Recommendation Systems in a Conversational Web

Konstantinos N. Vavliakis

1,2

, Maria Th. Kotouza

, Andreas L. Symeonidis

and Pericles A. Mitkas

Department of Electrical and Computer Engineering, Aristotle University of Thessaloniki, Thessaloniki, Greece

Pharm24.gr, Greece

Keywords:

Personalization, Recommendation, Conversational Web, e-Commerce, RFM, Recurrent Neural Networks.

Abstract:

In this paper we redeﬁne the concept of Conversation Web in the context of hyper-personalization. We argue

that hyper-personalization in the WWW is only possible within a conversational web where websites and users

continuously “discuss” (interact in any way). We present a modular system architecture for the conversational

WWW, given that adapting to various user proﬁles and multivariate websites in terms of size and user trafﬁc

is necessary, especially in e-commerce. Obviously there cannot be a unique ﬁt-to-all algorithm, but numerous

complementary personalization algorithms and techniques are needed. In this context, we propose PRCW,

a novel hybrid approach combining ofﬂine and online recommendations using RFMG, an extension of RFM

modeling. We evaluate our approach against the results of a deep neural network in two datasets coming

from different online retailers. Our evaluation indicates that a) the proposed approach outperforms current

state-of-art methods in small-medium datasets and can improve performance in large datasets when combined

with other methods, b) results can greatly vary in different datasets, depending on size and characteristics,

thus locating the proper method for each dataset can be a rather complex task, and c) ofﬂine algorithms should

be combined with online methods in order to get optimal results since ofﬂine algorithms tend to offer better

performance but online algorithms are necessary for exploiting new users and trends that turn up.

1 INTRODUCTION

Personalization (or customization) systems focus on

tailoring a service or a product to accommodate spe-

ciﬁc individuals, sometimes tied to groups of indi-

viduals. They have become increasingly popular in

recent years and are considered key elements in a

variety of areas including e-commerce, movies, mu-

sic, news, research articles, search queries and social

tags. They are broadly used for improving customer

satisfaction, sales conversion and marketing results.

In speciﬁc, web personalization dynamically serves

the most relevant content, call-to-action elements, and

messaging for stakeholders’ unique interests. The im-

portance of personalization in e-commerce has been

undisputed, already from 1998 in an interview with

the Washington Post; Jeff Bezos made a visionary

statement about the web: “If we have 4.5 million cus-

tomers, we shouldn’t have one store. We should have

4.5 million stores”.

Nowadays it is state of practice for web compa-

nies some sort of personalization in their websites,

based on IP address, browser language, the referral

link and user’s history. The vast majority of them

employ customer proﬁles which sometimes may be

dynamic. This means that personalization targeted to

segments of users rather than individuals requires spe-

ciﬁc steps in order to launch an effective strategy, such

as to identify audience, understand visitors, plan and

create different experience for each audience.

Recently a new term has appeared, “hyper per-

sonalization”, deﬁned as: “the use of data to pro-

vide more personalized and targeted products, ser-

vices and content”. Hyper-Personalization means to

rethink customer interaction on a one-to-one basis,

where we treat each and every customer uniquely and

design a customized experience for each one. The key

element for hyper-personalization is interacting one-

to-one with individuals, not the customer segments

they fall in. To anticipate an individual’s desires at

any point in time, however, requires having deep cus-

tomer insight, which comes from analyzing granular

and big data.

Hyper-personalization is the next era of digital

marketing; emails that change content based on where

a customer is and when the email is opened. Context-

aware messages and segments that are build for more

relevant communications with your customer, push-

ing only those messages he/she should like to receive,

this way targeting to increased revenue. Although of

Vavliakis, K., Kotouza, M., Symeonidis, A. and Mitkas, P.

Recommendation Systems in a Conversational Web.

DOI: 10.5220/0006935300680077

In Proceedings of the 14th International Conference on Web Information Systems and Technologies (WEBIST 2018), pages 68-77

ISBN: 978-989-758-324-7

added value, there are numerous reasons why hyper-

personalization has not yet been adopted by the ma-

jority of websites. Some of these are: a) the over-

abundance of non-actionable data, as most companies

have an abundance of data but cannot use it to person-

alize digital experiences, b) not knowing who to per-

sonalize ﬁrst, as content is locked up in a content man-

agement system and controlled by developers, while

visitor data is not available for targeting in real time,

c) difﬁculties in measuring the impact of personaliza-

tion, as companies often lack a direct way to measure

the aggregate effect of that portfolio of customized

content across their site over time.

Alongside “hyper-personalization” another term,

“conversational web” has recently started to be used

in the context of user interfaces, also known as chat-

bots or virtual assistants, as well as in the context of

web services. Conversation interfaces interact with

users combining chat, voice or any other natural lan-

guage interface with graphical UI elements like but-

tons, images, menus, videos, etc. The new trend

to evolve from NLP (natural language processing)

to NLU (natural language understanding). On the

other hand, conversational web services (CWS) refer

to web services that communicate multiple times with

a client to complete a single task. Conversations pro-

vide a straightforward way to keep track of data be-

tween calls and to ensure that the Web Service always

responds to the correct client.

In this paper we redeﬁne the term “Conversa-

tional Web” in the context of hyper-personalization.

Conversational Web refers to dynamic, multiple and

asynchronous interactions (implicit conversations)

between users and websites. These conversations al-

low both sides to understand each other and com-

municate efﬁciently. We argue that only in a truly

conversational system is hyper-personalization pos-

sible, as in order to be able to create absolutely tar-

geted messages, offers, interfaces, and recommenda-

tions that resonate and connect differently with each

individual, one must ﬁrst listen the needs and wills

of each and every individual. This is only possi-

ble within a conversational web where websites and

users continuously “discuss” (interact). This discus-

sion takes place in the forms of clicks, mouse move-

ment, scrolling, purchases, back or forward move-

ments and time of each page on behalf of customers.

On the other hand, websites “hear” customer’s talk-

ing and respond in the form of personalized product

recommendations, offers, coupons, order appearance

in search, newsletters communications, popups and

push notiﬁcations. Users in turn react to these re-

sponses and a new cycle of communication begins.

In order to produce accurate predictions and rec-

ommendations, big data analysis is necessary for

identifying trends and patterns in data. This analy-

sis can only take place in ofﬂine mode as it is both a

time and resource consuming process. On the other

hand new customers, products and trends continu-

ously emerge, thus achieving hyper-personalization

requires more than just analysis of historical data. The

“discussion” between users and websites should con-

tinuously be analyzed for improving customer expe-

rience, and online analysis should also take place and

complement the results of the ofﬂine processes.

In this context, we propose a modular architec-

ture for conversational websites. We acknowledge

that the conversational web needs to adapt to various

user proﬁles and independent websites with varying

context, size and user trafﬁc, thus there cannot be a

unique ﬁt-to-all algorithm, but numerouscomplemen-

tary personalization algorithms and techniques are re-

quired, as well as a framework to decide when and

where to use each algorithm. For this reason, we pro-

pose PRCW (Product Recommendations for Conver-

sation Web), a novel hybrid approach combing ofﬂine

and online recommendations using RFMG (Recency-

Frequency-Monetary-Gender), an extension of the

well-known RFM method. Through PRCW, mod-

eling and partial matching recommendations can be

combined with existing deep neural networks and

provide improved results. We evaluate the proposed

methodology on two discrete datasets, with different

characteristics to test how the proposed method per-

forms. Then we combine the proposed method with

the deep neural network and we show that this com-

bination leads to improved results.

The remainder of this paper is structured as fol-

lows. Related work on personalization and recom-

mender systems, is discussed in Section 2. Section

3 describes in detail a framework for the Conversa-

tional Web, while Section 4 introduces a novel hy-

brid approach for recommendations, which is evalu-

ated in Section 5. Section 6 summarizes work done,

discusses future work and concludes the paper.

2 RELATED WORK

Web personalization implies tailoring a website to

accommodate speciﬁc individuals or groups of in-

dividuals. Recommender systems are key elements

in almost every personalization system and are di-

vided in online and ofﬂine systems. Ofﬂine recom-

mendation systems (Koren et al., 2009) either con-

sisting of content-based recommendations (Pazzani

and Billsus, 2007) or collaborative ﬁltering (Sarwar

et al., 2001), have weaknesses. They require signif-

Recommendation Systems in a Conversational Web

icant training time; data updates usually require re-

training the whole model and cannot take into ac-

count frequent changes in interests and proﬁle of

users. In more detail, the techniques used in the

ﬁeld of recommendation systems can be categorized

into four general types: content-basedﬁltering (CBF);

collaborative ﬁltering (CF); rule-based approaches;

hybrid approaches (Ranjbar Kermany and Alizadeh,

2017). Collaborative Filtering techniques look for

patterns in the overall user activity to produce rec-

ommendations, and can be further categorized into

Neighborhood-based, Model-based, Clustering and

Association Rules methods. In recent years, with the

rise of big data, deep model-based approaches have

been applied in this ﬁeld with promising results.

Although session-based recommendation was un-

til recently a relatively unappreciated problem, in the

last few years it has attracted interest (Hidasi et al.,

2015). This is because the behavior of users shows

session-based traits, or users often have only one ses-

sion. Recommendation systems widely use factor

models (Koren et al., 2009) or neighborhood methods

(Sarwar et al., 2001). Factor models are hard to apply

in session-based recommendation due to the absence

of user proﬁles, while neighborhood models, such as

item to item similarity, ignore the information of the

past clicks.

Drawbacks of ofﬂine recommender systems have

been acknowledged and for that reason various on-

line recommender methods (Ying et al., 2006) have

been proposed, which need less processing power and

do not require training. Nevertheless, online recom-

menders are less accurate than ofﬂine methods, thus

hybrid approaches (Burke, 2002) have been proposed

that combine the advantages of online and ofﬂine rec-

ommendation methods. Preference elicitation is also

a popular personalization technique. In the context

of preference elicitation, questionnaires, reviewing

pre-selected items, dynamic learning (Rubens et al.,

2011), entropy optimization (Salimans et al., 2012)

and latent factor models (Huang, 2011) have been em-

ployed. Nevertheless, preference elicitation is not al-

ways efﬁcient and it is recommended only in speciﬁc

problems (Zhao et al., 2013). Interactive systems are

another popular group of methods relative to our case.

In interactive systems users play an active role, they

are usually based on reviews (Chen and Pu, 2012),

constrains (Felfernig et al., 2011), and questionnaires

(Mahmoodand Ricci, 2009). A common method used

in interactive systems, is when users are asked to re-

view a predeﬁned selection of items, in order to cope

with the cold-start problem. These requirements may

frustrate users.

Recently deep learning and especially recurrent

neural networks allow sequential data modeling and

have shown remarkable results (Quadrana et al.,

2017). Embedding deep learning techniques into rec-

ommender systems is gaining traction due to its state-

of-the-art performances and high-quality recommen-

dations. Deep learning (Zhang et al., 2017) provides

a better understanding of user’s demands, item’s char-

acteristics, historical interactions and relationships

between them than traditional methods do. Recursive

Neural Networks (RNNs) (Goodfellow et al., 2016)

are a family of neural networks suitable for modeling

variable-length sequential data and can scale to much

longer sequences than other neural networks. Unlike

feed-forward neural networks, RNNs use loops which

allow information to be passed from one state of the

network to the next one. This way former computa-

tions and previous states of the sequence are remem-

bered and taken into account. Variants such as Long

Short Term Memory (LSTM) and Gated Recurrent

Unit (GRU) network are often deployed to overcome

the vanishing gradient problem.

Obviously a lot of progress has been made in per-

sonalization and recommendation systems; however

there is no integrated solution that can semantically

understand user’s intentions and dynamically evolve

based on them. Advancing the state-of-the-artwe pro-

pose the combination of state-of-the art deep learn-

ing techniques, which have shown impressive results,

with online recommenders, that use partial matching.

This combination not only achieves better results, but

is also more adaptable.

3 A FRAMEWORK FOR THE

CONVERSATIONAL WEB

Although from the system point of view, creating a

truly conversational website involves a rather com-

plex multi-step procedure; from a user point of view

an e-commerce website supporting conversation web

technologies is just an ordinary website. The only dif-

ference is that somehow it seems so much easier to

use and ﬁnd products and everything seems simple

and intuitive both in terms of UX elements and prod-

uct search.

3.1 A Use-case Scenario

Consider a customer, Irene that wants to buy the new

brand X1 night face cream. Irene performs a web

search and clicks the ﬁrst result that redirects her to

an e-commerce site she has never visited before. At

this point the implicit conversation between the cus-

tomer and the user has already begun. The web-

WEBIST 2018 - 14th International Conference on Web Information Systems and Technologies

site “listens” that a new customer landed from an or-

ganic source, searching for brand X1 night cream, so

it responds with recommendations about other night

creams that are popular among users coming from this

source type, together with other Brand X1 products,

as the user seems interested to this brand. In addi-

tion, the site recognizes that this is a new user, so it

displays the “subscribe to our newsletter” banner in a

more prominent location. Next, Irene adds the prod-

uct to her basket and then hovers for some time over

a shampoo for oily hair, but ﬁnally clicks on a brand

X2 serum she noticed in a banner of the main page.

These actions alone comprise four discrete messages:

as the user has stated that she is actually a) very inter-

ested in the brand X1 night cream (with intent to buy),

b) she is also interested in general for brand X2 and

c) more speciﬁcally in serums, and d) she may need a

shampoo for oily hair.

The website once again “listens” and responds

with even more personalized results as it quickly

learns the interests of the user, for example it recom-

mends cheaper shampoos for oily hair as the ones dis-

played before are considered premium products and

are probably too expensive. In case Irene clicks on

a cheaper shampoo the website will classify Irene as

a customer interested in mid-level products (at least

until she starts showing interest for premium prod-

ucts). This is a continuous and everlasting process;

the website not only adapts to better serve Irene’s in-

terest but also learns from her behavior and the behav-

ior of other users, aggregating this collective wisdom

into actionable insights for improving the overall e-

commerce UX of the site.

3.2 The Proposed Framework

Next, we propose an integrated framework for creat-

ing a conversational website that consists of four py-

lons: a) behavior analysis; b) user experience analy-

sis; c) big data warehousing and d) personalization.

Figure 1 illustrates the overall proposed architecture,

detailed at a module level, as well as the data ex-

change means between subsystems.

a) The Behavior analysis module is responsible

for dynamically analyzing user behavior. Data from

analytics tools (e.g. Google Analytics, Yandex, etc.)

along with scroll maps and mouse gestures should

be combined in order to effectively recognize differ-

ent patterns and user groups, such as novice or ex-

perienced users, users that are just browsing or in-

tent to buy, and categories, brands or speciﬁc prod-

ucts users are interested in. For this task, classiﬁca-

tion and support vector machines have provided em-

inent results in the past (Sun et al., 2002), while re-

cently deep learning and especially recurrent neural

networks have shown improvedperformance. Seman-

tic analysis is also required, as topic modelling and la-

tent dirichlet allocation are useful for analyzing user’s

interests

b) User experience analysis is necessary in order

to better understand users and be able to adapt to their

needs. User experience is a multifactor parameter,

such as website structure, marketing, trust, interactive

and information elements, colors and ease of use. All

these factors are hard to be deﬁned as they contain

strongly subjective elements, key performance indi-

cators, such as bounce rate, average time on site, con-

version rate, and depth of search can provide accurate

metrics for calculating user experience.

c) Consideringspecial requirements for data ware-

housing is necessary in order to build a conversational

system. Due to the nature of conversation, which

is continuous, lengthy and heterogeneous, data ware-

housing should be able to cope with big data and ex-

tremely low response times in queries that will allow

real time queries, as well as different type of informa-

tion, including product data, user click history, mouse

movements, scroll data, e-commerce data including

buys, add to cart, and favorites, visual elements and

statistics about their use. This information should be

combined in ofﬂine operations where intelligent mod-

els will be trained, as well as in real-time situations for

delivering personalized services and UI/UX. Luckily

there are plenty of open source tools able to handle

this type of data, including Apache Hadoop, Elastic

Server, and MongoDB.

d) Finally, the key module of our proposed frame-

work is the personalization module which is responsi-

ble for dynamically integrating information data and

user actions originating from user experience and

behavior and composing different recommendations,

website user interfaces and content tailored to the in-

dividual needs of every visitor. Time performance is

crucial for this step, as most operations are real-time.

This step also includes feedback and dynamic learn-

ing using user-website conversations, thus it’s a self-

improving process.

4 A HYBRID APPROACH FOR

RECOMMENDATIONS

In this Section we propose a hybrid approach for

product recommendations in e-commerce sites that

we call PRCW (Product Recommendations for the

Conversational Web). As there is not a universally

good solution that can ﬁt all circumstances and solve

any problem in product recommendation, different

Recommendation Systems in a Conversational Web

Figure 1: Overall system architecture in a conversational website (detailed at module level).

approaches have to be used depending on the dataset

and the kind of the target e-commerce site. One of the

main parameters that have to be taken into account

is the size of the e-shop in terms of trafﬁc, number

of orders and available products. For this reason, we

propose a hybrid approach and we apply a deep model

to two different click stream datasets, one originating

from a small-to-medium e-commerce site and another

coming from a large European retailer’s website.

Recommendations need to satisfy two fundamen-

tal principles in a conversational e-commerce site, a)

must be relevant and b) must be provided in real-time.

Thus, hybrid approaches are required, that can pro-

vide recommendation in online-mode and date pro-

cessing for improvingresults in ofﬂine-mode. For this

reason, we introduce a new hybrid approach using of-

ﬂine and online processing that combines a cluster-

ing algorithm with a rule-based method. Clustering is

applied to perform consumer segmentation based on

consuming behavior, using RFMG, a modiﬁed ver-

sion of RFM modeling that combines recency, fre-

quency and monetary with gender, whereas the pro-

posed rule based approach that uses four different par-

tial matching processes focuses on solving the prob-

lem of unknown user history.

4.1 Ofﬂine Phase

The ofﬂine phase consists of data preprocessing,

clustering via RFMG analysis and post-processing

analysis. Figure 2 depicts this phase. Data Prepro-

cessing: Data preprocessing is necessary to make

knowledge discovery easier and more accurate. In

this step, data are processed in order to follow the

desired format, attributes are selected, and auxiliary

operations like outlier detection, normalization and

discretization are performed. Users from whom

information is not of adequate value (e.g. users that

have only one or even no product views) are removed

from the dataset.

RFMG Analysis: RFM (Recency, Frequency,

Monetary) analysis is a marketing model that pro-

vides information about customers’ consumption be-

havior and widely used for customer segmentation

(Birant, 2011). The three variables are computed on

the transaction history and measure howrecently, how

often and how much do the consumers buy. These

three attributes are not only computed on the prod-

uct orders (RO, FO, MO), but on the product views

as well (RV, FV, MV), because of the more exten-

sive amount of information that page views provide.

In our work we extend the traditional RFM model to

the RFMG model, by adding the Gender attribute, as

gender is a major factor for decision making in almost

every e-commerce environment. The six attributes are

deﬁned as follows:

• RO/RV: the number of days passed since the cus-

tomer last viewed/purchased a product. Range

[0,d]

• FO/FV: the number of purchases/product views

made by the customer in the last d days. Range

[0,f]

• MO/MV: the summary of the prices of the prod-

ucts that were ordered/viewed by the customer in

the last d days. Range [0,m]

If an attribute value is higher than the maximum

upper threshold, the maximum allowed value is used

instead. Then, normalization is performed in the

range [0,1].

Clustering: Consumer segmentation is an un-

supervised machine learning process that allows

WEBIST 2018 - 14th International Conference on Web Information Systems and Technologies

Figure 2: Ofﬂine phase of PRCW.

understanding consumer characteristics and grouping

consumers according to their behavior. Thus, it’s an

excellent start for designing personalized solutions

according to each customer group’s needs.

Preference Table Creation: A user’s preference

for an item is obtained by the number of times the user

has viewed or ordered the target product. Product list

extraction: For each consumer segment, a top list of

ordered products is extracted by ﬁnding the most pre-

ferred items of the users that belong to the particular

cluster.

4.2 Online Phase

The proposed online phase of our approach is de-

picted in Figure 3. Prediction by Partial Matching

(PM), or else known as Markov Model, is a method

used to predict the next state of the model taking

into account the n previous states (Gellert and Florea,

2016). This means that the current state depends on

the previous n states. The number of previous states

determines the order of the PM model. Assuming that

is the state at time t, an R-order model is deﬁned as

in Eq 1.

P[q

t−1

, ...,q

] = P[q

t−1

, ..., q

t−R

] (1)

In our case the states are represented by prod-

uct views. When the target user views the prod-

uct q

, partial matching can be used to ﬁnd the pat-

tern < q

t−1

, q

> within the history of all the users.

The products found to follow the matched pattern are

saved and their frequencies are computed in order to

extract the topM products. Needless to say, when

the order of the model, R, increases, the possibilities

to ﬁnd the desired pattern into the history becomes

lower. So, in our situation the second ordered model

is used.

However, because of the limited number of obser-

vations in smaller e-shops, the non-matching pattern

possibility remains high. In order to address this prob-

lem, we introduce two elastic variants of the partial

matching procedure:

• The ﬁrst one is called PM by intervals and looks

for the pattern < q

t−1

, ..., q

> within the history,

with the constriction that the time interval be-

tween the product views q

t−1

and qt is less than

a time period T. In this case, the topM list is com-

puted using the products that were viewed within

the time period T and after the product view q

• The second one is called PM by session and looks

for the pattern < q

t−1

, ..., q

> within the history,

with the constriction that the product views q

t−1

and q

occurred within the same session. The

topM list is computed using the products that were

viewed within the same session and after the prod-

uct view q

Assume that the target user views the sequence

data < i

, i

>. If sessions [Session1-Session5] have

been extracted by the history, the topM recommenda-

tion list using the four algorithms is presented in Table

Session1: <i3><i5>

<i1>

<i2>

Session2: <i4>

<i3>

Session3: <i6><i4><i9><i4>

Session4: <i4>

<i1>

Session5:

<i9>

<i2>

<i1>

<i4>

Recommendation Systems in a Conversational Web

Figure 3: Online phase of PRCW.

Table 1: Example of topM recommendation list using the different PM algorithms.

Method Recall@1Next Recall@AllNext PrecisionR

PM R=1 i2 40% i3 20% i4 20%

PM R=2 i3 20% - -

PM by intervals i3 20% i4 30% -

PM by session i2 40% i3 20% i4 20%

5 EXPERIMENTAL EVALUATION

5.1 Experimental Setup

We evaluate the proposed hybrid recommendation

method on two different datasets. The ﬁrst dataset

originated from Pharm24.gr, a small-medium (in

terms of trafﬁc) retailer in Greece. The available click

stream contained data from a period of 9 months.

Data from the ﬁrst 7 months were used as the training

set whereas data from the last 2 months where used as

the test set. Items with less than 5 views were ﬁltered

out from the training set, as well as sessions with less

than two item views. Sessions with less than one item

view were also removed from the test set, as well as

item views that do not exist in the training set. After

the preprocessing, the training set contained 53,071

sessions of 875,366 events and 9,733 items, whereas

the test set contains 86 sessions of 585 events and 244

items.

The second dataset is the RecSys dataset that

was provided for the RecSys Challenge 2015 (Ben-

Shimon et al., 2015). This dataset contains click-

streams of a big e-commerce site, organized in ses-

sions. The training set contains all but the last 10 days

of the dataset, whereas the test set contains the ses-

sions of the last 10 days. After the same preprocess-

ing phase, the training set contains 7,802,137 sessions

of 30,958,148 events and 37,331 items, while the test

set contains 71,060 sessions of 217,014 events and

10,829 items. The evaluation was performed by pro-

viding the events of each session of the test set one by

one and making recommendations applying the pro-

posed algorithm to the training set (Algorithm 1).

5.2 Evaluation Metrics

Precision and recall, two commonly used metrics in

the ﬁeld of recommender systems were employed for

evaluating the performance of our algorithms. Sup-

pose that U is the set of users that are examined, R(u)

is the set of items recommended to user u, V(u) is the

set of items viewed by user u after the recommenda-

tion and V(u,1) is the ﬁrst product that user u viewed

after the recommendation. We deﬁne PrecisionR (Eq.

2) as the percentage of recommended items viewed by

the user over the number of recommended products

and PrecisionV (Eq. 3) as the percentage of recom-

mended items viewed by the user

PrecisionR =

∑

|R(u) ∩V(u)|

∑

|R(u)|

(2)

WEBIST 2018 - 14th International Conference on Web Information Systems and Technologies

Algorithm 1: Partial pseudo-code for “UserPersonomy”.

Input:

(test set: table Tx3 [user_id, timestamp, item])

Output:

(topM1, topM2, topM3, topP, topN, next_views)

#Preprocessing

1. Filter out products with less than 5 views

2. Filter out sessions with less than 2 products

3. Separate the dataset into training and test set

4. Filter out products from the test set that do not belong to the training set

#Iteration

For

each user_id-u:

For

each timestamp-t:

Find current item i

, previous item i

c-1

and the next items i

c+1

, .., i

Perform PM R2 using i

c-1

: topM1 list is returned

Perform PM by intervals using i

c-1

: topM2 list is returned

Perform PM by session using i

c-1

: topM3 list is returned

user_id[u] belongs to any clusterRFM

then

Get the topP[u,t] list from the corresponding element of clusterRFM

Add a score to topM[u,t] belonging to topP[u,t] and (topM[u,t] & topP[u,t])

End_if

Merge topM1[u,t], topM2[u,t], topM3[u,t] lists

Update (topN[u,t] list, next_views[u,t] list< i

c+1

, .., i

End_For

Return

(topN, next_views)

PrecisionV =

∑

|R(u) ∩V(u)|

∑

|V(u)|

(3)

We deﬁne recall as the percentage of users

that viewed recommended items at next times-

tamps. Three variants of recall are deﬁned: Re-

call@1Next (Eq. 4), the strictest one, which deter-

mines only the ﬁrst next view after recommendation,

Recall@AllNext (Eq. 5), which determines all next

views after recommendation, and Recall@Positive

(Eq. 6), which considers only the cases the recom-

mendation list has at least one item. The evaluation

process is depicted in Algorithm 1.

Recall@1Next =

∑

|R(u) ∩V(u, 1)|

|U|

(4)

Recall@AllNext =

∑

|R(u) ∩V(u)|

|U|

(5)

Recall@Positive =

∑

|R(u) ∩V(u)|

∑

|R(u) 6= 0|

(6)

5.3 Results

Tables 2 and 3 present the results achieved by the pro-

posed algorithm PRCW, the RNN and the combina-

tion of them using the Pharm24.gr and the RecSys

dataset, accordingly. For deep model evaluation we

used a GRU-based RNN model (Hidasi et al., 2015)

for session-based recommendations. The input of the

network was the actual state of the session represented

by a 1-of-N encoding, where N is the number of items

(a vector with 1 to the active items and 0 elsewhere),

and the output was the likelihood for each item to be

part of the next session. Session-parallel mini-batches

and mini-batch based output sampling were used for

the output.

Taking into account Tables 2 and 3 one can ob-

serve that the RNN model could not achieve good

enough results in a smaller and sparse dataset, while

the proposed approach not only demanded consid-

erable less RAM and CPU recourses, but also per-

formed better as PRCW achieved better results than

RRN for the Pharm24 dataset, both in terms of Re-

call and Precision. On the other hand, the RNN has

better performance in the RecSys dataset which con-

tains more data both in terms of quantity and den-

sity. Nevertheless, the combination of both meth-

ods (PRCW+RNN) achieves improved performance

in both datasets.

Comparing the results, one can better understand

the difference between the algorithms and datasets.

Bigger datasets have improved chances to get better

recommendations, due to the larger amount of infor-

Recommendation Systems in a Conversational Web

Table 2: Results of the Pharm24 dataset using the hybrid approach.

Method Recall@1Next Recall@AllNext Prec.R Prec.V Pos.Recall

PRCW 0.2880 0.5247 0.0518 0.1414 0.5247

RNN 0.1993 0.3101 0.0348 0.0936 0.3101

PRCW+ RNN 0.3901 0.6065 0.0734 0.1737 0.6065

Table 3: Results of the RecSys dataset using the hybrid approach.

Method Recall@1Next Recall@AllNext Prec.R Prec.V Pos.Recall

PRCW 0.0868 0.1711 0.0273 0.0229 0.1711

RNN 0.8120 0.8886 0.0998 0.6380 0.8886

PRCW+RNN 0.8366 0.9037 0.1139 0.7069 0.9037

mation that contain, and achieve worse results at the

PrecisionR metric, as there are too many products in

the dataset. On the other hand, smaller datasets have

shorter sessions and achieve worse results at the Pre-

cisionV metric. Deep learning can perform excep-

tionally well, as long as there are enough data and

processing power to feed the neural network. On

the other hand, the proposed method PRCW works

better on smaller datasets. In any case combining

both PRCW and RNN delivers the best results in both

datasets, which leads us to the conclusion that both

methods deliver useful results that should be com-

bined for optimal performance.

6 CONCLUSION

In this paper we redeﬁned the concept of Conversa-

tion Web in the context of hyper-personalization. We

proposed a generic design for conversational web that

may be expanded in terms of hyper-personalization,

such as product recommendation, UI/UX personal-

ization, as well as individual messages and promos

per customer. We argued that in a high-demanding

and versatile environment, such as the WWW there is

not a unique ﬁt-to-all solution, thus various solutions

have to be evaluated and blended in order to provide

relevant results in live environments. For this rea-

son, we proposed a novel hybrid method that extends

the RFM model by introducing the Gender factor,

combined with Partial Matching. The method pro-

vided improved results for small-to-medium datasets.

In addition, we combined the proposed algorithm

with a deep learning method and we showed that

they can work complementary as improved results are

achieved when combined of both methods.

Future work includes working on the decision

module for applying the optimum algorithm based on

the dataset characteristics. In addition, we plan to ex-

plore the possibility of further integrating our hybrid

approach with RNNs. Another issue for future con-

sideration is privacy concerns that may arise and how

to tackle them.

ACKNOWLEDGEMENTS

This work was partially funded by an IKY scholar-

ship funded by the “Strengthening of Post-Academic

Researchers” Act from the resources of the OP “Hu-

man Resources Development, Education and Lifelong

Learning” with Priority Axes 6,8,9 and co-funded by

the European Social Fund ECB and the Greek gov-

ernment.

REFERENCES

Ben-Shimon, D., Tsikinovsky, A., Friedmann, M., Shapira,

B., Rokach, L., and Hoerle, J. (2015). Recsys chal-

lenge 2015 and the yoochoose dataset. In Proceedings

of the 9th ACM Conference on Recommender Systems,

RecSys ’15, pages 357–358, New York, NY, USA.

ACM.

Birant, D. (2011). Data mining using rfm analysis. In Fu-

natsu, K., editor, Knowledge-Oriented Applications in

Data Mining, chapter 6. Rijeka.

Burke, R. (2002). Hybrid recommender systems: Survey

and experiments. User Modeling and User-Adapted

Interaction, 12(4):331–370.

Chen, L. and Pu, P. (2012). Critiquing-based recom-

menders: survey and emerging trends. User Modeling

and User-Adapted Interaction, 22(1):125–150.

Felfernig, A., Friedrich, G., Jannach, D., and Zanker, M.

(2011). Developing Constraint-based Recommenders,

pages 187–215. Springer US, Boston.

Gellert, A. and Florea, A. (2016). Web prefetching through

efﬁcient prediction by partial matching. World Wide

Web, 19(5):921–932.

Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep

Learning. MIT Press.

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

(2015). Session-based recommendations with recur-

rent neural networks. CoRR, abs/1511.06939.

WEBIST 2018 - 14th International Conference on Web Information Systems and Technologies

Huang, S.-L. (2011). Designing utility-based recommender

systems for e-commerce: Evaluation of preference-

elicitation methods. Electron. Commer. Rec. Appl.,

10(4):398–407.

Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factor-

ization techniques for recommender systems. Com-

puter, 42(8):30–37.

Mahmood, T. and Ricci, F. (2009). Improving recom-

mender systems with adaptive conversational strate-

gies. In Proceedings of the 20th ACM Conference

on Hypertext and Hypermedia, HT ’09, pages 73–82,

New York, NY, USA. ACM.

Pazzani, M. J. and Billsus, D. (2007). Content-Based

Recommendation Systems, pages 325–341. Springer

Berlin Heidelberg, Berlin, Heidelberg.

Quadrana, M., Karatzoglou, A., Hidasi, B., and Cremonesi,

P. (2017). Personalizing session-based recommenda-

tions with hierarchical recurrent neural networks. In

Proceedings of the Eleventh ACM Conference on Rec-

ommender Systems, RecSys 2017, Como, Italy, August

27-31, 2017, pages 130–137.

Ranjbar Kermany, N. and Alizadeh, S. H. (2017). A hy-

brid multi-criteria recommender system using ontol-

ogy and neuro-fuzzy techniques. Electron. Commer.

Rec. Appl., 21(C):50–64.

Rubens, N., Kaplan, D., and Sugiyama, M. (2011). Active

Learning in Recommender Systems, pages 735–767.

Springer US, Boston.

Salimans, T., Paquet, U., and Graepel, T. (2012). Collab-

orative learning of preference rankings. In Proceed-

ings of the Sixth ACM Conference on Recommender

Systems, RecSys ’12, pages 261–264, New York, NY,

USA. ACM.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2001).

Item-based collaborative ﬁltering recommendation al-

gorithms. In Proceedings of the 10th International

Conference on World Wide Web, WWW ’01, pages

285–295, New York, NY, USA. ACM.

Sun, A., Lim, E.-P., and Ng, W.-K. (2002). Web classiﬁ-

cation using support vector machine. In Proceedings

of the 4th International Workshop on Web Informa-

tion and Data Management, WIDM ’02, pages 96–99,

New York, NY, USA. ACM.

Ying, Y., Feinberg, F., and Wedel, M. (2006). Leveraging

missing ratings to improve online recommendation

systems. Journal of Marketing Research, 43(3):355–

365.

Zhang, S., Yao, L., and Sun, A. (2017). Deep learning based

recommender system: A survey and new perspectives.

CoRR, abs/1707.07435.

Zhao, X., Zhang, W., and Wang, J. (2013). Interac-

tive collaborative ﬁltering. In Proceedings of the

22Nd ACM International Conference on Information

& Knowledge Management, CIKM ’13, pages 1411–

1420, New York, NY, USA. ACM.

Recommendation Systems in a Conversational Web