frequent

N(N ∈ (10, 20, 30)) items are selected as ﬁnal recom-

mendation list.

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

182

4 EXPERIMENTS AND RESULTS

We conduct a set of experiments to validate the per-

formance of our proposed framework on cold-start

sessions. For evaluation and comparison of our

method, we use two performance metrics based on

Item-Item KNN, and Random and Session Popularity

based models.

4.1 Dataset Speciﬁcations

In this paper, we use two real world datasets collected

from e-commerce platforms by a UK based personal-

isation company. For pre-processing, sessions which

have less than two item interactions are eliminated

to have better correlation between items. Table 1

shows the number of sessions, items and interactions

of datasets before and after pre-processing.

Table 1: Statistics about datasets.

Dataset #session #item #interaction

ds 1 2490591 10811 4892053

ds 2 1190897 2316 2211056

After Pre-processing

ds 1 461627 9605 2479139

ds 2 186239 2260 1052636

Since we use session features (time, day of the

week, operating system, etc.) in order to create ses-

sion similarities, we provide statistics about these fea-

tures.

Figure 2 shows the number of the user interac-

tions with e-commerce platform by the day of week

for both dataset. Monday is the highest number of

user interactions for ds 1, while Thursday and Satur-

day are the highest number of user interactions days

for the ds 2. Moreover, Figure 3 presents statistical

analysis of the number of user interactions by used

platform by users. It can be seen that both datasets

have similar trend for the platforms.

Figure 2: User interaction statistics by day of week.

Figure 3: User interaction statistics by platform.

Figure 4 shows the number of user interactions by

operating systems used. Similar to Figure 3, Figure 4

also shows similar trends for number of user interac-

tions by used operating systems.

Figure 4: User interaction statistics by platform.

4.2 Evaluation Metrics

To assess the performance of the proposed frame-

work, precision and recall metrics are used. In RS

domain, recall@n is the ratio between the length of

correct predicted items within top n recommenda-

tions and the length of test items (ground truth) and

precision@n is the ratio between the length of cor-

rect predicted items within top n recommendations

and the length of total recommended items. They are

computed as follows:

recall@n =

|relevant recommendations|

|relevant items|

(1)

precision@n =

|relevant recommendations|

|total recommendations|

(2)

After having ﬁnal ranked recommended items,

we evaluate the proposed models with above men-

tioned metrics for top n recommendations where n ∈

{10, 20, 30} since users are more interested in the rec-

ommendations in top of the recommendation list.

Session Similarity based Approach for Alleviating Cold-start Session Problem in e-Commerce for Top-N Recommendations

183

Table 2: Performance comparison of RS models on proposed framework with different N

parameter.

Models p@10 r@10 p@20 r@20 p@30 r@30 N s

ds 1

KNN 0.028 0.278 0.018 0.363 0.014 0.418

Session Pop 0.005 0.045 0.003 0.068 0.003 0.090

Random RS 0.000 0.000 0.000 0.000 0.000 0.001

KNN cold-start 0.001 0.007 0.001 0.013 0.001 0.021

n=10

Session Pop cold-start 0.001 0.013 0.001 0.022 0.001 0.030

KNN cold-start 0.002 0.021 0.002 0.033 0.001 0.038

n=20

Session Pop cold-start 0.001 0.007 0.001 0.016 0.001 0.037

KNN cold-start 0.002 0.021 0.002 0.041 0.002 0.062

n=30

Session Pop cold-start 0.001 0.011 0.001 0.018 0.001 0.027

KNN cold-start 0.002 0.023 0.002 0.036 0.002 0.05

n=40

Session Pop cold-start 0.001 0.011 0.001 0.019 0.001 0.026

ds 2

KNN 0.038 0.382 0.025 0.491 0.019 0.558

Session Pop 0.007 0.073 0.010 0.202 0.008 0.236

Random RS 0.001 0.006 0.000 0.009 0.000 0.013

KNN cold-start 0.004 0.044 0.004 0.082 0.004 0.116

n=10

Session Pop cold-start 0.003 0.026 0.002 0.041 0.003 0.075

KNN cold-start 0.007 0.067 0.006 0.109 0.005 0.147

n=20

Session Pop cold-start 0.004 0.044 0.004 0.071 0.004 0.112

KNN cold-start 0.006 0.064 0.006 0.120 0.005 0.164

n=30

Session Pop cold-start 0.002 0.023 0.002 0.041 0.002 0.061

KNN cold-start 0.007 0.069 0.006 0.125 0.005 0.164

n=40

Session Pop cold-start 0.002 0.018 0.002 0.041 0.002 0.062

4.3 Experiments

The main aim of the experiments is to measure the

accuracy of the recommendations of the proposed

framework for cold-start sessions comparing to ran-

dom item recommendations. The fundamental idea

of our approach is that similar sessions which have

similar features can share similar user preferences.

Nearest neighbour and RS models are created

using train dataset. Moreover, for each n

∈

{10, 20, 30, 40} value, a new nearest neighbour model

is created. For each created nearest neighbour model,

ﬁnal recommendation list is created using two differ-

ent RS model.

For each similar session s ∈ s

, s

...s

, we derive

100 product recommendations from each RS model

and create a ﬁnal recommendation list based on ma-

jority voting method. We illustrate how majority vot-

ing works in small example; lets say recommended

items from session s

are {i

, i

100

, i

}, and rec-

ommendations from s

are {i

, i

100

, i

}. So, items

100

} are the same recommended times found in

both sessions {s

, s

...s

}. Therefore, items i

and

100

are selected and added to the ﬁnal recommen-

dation list. The ﬁnal recommendation list is ranked

based on the the number of vote each item gets. For

instance, if item i

appeared 10 times in the recom-

mendations while i

appeared 6 times, then i

is listed

before i

. After the recommendation list is ranked

based on the votes, we evaluate the RS accuracy by

selecting top n items from the ranked list using recall

and precision metrics.

4.4 Results

Table 2 shows the results of the proposed framework

applied to two different RS models. We report RS

models’ performance on top n (n ∈ {10, 20, 30}) rec-

ommendations using recall@n (r), precision@n (p)

metrics. N

shows the selected number of similar ses-

sions for each test session, which are derived from

the nearest neighbour model based on test session fea-

tures. In our experiments, we use random RS model

as the baseline to evaluate the performance of our pro-

posed framework. Random RS models are commonly

used in the case where there is a lack of prior infor-

mation (e.g. item interactions, user preferences, asso-

ciated social media account, etc.) about the user (Ne-

gre et al., 2013; Rohani et al., 2014; Castillejo et al.,

2012).

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

184

Experiments results show that regardless of the N

parameter, when the proposed framework applied, RS

models show better performance than Random RS for

both datasets. It can be seen from the results that

when RS does not have any information about the

new session (cold-start session), the performance of

RS is low. However, where a session has interacted

items, the performance of the RS models improves

signiﬁcantly. For example, when sessions have inter-

acted items, experiment results show that Item-Item

KNN based RS can show around 41% and 55 % re-

call scores in ds 1 and ds 2 respectively.

On the other hand, when the proposed framework

is applied to two well-known RS models, it can be

seen that the recall score can increase signiﬁcantly in

comparison with random RS model in cold-start ses-

sions. In the results table, we also show the impact of

parameter on RS performance. The results indicate

that increasing N

parameter after a certain point does

not help improve the results. We found that N

= 30

is the best value for achieving the best RS accuracy

score. Moreover, in each N

parameter and dataset,

the experiment results showed that Item-Item KNN

based RS is better than Session Popularity based RS

model.

4.5 Conclusion

Cold-start problems are major problems of the RS. In

this paper, we introduced a novel framework that can

deal with the cold-start problem for the new sessions

that have no prior user information such as item in-

teractions. The proposed framework can ﬁnd similar

sessions based on session features such as time ses-

sion started, users’ location, users’ platform, etc. and

create recommendations from these sessions. Com-

putational experiments using two datasets proved that

the proposed framework signiﬁcantly improved rec-

ommendation accuracy and alleviate cold-start user

session problem. We applied the proposed framework

on Item-Item KNN and Session Popularity based RS

models. Nevertheless, the proposed framework is a

general framework that can be applied to different

kinds of RS model, including Deep-Learning based

models (Hidasi et al., 2015; Gabriel De Souza et al.,

2019; Ludewig et al., 2019).

For future work, other session-based datasets

could be tested on the framework to substantiate the

robustness of the proposed framework. Also, clus-

tering approach (Zhang et al., 2013; Yanxiang et al.,

2013; Bader-El-Den et al., 2018) can be applied to

ﬁnd similar sessions. For example, in the ﬁrst step,

sessions can be clustered into groups, and based on

new session features the cluster group that session

belongs to can be identiﬁed then most similar ses-

sions from that cluster can be used in order to create

a recommendation list. Also, other available session

features can be included for nearest neighbour model

creation to ﬁnd similar sessions that can represent new

sessions more accurately.

REFERENCES

Aharon, M., Anava, O., Avigdor-Elgrabli, N., Drachsler-

Cohen, D., Golan, S., and Somekh, O. (2015). Ex-

cuseme: Asking users to help in item cold-start rec-

ommendations. In Proceedings of the 9th ACM Con-

ference on Recommender Systems, pages 83–90.

Aledo, J. A., G

amez, J. A., and Molina, D. (2017). Tack-

ling the supervised label ranking problem by bagging

weak learners. Information Fusion, 35:38–50.

Anwaar, F., Iltaf, N., Afzal, H., and Nawaz, R. (2018). Hrs-

ce: A hybrid framework to integrate content embed-

dings in recommender systems for cold start items.

Journal of computational science, 29:9–18.

Bader-El-Den, M., Teitei, E., and Perry, T. (2018). Biased

random forest for dealing with the class imbalance

problem. IEEE transactions on neural networks and

learning systems, 30(7):2163–2172.

Bouadjenek, M. R., Hacid, H., and Bouzeghoub, M. (2016).

Social networks and information retrieval, how are

they converging? a survey, a taxonomy and an analy-

sis of social information retrieval approaches and plat-

forms. Information Systems, 56:1–18.

Carrer-Neto, W., Hern

andez-Alcaraz, M. L., Valencia-

Garc

ıa, R., and Garc

ıa-S

anchez, F. (2012). Social

knowledge-based recommender system. application to

the movies domain. Expert Systems with applications,

39(12):10990–11000.

Castillejo, E., Almeida, A., and L

opez-De-Ipi

na, D. (2012).

Alleviating cold-user start problem with users’ social

network data in recommendation systems. In Work-

shop on Preference Learning: Problems and Applica-

tions in AI (PL-12) at ECAI, pages 28–33.

Esmeli, R., Bader-El-Den, M., and Abdullahi, H. (2019a).

Improving session based recommendation by diver-

sity awareness. In UK Workshop on Computational

Intelligence, pages 319–330. Springer.

Esmeli, R., Bader-El-Den, M., Abdullahi, H., and Hender-

son, D. (2020). Improving session-based recommen-

dation adopting linear regression based re-ranking. In

IEEE World Congress on Computational Intelligence

(WCCI) 2020. Institute of Electrical and Electronics

Engineers.

Esmeli, R., Bader-El-Den, M., and Mohasseb, A. (2019b).

Context and short term user intention aware hybrid

session based recommendation system. In 2019 IEEE

International Symposium on INnovations in Intelli-

gent SysTems and Applications (INISTA), pages 1–6.

IEEE.

Gabriel De Souza, P. M., Jannach, D., and Da Cunha, A. M.

(2019). Contextual hybrid session-based news recom-

Session Similarity based Approach for Alleviating Cold-start Session Problem in e-Commerce for Top-N Recommendations

185

mendation with recurrent neural networks. IEEE Ac-

cess, 7:169185–169203.

Guy, I., Zwerdling, N., Ronen, I., Carmel, D., and Uziel, E.

(2010). Social media recommendation based on peo-

ple and tags. In Proceedings of the 33rd international

ACM SIGIR conference on Research and development

in information retrieval, pages 194–201.

He, J. and Chu, W. W. (2010). A social network-based rec-

ommender system (snrs). In Data mining for social

network data, pages 47–74. Springer.

Herce-Zelaya, J., Porcel, C., Bernab

e-Moreno, J., Tejeda-

Lorente, A., and Herrera-Viedma, E. (2020). New

technique to alleviate the cold start problem in recom-

mender systems using information from social media

and random decision forests. Information Sciences.

Hidasi, B., Karatzoglou, A., Baltrunas, L., and Tikk,

D. (2015). Session-based recommendations

with recurrent neural networks. arXiv preprint

arXiv:1511.06939.

Lika, B., Kolomvatsos, K., and Hadjiefthymiades, S.

(2014). Facing the cold start problem in recom-

mender systems. Expert Systems with Applications,

41(4):2065–2073.

Ludewig, M., Mauro, N., Latiﬁ, S., and Jannach, D. (2019).

Performance comparison of neural and non-neural ap-

proaches to session-based recommendation. In Pro-

ceedings of the 13th ACM Conference on Recom-

mender Systems, pages 462–466.

Negre, E., Ravat, F., Teste, O., and Tournier, R. (2013).

Cold-start recommender system problem within a

multidimensional data warehouse. In IEEE 7th Inter-

national Conference on Research Challenges in Infor-

mation Science (RCIS), pages 1–8. IEEE.

Pereira, A. L. V. and Hruschka, E. R. (2015). Simultane-

ous co-clustering and learning to address the cold start

problem in recommender systems. Knowledge-Based

Systems, 82:11–19.

Ralph, D., Li, Y., Wills, G., and Green, N. G. (2020). Rec-

ommendations from cold starts in big data. Comput-

ing, pages 1–22.

Rohani, V. A., Kasirun, Z. M., Kumar, S., and Shamshir-

band, S. (2014). An effective recommender algorithm

for cold-start problem in academic social networks.

Mathematical Problems in Engineering, 2014.

Rosli, A. N., You, T., Ha, I., Chung, K.-Y., and Jo, G.-S.

(2015). Alleviating the cold-start problem by incor-

porating movies facebook pages. Cluster Computing,

18(1):187–197.

Safoury, L. and Salah, A. (2013). Exploiting user demo-

graphic attributes for solving cold-start problem in

recommender system. Lecture Notes on Software En-

gineering, 1(3):303–307.

Sahebi, S. and Cohen, W. W. (2011). Community-based

recommendations: a solution to the cold start prob-

lem. In Workshop on recommender systems and the

social web, RSWEB, page 60.

Sedhain, S., Sanner, S., Braziunas, D., Xie, L., and Chris-

tensen, J. (2014). Social collaborative ﬁltering for

cold-start recommendations. In Proceedings of the

8th ACM Conference on Recommender systems, pages

345–348.

Silva, N., Carvalho, D., Pereira, A. C., Mourao, F., and

Rocha, L. (2019). The pure cold-start problem: A

deep study about how to conquer ﬁrst-time users

in recommendations domains. Information Systems,

80:1–12.

Son, L. H. (2016). Dealing with the new user cold-start

problem in recommender systems: A comparative re-

view. Information Systems, 58:87–104.

Tian, H. and Liang, P. (2017). Personalized service rec-

ommendation based on trust relationship. Scientiﬁc

Programming, 2017.

Wu, C. and Yan, M. (2017). Session-aware information em-

bedding for e-commerce product recommendation. In

Proceedings of the 2017 ACM on conference on in-

formation and knowledge management, pages 2379–

2382. ACM.

Yanxiang, L., Deke, G., Fei, C., and Honghui, C. (2013).

User-based clustering with top-n recommendation on

cold-start problem. In 2013 third international con-

ference on intelligent system design and engineering

applications, pages 1585–1589. IEEE.

Zhang, D., Hsu, C.-H., Chen, M., Chen, Q., Xiong, N.,

and Lloret, J. (2013). Cold-start recommendation us-

ing bi-clustering and fusion for large-scale social rec-

ommender systems. IEEE Transactions on Emerging

Topics in Computing, 2(2):239–250.

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

186