Recommendation Recovery with Adaptive Filter for Recommender

Systems

Jos

e Miguel Blanco

1 a

, Mouzhi Ge

and Tom

s Pitner

Faculty of Informatics, Masaryk University, Brno, Czech Republic

Deggendorf Institute of Technology, Germany

Keywords:

Recommender Systems, Recommendation Recovery, Adaptive Filter, User-oriented Recommendation.

Abstract:

Most recommender systems are focused on suggesting the optimal recommendations rather than ﬁnding a way

to recover from a failed recommendation. Thus, when a failed recommendation appears several times, users

may abandon to use a recommender system by considering that the system does not take her preference into

account. One of the reasons is that when a user does not like a recommendation, this preference cannot be

instantly captured by the recommender learning model, since the learning model cannot be constantly updated.

Although this can be to some extent alleviated by critique-based algorithms, ﬁne tuning the preference is not

capable of fully expelling not-preferred items. This paper is therefore to propose a recommender recovery

solution with an adaptive ﬁlter to deal with the failed recommendations while keeping the user engagement

and, in turn, allow the recommender system to become a long-term application. It can also avoid the cost of

constantly updating the recommender learning model.

1 INTRODUCTION

Recommender Systems (RS) play an important role

in our life, from recommending the next favourite

movie or song to suggesting a particular ﬁtness ex-

ercise (Chedrawy and Abidi, 2009; Lavanya et al.,

2021; Sanchez et al., 2020). But their use has ex-

panded beyond those common uses to reach the point

on which their use has been linked to more difﬁcult

questions as where to travel (Mahmood et al., 2008).

Their implementation has gotten to the point where

they aim to have a real conversation with the user

(Jannach et al., 2020).

Most have assumed that ensuring a higher satis-

faction from the user is achieved by making the best

possible prediction (Fang et al., 2020). However, once

RS fail to recommend the optimal item, they may ne-

glect considering the reasons for the unsuccessful rec-

ommendation. Thus, this may lead to a dichotomy in

which a set of users is left behind. To alleviate this is-

sue, explanations can be used to improve RS (Naiseh

et al., 2020) but it does not address the failure itself.

Also, it can be partly attributed to the reason that RS

are considered to be used as one-shot services rather

than long-term services (Mimoun et al., 2017). There-

https://orcid.org/0000-0001-9460-8540

fore, for RS, it is not only important to predict the

best possible recommendation, but also is critical to

recover from the failed recommendations.

Therefore, catching the user preference has be-

come a hot topic in the ﬁeld. While analyzing the

reasons for failure (Ben Mimoun et al., 2012) might

help to understand the next step to be taken, there are

other approaches focused on obtaining direct results.

For example (Narducci et al., 2018) has focused on

using conversational RS to improve the user expe-

rience. Similarly, (Kang et al., 2017) explains how

the users utilize natural language to look for a recom-

mendation, thus being able to ﬂag certain behaviours

that point directly to catching their preference. It is

worth noting, that to catch user preference, a point

in which their personalities are modelled has been

reached (Alves et al., 2020).

In order to tackle the recommendation recovery,

the main aim of this paper is to propose a solution for

recovering failed recommendations, i. e., when RS

recommend a suboptimal item, it can be used to ad-

dress the user un-satisfaction for recommendations.

Further, we aim to make the RS sustainable for a

longer period of time without increasing the compu-

tational complexity of it or having to deal with a cold

start again (Han et al., 2019). This solution is driven

Blanco, J., Ge, M. and Pitner, T.

Recommendation Recovery with Adaptive Filter for Recommender Systems.

DOI: 10.5220/0010653600003058

In Proceedings of the 17th International Conference on Web Information Systems and Technologies (WEBIST 2021), pages 283-290

ISBN: 978-989-758-536-4; ISSN: 2184-3252

283

by an adaptive ﬁlter, which after a recommendation

failure, ﬁlters out all items that are similar to the one

disliked by the user. We also validate the solution in

the healthcare RS domain.

The structure of the paper is organized as follows:

Section 2 reviews the state-of-the-art works that are

related to recommendation reﬁnement and recovery.

Based on the review, Section 3 introduces a series of

deﬁnitions and technical notions that help to under-

stand the adaptive ﬁltering technique that we aim to

implement. We also include the process model of how

the proposed adaptive ﬁlter should work with a visual

representation of the solution. Section 4 is dedicated

to the preliminary validation in which we show how a

RS is prepared to fail and discard suboptimal choices

in the concrete case of the use of medical drugs in

the Emergency Room. In Section 5 we discuss the

proposed solution by comparing with critique-based

RS and highlighting the novelty of the work. Finally,

Section 6 concludes the paper and outlines future re-

search work.

2 RELATED WORKS

There are several research works about RS that try

to improve the user experience. Most of those works

are to ﬁnd a characteristic related to the recommenda-

tion and use it as the main pillar to improve recom-

mendations. Those recommenders can be classiﬁed

as characteristic-based ﬁne tuning, boost factor im-

provement, and interaction-based recovery.

Previous papers such as (Calero Valdez et al.,

2016), (Sidana et al., 2021), (Burke, ), (Ziegler, 2005)

and (Codina and Ceccaroni, 2010) show that there

are different valid technical approaches when devel-

oping a RS. This is related to how the novelty of ex-

cluding the items that our approach proposes is worth

developing. (Ziegler, 2005) and (Codina and Cecca-

roni, 2010) are focused on how RS works within the

framework that the semantic web provides. While the

ﬁrst is a compendium of all the techniques that can

be applied, the latter is a direct application of given

techniques. Also, (Sidana et al., 2021) provides a

new framework for collaborative ﬁltering RS for a

smaller user-loss despite a minimal choosing of ele-

ments. The work is further expanded into a Neural-

Network model to support the analysis of the data.

Another type of framework is to ﬁne tune the recom-

mendation. For example, (Calero Valdez et al., 2016)

shows the use of RS in Health Informatics. They fo-

cus on explaining how the use of a doctor-in-the-loop

ﬁgure would help the systems to ﬁne tune the system

and provide a more clear recommendation. They also

propose a framework for evaluating this kind of RS.

To mix the frameworks and achieve high quality rec-

ommendation, (Burke, ) shows how hybrid web RS,

those that use multiple approaches instead of just one

when generating a recommendation, are better suited

than those that are not. This leads to a higher satisfac-

tion of the user.

The works from (De Pessemier et al., 2010) and

(Mendoza and Torres, 2020) show how the best possi-

ble recommendation is generated based on the result

of certain boost factor in the RS. In (De Pessemier

et al., 2010), the authors focus on showing how time

affects the quality of data collected in collaborative

algorithms to provide the end user with the best rec-

ommendation possible. On the other hand, (Mendoza

and Torres, 2020) introduces a framework on which

the bias for popular items is alleviated by introduc-

ing an evaluation tool on new items regarding their

novelty with respect to the characteristics of the most

popular items.

Finally, most of the work on RS is based on how to

process the data generated by the user proﬁles. Papers

are focused on obtaining the best recommendation a

priori, even though there are exceptions as (Li et al.,

2018). These papers show the importance of listen-

ing to the choice that the user is making and being

able to respond afterwards. For example, (Adomavi-

cius et al., 2011) goes in detail into Context Aware

RS that make use of the context generated data for the

user. The position that the authors hold is that includ-

ing all the context data the recommendation will be

more tailored for a speciﬁc user. (Li et al., 2018) de-

velops a method focused on group recommendation

and interactive preference. They include a mecha-

nism to generate feedback from a post-rating system.

They evaluate the work done by comparing with tra-

ditional collaborative ﬁltering. (Vajjhala et al., 2021)

built a RS based on the analysis of the user Twit-

ter’s proﬁle, therefore recommending items and ser-

vices catered to their tastes. They are able to show a

strong correlation between the user’s tweets and the

category of items/services the user consumes. (Jin

et al., 2020) focuses on showing how the character-

istics of the users, namely visual memory and musi-

cal sophistication help the RS to provide a much bet-

ter recommendation. They found that modifying the

design could help to get better acceptance from the

user while the musical sophistication can play a role

against the recommendation. (Narang et al., 2021)

aims to combine multiple data from the characteris-

tics of the user in an interpretable manner to obtain

a much better recommendation. The model proposed

offers a perspective on the importance of each char-

acteristic depending on which dataset is being used.

WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies

284

(Ayub et al., 2020) propose an uniﬁed approach on

which explicit trust, implicit trust and user preference

similarity get uniﬁed in a rating proﬁle for the target

user. This leads to produce more accurate recommen-

dation. (Nalmpantis and Tjortjis, 2017) shows a RS

in which personality tests are combined with a pre-

existing movie RS. This leads to an increase in the

satisfaction of the user when comparing with current

and available RS. (Tian et al., 2021) proposes a frame-

work that takes into consideration the order on which

the user has interacted with different items. It reduces

the list of all the items to just the most relevant, mak-

ing the choosing of the user more simple. The exper-

iments show how beneﬁcial this approach is. Those

works are specially critical for our case, as we are

focusing on how the RS responds to the interaction.

(Prasad, 2005) intends to solve the ”sequence recog-

nition problem”, the situation on which the RS is un-

able to process the fact that the user already has pur-

chased items more advanced than those that are being

recommended. The proposed solution is based on a

hybrid model that reinforces a collaborative ﬁltering

with a case-based reasoning tool for e-commerce. We

group the related works according to their interests as

shown in Table 1.

3 MODEL OF ADAPTIVE FILTER

As we have seen in Section 2, there is a need for newer

approaches on how to interpret user interaction and

her satisfaction. Also, the lack of references on how

to act when a recommendation has failed propels even

more the proposal of this adaptive ﬁlter.

In this section we introduce the technical deﬁni-

tions of the solution we are presenting. First of all,

we deﬁne the set of items to be recommended and the

set of users that are getting recommended an item:

Deﬁnition 3.1 (Items and their Set). I is a non-empty

and non-trivial set such that I =

{

, i

, ..., i

}

Each i

represents a different item; therefore, I should

be regarded as the set of items to be recommended.

Additionally, for any item i

, it is built as follows:

= {c

, c

, ..., c

}. Each c

represents a differ-

ent characteristic of the item i

Deﬁnition 3.2 (Set of Users). U is a non-empty

and non-trivial set such that U =

{

, u

, ..., u

}

Each u

represents a different user; therefore, U

should be regarded as the set of users that are getting

recommended an item.

Now we deﬁne the core notion of a RS upon which

will deﬁne the adaptive ﬁlter for suboptimal recom-

mendations.

Table 1: Recovering user experience in recommender sys-

tems.

Characteristic-based ﬁne tuning

(Ziegler, 2005)

(Codina and Ceccaroni, 2010)

(Sidana et al., 2021)

(Calero Valdez et al., 2016)

(Burke, )

Boost factor improvement

(De Pessemier et al., 2010)

(Mendoza and Torres, 2020)

Interaction-based recovery

(Adomavicius et al., 2011)

(Li et al., 2018)

(Vajjhala et al., 2021)

(Jin et al., 2020)

(Narang et al., 2021)

(Ayub et al., 2020)

(Nalmpantis and Tjortjis, 2017)

(Tian et al., 2021)

(Prasad, 2005)

Deﬁnition 3.3 (Core Recommender System). A core

RS (cRS) is a function that for each user u

in U, or-

ders the elements of I from 1 to n, where n = |I |,

according to a certain criteria. The criterion on which

the items are ranked is dependant on how the RS is

set-up. For example, in a neighborhood RS, the items

Recommendation Recovery with Adaptive Filter for Recommender Systems

285

would be ranked according to how other users have

evaluated those items. Nevertheless, given our in-

tention to make the results as universal as possible,

these criteria can be any one that the reader might

like. Therefore, the set I is transformed into I

, the

ranked set of items. We deﬁne I

as a well-ordered

set, the result of applying an ordering operation (the

base algorithm of the RS), on I . Afterwards, u

is rec-

ommended the ﬁrst element of the ordered set, I

. By

R(u

, i

) we mean that the user u

is recommended

the item i

. Furthermore, by i

= q we mean that the

item i

is in the position q in the set of ranked items.

All of the above can be expressed as follows:

cRS = ∀u

, u

∈ U, ∃i

, i

∈ I

, i

= 1, and R



, i



It is important to take into account that the proposed

deﬁnition of a cRS is done with an RS that only rec-

ommends the ﬁrst item in mind. Nevertheless, it could

be easily modiﬁed so that the cRS actually recom-

mends a subset of items from I

. In the case we are

working with more than one item, the user would be

recommended the items ordered from the 1st position

to pth position, where p is the number of items to be

recommended.

After we have deﬁned the notion of cRS, we need to

deﬁne how the evaluation of an item happens and the

similarity between items.

Deﬁnition 3.4 (Evaluation of an Item). E (u

, i

)

means that the user u

evaluates the item i

. This has

two possible outcomes: (1) E (u

, i

) = Sat and (2)

E (u

, i

) = ¬Sat. If the result of the evaluation is (1),

the RS has succeeded in recommending an item; i. e.,

the user is satisﬁed and the recommended item was

optimal. If the result is (2), we say that the RS has

failed; i. e., the user is not satisﬁed and the recom-

mended item was suboptimal.

Deﬁnition 3.5 (Similar Items). Any two items i

and

are similar, in symbols S(i

, i

), if and only if they

share most of their characteristics.

After introducing the previous concepts, we deﬁne the

adaptive ﬁlter.

Deﬁnition 3.6 (Adaptive Filter). An adaptive ﬁlter

is a cRS in which, after an evaluation such that

E(u

, i

) = ¬sat, the set I

is revised via ﬁltering. For

this, the set I

is modiﬁed so that all the items that are

similar to the suboptimal item i

are pulled out, thus

giving birth to a new ranked set of items I

. As the

set I

is built, so is the set I

, the set of discarded

items. This set is built as follows:

= {∀i



, i



, I

∪ {i

}}

Then, for any user u

and a selected item i

, the new

ranked set I

is built from the original set I

as fol-

lows:

= {∀i



, i



, I

− {i

}}

This can be iterated a ﬁnite number of times p at most,

where p = |I

|. Also, it is obvious that I

can be de-

ﬁned from I

and I

, but we have included its deﬁni-

tion as standalone to make everything clearer. From

an implementation point of view it should be deﬁned

from both sets so it is less taxating for the system.

Deﬁnition 3.7 (Similarity Threshold). Deﬁnition 5

is an obvious application of the Jaccard Index (Lee,

2017) and, therefore, we need to establish a Similar-

ity Threshold (ST). This ST is equal to the ratio of the

summatory of the constants of preference (k) of each

characteristic, per item, multiplied by the fraction of

the set of discarded items. Thus, ST is as follows:

ST =



∑

, ..., k

|I |







For qualitative characteristics, the previous applies

automatically. For quantitative characteristics, they

are to be considered equal if and only if, for charac-

teristics c

and c

, c

= c

± 15% follows. All of the

above means that whenever an RS fails consecutively,

less and less items are pulled from the set. Therefore,

even if the RS fails too many times consecutively, the

set I

would not be emptied.

Remark 3.1 (Constant of Preference). The constants

of preference (k) are introduced in the previous deﬁni-

tion to ponder the characteristics to each user. These

constants of preference are to be obtained from the

user in the same way the RS would deal with a cold

start (Han et al., 2019). Some of the tools that can

be used for that matter include small surveys at the

beginning of the use of the RS, or emotion detection

on the user’s previous reviews (Ishwarya et al., 2019)

among others. The reader might choose the one that

feels more adequate.

3.1 Adaptive Filtering Process

After deﬁning the adaptive ﬁlter, we show its func-

tionality more clearly with a process model. For that

matter, Figure 1 is the representation of a compari-

son of the process model of how an adaptive ﬁlter

works against a regular recommender system without

the adaptive ﬁlter.

A starting set of items I is passed through the pref-

erences of the user u thus, obtaining a ranked set of

items I

. This ranked set, I

, allows us to present

an item i

as a recommendation to the user u. Af-

ter the user makes the purchase of the recommended

item, there are two different possibilities: either u

is satisﬁed with the item, E(u, i

) = Sat, and so the

ranked set of items is just missing the previously rec-

ommended item; or either the user is not satisﬁed with

WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies

286

Set of

Items I

User u

Ranked set

of Items I

Evaluation

of top

ranked

item, i

from I

minus i

Similar

items to i

are ﬁltered

out of I

Discarded

set of

items I

New ranked

set of

items I

Set of

Items I

User u

Ranked set

of Items I

Evaluation

of top

ranked

item, i

from I

Rearranged

Sat

¬Sat Sat or ¬Sat

Figure 1: Adaptive Filter Process Model (l) and Recommender System without the Adaptive Filter Process Model (r).

the item, E(u, i

) = ¬Sat, and two new sets are cre-

ated: a set of discarded items, I

, containing all the

items from the original ranked set that were similar to

the non-satisfactory item, and another, I

, that is ob-

tained subtracting the elements of the set of discarded

items from the original set of items I .

4 PRELIMINARY VALIDATION

To further describe the implementation of an adaptive

ﬁlter, we describe a validation from a within-subject

study design perspective. This is due to the fact that

it allows for a quicker and better identiﬁcation of

differences rather than a between-subject design.

We are validating two different scenarios based on

a fragment of a real-world dataset (U. S. Food &

Drug Administration, 2021). In both scenarios, a

patient has reached the Emergency Room with an

unspeciﬁc disease. After a preliminary diagnosis

the medical team is using two different systems that

allow them to rate and ﬁnd the perfect medical drug

to be administered. Both systems can be inputted

if the treatment worked or not. Both systems have

a built-in medical drug RS, that is selecting which

the recommendation is, takes care of the symptoms

and compares other people cases. Each time the

systems recommend just one medical drug to be

used, this is supported by the concern that to obtain

a better diagnosis just one drug can be used at the

same time. Furthermore, drug use usually needs to

be carefully planned and one usually does not work

on various options at the same time. In the second

scenario, the medical team uses a system whose RS

has the adaptive ﬁlter technique from Deﬁnition 6.

The fragment of interest of the set of drugs that the

systems are using is the following:

DrugsDataset={Ashlyna, Daysee, Jaimiess, Mal-

morede, Namenda, Namzarinc, Prozac, Sarafem,

Zovia}

Each of these drugs has some characteristics, as

instantiated from Deﬁnition 1. These characteristics

are their active ingredient, strength and route form.

Furthermore, each one of them shall be used in

obtaining a recommendation. These characteristics

have been selected as basic elements to make a drug

itself: the active ingredient represents what the drug

focuses on and, therefore, what it would be useful for;

the strength is a clear indicator of the performance;

and the route is sign of how easy or hard would be to

apply that drug. Nevertheless, it is worth mentioning

that these characteristics are just a representation of a

much bigger set and the reader might feel compelled

to choose different ones for a different validation. All

this is summarized in Table 2.

The medical team has administered Malmorede

and Ashlyna; the patient has a good response to

those. Given that the patient is improving with the

Recommendation Recovery with Adaptive Filter for Recommender Systems

287

Table 2: Dataset Fragment of Medical Drugs.

Drug Name Active Ingredient Strength Route

Ashlyna Ethinyl Estradiol; Levonorgestrel 0.03mg; 0.01mg Tablet; Oral

Daysee Ethinyl Estradiol; Levonorgestrel 0.03mg; 0.02mg Tablet; Oral

Jaimiess Ethinyl Estradiol; Levonorgestrel 0.02mg; 0.01mg Tablet; Oral

Malmorede Ethinyl Estradiol; Ethynodiol Dyacetate 0.05mg; 1mg Tablet; Oral

Namenda Memantine Hydrochloride 5mg Tablet; Oral

Namzaric Donepezil Hydrochloride; Memantine Hydrochloride 10mg; 14mg Capsule; Oral

Prozac Fluoxetine Hydrochloride 10mg Capsule; Oral

Sarafem Fluoxetine Hydrochloride 20mg Capsule; Oral

Zovia Ethinyl Estradiol; Ethynodiol Dyacetate 0.05mg; 1mg Tablet; Oral

drugs administered, both systems have ordered the

Drugs Dataset as follows:

DrugsDataset

={Daysee=1, Jaiminess=2, Zovia=3,

Prozac=4, Namenda=5, Sarafem=6}

Thus, the medical team is recommended to con-

tinue the treatment with Daysee, as it comes from

Deﬁnition 3. The main reason is that it shares active

ingredient and has a similar strength to the last drug

used, Ashlyna. Trusting the system of each scenario,

Daysee is administered. However, in this case,

the drug has little to no effect, and so the medical

team evaluates the drug as in Deﬁnition 4. In the

ﬁrst scenario, after letting the system know that

the drug was not cutting it, the recommendation is

administering Jaiminess, the next drug highest in

the ranked set. With this drug being quite similar

to Ashlyna and Daysee, there is a high probability

that the drug will not work properly, therefore, the

medical team will stop being satisﬁed with the RS.

In the second scenario, the one with a system with

a RS with the adaptive ﬁlter, after letting the system

know that the drug was not efﬁcient, the adaptive

ﬁlter built in the system ﬁlters out drugs with similar

characteristics, as it follows from Deﬁnition 6. In

this case, as Jaimiess has the same active ingredient,

route and the strength is in a 15% range of variation,

as speciﬁed in Deﬁnition 7. The newly ranked set of

drugs is as follows:

DrugsDataset

={Zovia=3, Prozac=4, Namenda=5,

Sarafem=6}

Also, a new set of discarded drugs is built:

DiscardedDrugs={Jaiminess}

These new sets are built according to each of

the algorithms of Deﬁnition 6. Finally, it recom-

mends them to use to Zovia, the best ranked Drug

in the updated set of drugs. It can be seen that

adaptive ﬁlter offers a viable alternative for a user

who is facing a suboptimal recommendation. In

this example, the medical team, after being given a

suboptimal recommendation, Daysee, is offered a

new option, Zovia, as the best ﬁt. This is because

Zovia shares most of its characteristics with the

previous experiences from the team. Furthermore,

they may trust in the adaptive ﬁlter RS for further

recommendations, as it considers recommendation

recovery. Therefore, the life-span of the RS has gone

from a one-shot to a multiple-use, and may become

a crucial part of the medical team’s diagnostic cycle.

We can further infer that the satisfaction of the user

may improve, even if there has been a suboptimal

recommendation, as they are offered a new alternative

after a bad experience. But this alternative has been

catered to their previous experiences. Additionally,

some drugs are pulled out of the drugs dataset and

they will not see them recommended again for the

same clinical case. This serves to improve their

conﬁdence in the agent and extend their use of the

RS, since they may feel that their preferences are

taken into account.

5 DISCUSSION

One of the effective solutions to address the unsatis-

ﬁed recommendations is the critique-based RS (Jan-

nach and Zanker, 2020). We name our solution as

adaptive ﬁlter RS and compare with critique-based

RS. Both solution are user-oriented and their core

techniques might rely in the same intuitions: that they

intend to let users instantly adjust the recommenda-

tions in the case that the recommended items are not

favored by the users.

There are several differences between the two so-

lutions. The main one is the intention behind each

system: while critique-based systems are focused on

ﬁne-tuning the users preferences, the adaptive ﬁlter-

ing focuses on avoiding any recommendation pitfalls

WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies

288

the user might ﬁnd. Critique-based RS focus on

ﬁne tuning the best options available (Ramnani et al.,

2018), whereas the adaptive ﬁlter RS attempt to hide

the worst recommendations.

On the technical aspect, adaptive ﬁltering happens

after the purchase, while critique-based systems are

set before or after the purchase. The purchase does

not make a difference for the latter. Additionally,

critique-based RS can only ﬂag items as low-priority,

while adaptive ﬁlter expels them out of the recom-

mendation pool. It is thus possible that critique-based

RS may recommend the same items again, even if

they are discarded by the user. On the other hand,

items discarded by an adaptive ﬁlter are not shown

again, therefore ﬁnding a working niche in which the

user does not want to be shown similar items ever

again. Also, it is worth mentioning that our solution,

as it decreases the set of items to be recommended

each time when the RS fails, makes the computational

and interactive costs also decrease.

Finally, as we have seen in Section 4, the value

that the adaptive ﬁlter RS can offer to some speciﬁc

domains is immense. In the particular case of health-

care, the proposed solution offers an efﬁcient method

to avoid recommending certain items that are of lim-

ited usage, which makes the system more intuitive and

when applied in healthcare domain, it might be crucial

to save a life. This is particularly important in situa-

tions of the Emergency Room, the proposed solution

is able to exclude a whole family of medical drugs

that has already shown limited effects on the patient.

This can further facilitate the doctors to conduct more

accurate diagnosis and treatment.

6 CONCLUSION

In this paper, we have proposed an adaptive ﬁlter for

recommender system that is designed to recover failed

recommendations. It can be easily integrated to the

existing RS without modifying the recommendation

algorithm, and the proposed adaptive ﬁlter technique

adds limited computational complexity. The addi-

tional computational complexity is the creation of the

set of discarded items. In order to validate the pro-

posed solution, we have conducted a preliminary eval-

uation with the medical recommender systems. The

evaluation result has showed that the proposed solu-

tion offers an efﬁcient method to avoid recommend-

ing certain items that are of limited effects on patients.

We believe that the adaptive ﬁlter can be further ap-

plied in other domains to recover the failed recom-

mendations.

To continue this research, there are multiple lines

of further investigation that are to be explored in the

future. Among others, we will include the deﬁni-

tion of tools that allow for an evaluation of RS that

implement this technique, and integrate this solution

into existing RS. The development of evaluation tools

would allow to quantify the performance and the im-

provement that the adaptive ﬁlter offers when com-

pared to traditional RS, something that cannot be done

in the time being. Furthermore, it would be interest-

ing to see its integration with semantic web RS as

those of (Ziegler, 2005) but from the perspective of

inconsistent data offered on (Blanco et al., 2021).

REFERENCES

Adomavicius, G., Mobasher, B., Ricci, F., and Tuzhilin, A.

(2011). Context-Aware Recommender Systems. AI

Magazine, 32(3):67–80. Number: 3.

Alves, P., Saraiva, P., Carneiro, J., Campos, P., Martins,

H., Novais, P., and Marreiros, G. (2020). Modeling

Tourists’ Personality in Recommender Systems: How

Does Personality Inﬂuence Preferences for Tourist At-

tractions? In Proceedings of the 28th ACM Confer-

ence on User Modeling, Adaptation and Personaliza-

tion, pages 4–13, New York, NY, USA. Association

for Computing Machinery.

Ayub, M., Ghazanfar, M. A., Mehmood, Z., Alyoubi,

K. H., and Alfakeeh, A. S. (2020). Unifying user

similarity and social trust to generate powerful rec-

ommendations for smart cities using collaborating

ﬁltering-based recommender systems. Soft Comput-

ing, 24(15):11071–11094.

Ben Mimoun, M. S., Poncin, I., and Garnier, M. (2012).

Case study—Embodied virtual agents: An analysis

on reasons for failure. Journal of Retailing and Con-

sumer Services, 19(6):605–612.

Blanco, J. M., Ge, M., and Pitner, T. (2021). Modelling

inconsistent data for reasoners in web of things (ac-

cepted). In Procedia Computer Science. Elsevier.

Burke, R. Hybrid Web Recommender Systems. In The

Adaptive Web: Methods and Strategies of Web Per-

sonalization, pages 377–408. Springer, Berlin, Hei-

delberg.

Calero Valdez, A., Zieﬂe, M., Verbert, K., Felfernig, A.,

and Holzinger, A. (2016). Recommender Systems for

Health Informatics: State-of-the-Art and Future Per-

spectives. In Machine Learning for Health Informat-

ics: State-of-the-Art and Future Challenges. Springer

International Publishing.

Chedrawy, Z. and Abidi, S. S. R. (2009). A Web Recom-

mender System for Recommending, Predicting and

Personalizing Music Playlists. In Web Information

Systems Engineering - WISE 2009, pages 335–342,

Berlin, Heidelberg. Springer.

Codina, V. and Ceccaroni, L. (2010). A Recommendation

System for the Semantic Web. In Distributed Comput-

Recommendation Recovery with Adaptive Filter for Recommender Systems

289

ing and Artiﬁcial Intelligence, Advances in Intelligent

and Soft Computing, Berlin, Heidelberg. Springer.

De Pessemier, T., Dooms, S., Deryckere, T., and Martens,

L. (2010). Time dependency of data quality for col-

laborative ﬁltering algorithms. In Proceedings of

the fourth ACM conference on Recommender systems,

RecSys ’10, pages 281–284, New York, NY, USA. As-

sociation for Computing Machinery.

Fang, H., Zhang, D., Shu, Y., and Guo, G. (2020).

Deep Learning for Sequential Recommendation: Al-

gorithms, Inﬂuential Factors, and Evaluations. arXiv:

1905.01997.

Han, D., Li, J., Li, W., Liu, R., and Chen, H. (2019). An app

usage recommender system: improving prediction ac-

curacy for both warm and cold start users. Multimedia

Systems, 25(6):603–616.

Ishwarya, M. V., Swetha, G., Saptha Maaleekaa, S., and

Anu Grahaa, R. (2019). Efﬁcient Recommender Sys-

tem by Implicit Emotion Prediction. In Advances in

Big Data and Cloud Computing, Advances in Intelli-

gent Systems and Computing, pages 173–178.

Jannach, D., Manzoor, A., Cai, W., and Chen, L. (2020).

A Survey on Conversational Recommender Systems.

arXiv:2004.00646 [cs]. arXiv: 2004.00646.

Jannach, D. and Zanker, M. (2020). Interactive and

Context-Aware Systems in Tourism. In Handbook

of e-Tourism, pages 1–22. Springer International Pub-

lishing, Cham.

Jin, Y., Tintarev, N., Htun, N. N., and Verbert, K. (2020).

Effects of personal characteristics in control-oriented

user interfaces for music recommender systems. User

Modeling and User-Adapted Interaction, 30(2):199–

249.

Kang, J., Condiff, K., Chang, S., Konstan, J. A., Terveen,

L., and Harper, F. M. (2017). Understanding How

People Use Natural Language to Ask for Recommen-

dations. In Proceedings of the Eleventh ACM Confer-

ence on Recommender Systems, pages 229–237, New

York, USA. ACM.

Lavanya, R., Khokle, T., and Maity, A. (2021). Review on

Hybrid Recommender System for Mobile Devices. In

Artiﬁcial Intelligence Techniques for Advanced Com-

puting Applications, pages 477–486, Singapore.

Lee, S. (2017). Improving Jaccard Index for Measuring

Similarity in Collaborative Filtering. In Information

Science and Applications 2017.

Li, B.-H., Zhang, A.-M., Zheng, W., Wan, S., Qin,

X.-L., Li, X., and Yin, H.-L. (2018). GRIP: A

Group Recommender Based on Interactive Preference

Model. Journal of Computer Science and Technology,

33(5):1039–1055.

Mahmood, T., Ricci, F., Venturini, A., and H

opken, W.

(2008). Adaptive Recommender Systems for Travel

Planning. In Information and Communication Tech-

nologies in Tourism, pages 1–11, Vienna, Austria.

Mendoza, M. and Torres, N. (2020). Evaluating content

novelty in recommender systems. Journal of Intelli-

gent Information Systems, 54(2):297–316.

Mimoun, M. S. B., Poncin, I., and Garnier, M. (2017). Vir-

tual Sales Agents: The Reasons of Failure. In The

Customer is NOT Always Right? Marketing Orienta-

tionsin a Dynamic Business World.

Naiseh, M., Jiang, N., Ma, J., and Ali, R. (2020). Explain-

able Recommendations in Intelligent Systems: Deliv-

ery Methods, Modalities and Risks. In Research Chal-

lenges in Information Science, Lecture Notes in Busi-

ness Information Processing, Cham. Springer Interna-

tional Publishing.

Nalmpantis, O. and Tjortjis, C. (2017). The 50/50 Rec-

ommender: A Method Incorporating Personality into

Movie Recommender Systems. In Engineering Appli-

cations of Neural Networks, pages 498–507. Springer

International Publishing.

Narang, K., Song, Y., Schwing, A., and Sundaram, H.

(2021). FuseRec: fusing user and item homophily

modeling with temporal recommender systems. Data

Mining and Knowledge Discovery, 35(3):837–862.

Narducci, F., de Gemmis, M., Lops, P., and Semeraro, G.

(2018). Improving the User Experience with a Con-

versational Recommender System. In AI*IA 2018 –

Advances in Artiﬁcial Intelligence, pages 528–538.

Springer International Publishing.

Prasad, B. (2005). HYREC: A Hybrid Recommendation

System for E-Commerce. In Mu

noz-

Avila, H. and

Ricci, F., editors, Case-Based Reasoning Research

and Development, Lecture Notes in Computer Sci-

ence, pages 408–420, Berlin, Heidelberg. Springer.

Ramnani, R. R., Sengupta, S., Ravilla, T. R., and Patil, S. G.

(2018). Smart entertainment - A critiquing based di-

alog system for eliciting user preferences and making

recommendations. In 23rd International Conference

on Applications of Natural Language to Information

Systems, Paris, France, volume 10859, pages 456–

463.

Sanchez, O. R., Torre, I., He, Y., and Knijnenburg, B. P.

(2020). A recommendation approach for user privacy

preferences in the ﬁtness domain. User Modeling and

User-Adapted Interaction, 30(3):513–565.

Sidana, S., Troﬁmov, M., Horodnytskyi, O., Laclau, C.,

Maximov, Y., and Amini, M.-R. (2021). User prefer-

ence and embedding learning with implicit feedback

for recommender systems. Data Mining and Knowl-

edge Discovery, 35(2):568–592.

Tian, X., Hao, Y., Zhao, P., Wang, D., Liu, Y., and

Sheng, V. S. (2021). Considering Interaction Se-

quence of Historical Items for Conversational Recom-

mender System. In Database Systems for Advanced

Applications, pages 115–131. Springer International

Publishing.

U. S. Food & Drug Administration (2021). Drugs@FDA:

FDA-Approved Drugs.

Vajjhala, N. R., Rakshit, S., Oshogbunu, M., and Salisu, S.

(2021). Novel User Preference Recommender System

Based on Twitter Proﬁle Analysis. In Soft Computing

Techniques and Applications, Advances in Intelligent

Systems and Computing, pages 85–93. Springer.

Ziegler, C.-N. (2005). Semantic Web Recommender Sys-

tems. In Current Trends in Database Technology -

EDBT 2004 Workshops, Lecture Notes in Computer

Science, pages 78–89, Berlin, Heidelberg. Springer.

WEBIST 2021 - 17th International Conference on Web Information Systems and Technologies

290