Explainable and unexpectable recommendations using relational learning on multiple domains

Abstract

In this research, we combine relational learning with multi-domain to develop a formal framework for a recommendation system. The design of our framework aims at: (i) constructing general rules for recommendations, (ii) providing suggested items with clear and understandable explanations, (iii) delivering a broad range of recommendations including novel and unexpected items. We use relational learning to find all possible relations, including novel relations, and to form the general rules for recommendations. Each rule is represented in relational logic, a formal language, associating with probability. The rules are used to suggest the items, in any domain, to the user whose preferences or other properties satisfy the conditions of the rule. The information described by the rule serves as an explanation for the suggested item. It states clearly why the items are chosen for the users. The explanation is in if-then logical format which is unambiguous, less redundant and more concise compared to a natural language used in other explanation recommendation systems. The explanation itself can help persuade the user to try out the suggested items, and the associated probability can drive the user to make a decision easier and faster with more confidence. Incorporating information or knowledge from multiple domains allows us to broaden our search space and provides us with more opportunities to discover items which are previously unseen or surprised to a user resulting in a wide range of recommendations. The experiment results show that our proposed algorithm is very promising. Although the quality of recommendations provided by our framework is moderate, our framework does produce interesting recommendations not found in the primitive single-domain based system and with simple and understandable explanations.

Keywords

Recommender system explainable multi-domain recommendation rule generation serendipity

1. Introduction

Recommender systems (RS) come closer to our daily life than ever, influencing us on the music we listen to, the movies we watch, the books we read, the persons we make friend with, for example. A recommender system has, thus, become one of the most basic supportive techniques in an online world and it has proven to be a major source of enhanced functionality, user satisfaction, and revenue improvement.

The most common critical issues found with RS include but not limited to maximizing prediction accuracy, solving cold-start problem, reducing sparsity, and providing novelty, diversity and serendipity [25]. However, solving one problem may create another problem, or a trade-off. All issues have not been perfectly solved since many current recommender algorithms seem to be locked away inside a black box [10]. Once an algorithm is processed, it is quite difficult to understand why it gives a particular recommendation to a set of data. If we can understand the reason behind the recommendation, we believe we will be able to possibly find the way to handle the problems mentioned above more effectively. One example is that the cold-start and sparsity problems can be managed using an explanation in the form of a general rule. We also believe that the explanation which indicates how the recommended items that relate to users’ preference items either directly or indirectly, can grab users’ attention and influence their decision to try out or to purchase eventually.

Moreover, the traditional recommendation approaches (i.e. Content-based Filtering and Collaborative Filtering) assume that users want to see the content similar to what they (or their friends) already rate highly. Thus, users are locked into clusters of similarity and may find recommendations provided practically useless for possibly a number of reasons i.e. the user has all items on the list, or the user feels bored by having too many similar items, or the user is no longer interested in the similar items. This is the long tail problem in RS [24], where the recommendations are centered around very small popular items in the head. There may be surprised items from the long tail that users may be interested in but they do not know about or are not aware that those items exist. To overcome this problem, RS should suggest broadly the novel and unexpected items from the long tail which is a challenging task.

Novel recommendations are recommendations of those items that the user does not know (previously unknown) [8]. Unexpected recommendations are those recommendations that significantly depart from the user’s expectations [25], either previously known or unknown. In fact, there is possibly an overlap between novel and unexpected recommendations. To achieve novel or unexpected recommendations, we need to look beyond a single domain that contains only a user’s preferences or ratings. We believe that the discovery of surprising correlations in the repositories coming from diverse domains can lead us to novel and unexpected recommendations. In addition, the recommendation should be able to be made in any domain depending on the generated relational rule, not limited to a particular domain specified by the user. In other words, there is no need for a user to define the target domain at the beginning of the process. The following examples attempt to illustrate our ideas.

Example 1: Suppose Kelly likes music “A” and enjoys movie “B” which has three music C, D, E. With the single-domain based RS, she will be recommended only the music which is similar to music “A”. Unfortunately, the music C, D, E will not be recommended to her since they are not considered as similar to “A”. We believe that if the RS can somehow relate the two different domains, music and movie, it can then recommend items besides the music similar to A to Kelly. For example, Kelly can be recommended with the music C, D, E or their similarities or the music which somehow relates to music A via its properties or attributes (i.e. music theme, movie genre, music artist). At any rate, no matter what the recommendation is, it will not be what the user expects.

Example 2: Suppose Kelly likes movie “A” and John likes movie “B”. Both “A” and “B” share the same genre “thriller”. John likes music “C” which is not the item that Kelly is familiar with (found in Kelly’s preference). With the links or the relations between this information, RS can possibly recommend music “C” to Kelly or its similarity or the music which somehow relates to music “C” via its properties or attributes. Again, the recommendation will be a surprise to Kelly.

Our other key supporting idea is the success of cross-domain recommender systems which suggest items in one domain using information from another domain or other multiple domains. User’s knowledge acquired in one domain can be transferred to and exploited in other domains instead of treating each domain separately. Cross-domain approach improves prediction accuracy by reducing data sparsity and offers added values to recommendations by providing diversity, novelty and serendipity predictions [6]. These benefits thus encourage us to develop our framework on multi-domain. However, unlike the typical cross-domain approach, a setting for specific domains (source and target domain) is not required prior to constructing the recommendation rules and no knowledge transfer across domains is needed.

Relational learning has already shown its use in RS. The evidences from the researches by Kouki et al. [26] and Catherine and Cohen [29] indicate that relational learning provides better recommendations by incorporating additional information, compared to traditional methods with a single dyadic relationship between the objects i.e. users and items. The most important capabilities of relational learning are that: it produces simple and understandable rules, it does not need to predefine the role of the domain and it constructs the general rule for recommendation. Consequently, the relational learning captures our interest to model and provide a potential solution for explainable multi-domain recommendations.

We propose a new formal framework for RS. The key components of our framework are: (i) the use of relational learning to learn the relations, possibly new, between users and items, (ii) the incorporation of user information and knowledge from various domains. With the advantages of relational learning, our framework is able to construct the general rules associating with probability for recommendations. These rules are used to create the recommendations for the users with transparency by describing how the recommendation was selected for a particular user. Moreover, the recommendation can be made in any domain, no need to pre-specify the source-target domain pair. With the information available in various domains, our framework is able to deliver a broad range of recommendations including novel and unexpected items. Additional domains, new data (e.g., features), and context (e.g., location, time, and mood) can simply be incorporated.

The remainder of this paper is organized as follows. In Section 2, the studies relating to explainable and cross-domain recommender systems are discussed. Our methodology and settings are described in Section 3. In Section 4, our experimental settings, evaluations, and results are presented. The results are summarized in Section 5 and the conclusion of this paper is presented in Section 6.

2. Related work

2.1 Explanations in RS

Although producing good recommendations is the primary goal of the recommender systems, it is also desirable that such systems provide an explanation accompanying the recommendation. Explanations can serve multiple purposes [23], one of which is “Transparency”. An explanation that describes how a recommendation was chosen makes the system transparent to the user as it provides clarity as to how a recommendation was picked for a user when the system shows ambiguous recommendations. “Customers Who Bought This Item Also Bought …” on Amazon is a case in point for a transparent recommendation. Explanations can also serve other purposes such as “Effectiveness” by helping users make good decisions and “Efficiency” by helping users make decisions faster. The rest of explanatory goals described in [23] are scrutability, trust, persuasiveness, and satisfaction. Some well-known examples of explanations in commercial and academic systems are described in Table 1 [22, 7]. Most of the explanations are presented in natural language which could possibly lead to a misunderstanding due to their complexity and lack of clarity. It is sometimes difficult to use language in a precise and unambiguous way without making the explanation wordy and difficult to read.

Table 1
Examples of explanations in commercial and academic systems

System	Item domain	Explanation
Amazon	e.g., Books, Movies	Content-based
Findory	News	Preference-based
LibraryThing	Books	Collaborative-based
LoveFilm	Movies	Content-based
OkCupid	People to date	Preference-based
Pandora	Music	Preference-based
StumbleUpon	Web pages	Preference-based
Qwikshop [14]	Digital cameras	Preference-based
LIBRA [19]	Books	Content-based, Collaborative-based
News Dude [2]	News	Preference-based
MovieLens [10, 3]	Movies	Collaborative-based
Top Case [4]	Holiday	Preference-based
ACORN [27]	Movies	Preference-based

2.2 Relational learning in RS

The success of relational learning has been shown in a number of researches. For example, the recent work of Hoxha and Rettinger [9] and Kouki et al. [26] which applied relational learning to the hybrid recommendations demonstrated that incorporating additional information for users and items was beneficial to the cold-start settings in particular [33]. Another progress in relational learning was presented in Hoxha et al. [9], a probabilistic graphical modeling representation using Markov Logic Networks to combine content-based with collaborative filtering. Its recommendation task began with predicting the existence probability of a relation between a particular user and a particular item, and then returning the recommendation for the query predicate by choosing the items with the high probability according to the specified threshold.

In relational learning, first-order logic rule representation is simpler and more concise compared to the rule representation in traditional item-based collaborative filtering algorithms as shown in the following example. The representation of the rule: “Similar items get similar ratings from a user” is “IF items $i_{1}$ and $i_{2}$ are similar, for a given similarity measure, AND user u rates item $i_{1}$ highly, THEN user u will rate item $i_{2}$ highly”, in traditional item-based collaborative filtering algorithms and Eq. (1) is its representation in first-order logic.

$\displaystyle\textit{similarItems}_{\textit{sim}}(i_{1},i_{2})\land\textit{% highRating}(u,i_{1})\implies\textit{highRating}(u,i_{2})$ (1)

In first-order logic rules, constants represent the objects in a domain of interest. Variable symbols range over the objects. Predicate symbols represent the relationship between objects or features of objects. Variables and constants might be typed, in which case variables only range over objects of the given type.

In 2015, the study by Kouki et al. [26] showed that relational learning framework could be used to develop a general and extensible hybrid RS framework called Hybrid Probabilistic Extensible Recommender. This framework also provided a mechanism to extend the system by incorporating and reasoning over unspecified types and similarity measures of additional information collected from several sources. A learning method used to appropriately balance the different input signals from many information sources was also discussed.

In 2016, probabilistic program showed its use in single-domain based recommendation by Catherine and Cohen [29]. They proposed three methods that used knowledge graphs for making personalized recommendations. Given that a user liked specific movies and entities in the past, ranked new movies (e.g., The Bridge of Spies) for that user using ProPPR. The methods gave large improvements compared to the state-of-the-art method that used knowledge graphs. The study of the behavior of the methods as rating matrix density was also discussed.

It is evident from the researches by Hoxha et al. [9], Kouki et al. [26] and Catherine and Cohen [29] above that a better recommendation performance is obtained by relational learning with additional information incorporated as opposed to a traditional method with a single dyadic relationship between the objects i.e. users and items. As a result, we were drawn into relational learning to model and to provide potentially a solution to construct rules for recommendation.

2.3 Cross-domain recommendations

Cross-domain recommendation problem is introduced, instead of performing a recommendation on a single domain and focusing on a certain market because users normally provide feedback for items of different types and express their opinions on different systems. With cross domain, prediction accuracy is improved because data sparsity is reduced. Cross domain also offers additional values to the recommendations as it gives diverse, novel and serendipitous predictions [6].

In cross-domain recommendation tasks, items in the target domain are recommended to users in the source domain. Aggregating knowledge approach and transferring knowledge approach are the two types of cross domain approaches and the difference between the two approaches lies in how the knowledge from the source domain is exploited. In the aggregating knowledge approach, user preferences or aggregate models from different domains may be merged, e.g., ratings for items in a book and a movie become a joint matrix and then common single domain approaches can be used to recommend particular items to users in the target domain.

The transferring knowledge approach first links the domains, e.g., linked through attribute by transferring function from a book domain to a movie domain, and then the knowledge among domains can be transferred for recommendation. However, as previously mentioned, these two approaches have their own limitations. Aggregating knowledge approach is designed for particular cross-domain scenarios which are quite difficult to generalize while transferring knowledge approach is computationally costly.

Cross-domain caught our attention due to its proven success in a lot of work. We can exploit and leverage the common attributes, semantics and other hidden knowledge across distinct domains to generate novel or unexpected recommendations. In this study, we also investigate whether the user’s preference in one domain, e.g., music, relates to other facets of users’ life and preferences (in other domains).

There are advantages and disadvantages in the explainable-based RS and the cross domain-based RS. One drawback of most explainable-based RS is the explanation form. Examples of explanation forms are textual sentence, tag cloud, visual image which seem to be complicated and require an extra afford to comprehend. In cross domain-based RS, the recommendation made is limited to the pre-specified domain. In addition, the learning task is computationally inefficient and quite difficult to generalize.

Our proposed framework is a new paradigm which takes advantage of ILP to provide recommendation rules in if-then logical format that allows us to form a clear and concise explanation, unlike most explainable-based RS. Our approach also allows the recommendation to be made in multiple domains, unlike the cross domain-based RS. With the ability of ILP, our approach is extensible since the re-training is not necessary. Other features of our approach which differ from the typical explainable and cross domain approach are discussed in the next section.

3. Research methodology

The main aim of our research is to develop a new formal framework for RS based on two aspects: including the explanation which explains why the system gives a particular recommendation to a user and providing a novel or an unexpected recommendation to the user. To achieve our goal, the following two questions need to be addressed: (i) What to explain and how? (ii) How to generate a novel and/or an unexpected recommendation?

To answer the first question, what to explain is to provide information which the users find helpful and beneficial and help them make decision i.e. to find items that they like. Therefore, we focus on the explanation which presents clearly why such recommendation is chosen for a user.

To deliver the explanation which is not only clear but also simple, understandable and unambiguous, we propose to use relational learning to construct rules for recommendation generation. As a result, the explanation is represented in the form of logic, the formal language.

Relational learning has received extensive attention in recent years since most of the data available is organized by the relations between entities. Relational learning refers to learning in a context where there may be relationships between learning examples, or where these examples may have a complex internal structure (i.e. consisting of multiple components and there may be relationships among these components) [13]. One of the main tasks of the relational learning is to make predictions of possible new relations. Relational learning algorithms learn the definition of a new relation in terms of existing relations in the database [12] since items in the world are connected by various relations. In this work, we use Inductive Logic Programming (ILP) in a probabilistic setting to perform relational learning from examples.

For the answer to the second question, what is new or what is unexpected to a user is the item which the user is not familiar with. In other words, in this research, new or unexpected items refer to the suggested items that differ from those recommended based on similarities either between users or items. To generate the new or unexpected items, we, therefore, need to go beyond the traditional approach since it deals mainly with single dyadic relationships between users and items on a single domain. Adding other domains to the system allows us to broaden our search space and provides us with more opportunities to discover items which are previously unseen or surprised to a user. To deliver a new or an unexpected item and which would still be useful to a user, we need to relate user’s information i.e. preference, demographic, user’s background with some information/knowledge provided in various domains either from other users or items.

The notions discussed above encourage us to develop our framework for a recommendation system using relational learning on multiple domains. In our point of view, we aimed at creating our framework with the following features.

1.
Our framework constructs a general recommendation rule, not an item, unlike a typical RS. The benefits of the rules in general form are that: the rules can be transformed into a clear and understandable explanation to accompany the recommended items, and the rules can possibly be used to manage the cold-start and sparsity problems. More importantly, an explanation is in if-then logical format, which is simpler and more concise compared to most of the explanation forms (e.g., textual sentence, tag cloud, visual image) currently available.
2.
The recommendation can be made in any domain led by the learned rule, unlike a typical cross-domain RS. For a single-domain based RS, a recommendation can be made only for that domain. For cross-domain based RS, a recommendation can be made for the pre-specified target domain only. The following are examples of rules constructed using our framework on three different domains; music, movie and book. Each rule can be used to recommend the items in each different domain. Rule Eq. (2) can be used to recommend “Music” composed by “Ed Sheeran” to the user if the user likes the book “Harry Potter”. Rule Eq. (2) is used to recommend the “Music” according to the conditions on both “Movie” and “Book” domains. Rules Eqs (2) and (2) can be used to recommend the “Movie” and the “Book” respectively.

$\displaystyle 0.37552::\textit{likeMusic}(\textit{User},\textit{Music})\text{ % :- }$ (2) $\displaystyle\textit{likeBook}(\textit{User},\textit{harrypotter}),\textit{% musicArtist}(\textit{edsheeran},\textit{Music}).$ $\displaystyle 0.13497::\textit{likeMusic}(\textit{User},\textit{backinblack})% \text{ :- }\textit{likeMovie}(\textit{User},\textit{Movie}),$ $\displaystyle\textit{movieProductionStudio}(\textit{Movie},\textit{% marvelstudios}),$ (3) $\displaystyle\textit{likeBook}(\textit{User},\textit{Book}),\textit{bookAuthor% }(\textit{Book},\textit{stanlee}).$ $\displaystyle 0.19461::\textit{likeMovie}(\textit{User},\textit{Movie})\text{ % :- }\textit{likeBook}(\textit{User},\textit{Book})$ (4) $\displaystyle\textit{bookGenre}(\textit{Book},\textit{science}),\textit{% movieGenre}(\textit{Movie},\textit{biography}).$ $\displaystyle 0.21311::\textit{likeBook}(\textit{User},\textit{Book})\text{ :-% }\textit{likeMovie}(\textit{User},\textit{Movie}),$ (5) $\displaystyle\textit{bookGenre}(\textit{Book},\textit{computer}),\textit{% movieGenre}(\textit{Movie},\textit{sci-fi}).$
3.
The role of source or target domain needs not be pre-defined, due to the nature of relational learning, unlike a typical cross-domain RS which the source and target domains must be pre-defined. However, our framework allows a user to specify the particular target and source domains if he/she desires in the setting at the beginning of the process.
4.
A new domain can be added into our framework without requiring a re-training, unlike a typical cross-domain RS which re-training is required for every domain pairs. Since number of domains used in our framework are not limited to one or two, any new domain can be added by simply specifying the new domain’s predicates and changes in the setting.
5.
In our framework, all domains can be processed simultaneously, unlike the typical cross-domain RS which only two domains can be processed each time.
6.
Our framework can derive negative recommendation rules automatically since we adopt ProbFOIL to be used in learning the rule, unlike the typical RS. An introduction to ProbFOIL is briefly summarized in the next section, Section 3.1. The negative rules can ensure that the item will not be included in the list and can generate negative examples for cold-start problem.

3.1 ProbFOIL and relational learning

In this research, we adopt the probabilistic first order rule learner (ProbFOIL) algorithm to learn the recommendation rule. ProbFoil was first introduced in 2010 by De Raedt and Thon [18] and was upgraded in 2015 [17]. It is a probabilistic extension of the traditional rule-learning system FOIL (First-Order Inductive Learner) that is capable of learning probabilistic rules from probabilistic data.

ProbFOIL supports the probabilistic data by generalizing the concept of true/false, positive/negative to a probabilistic context [15]. Moreover, it performs an additional step of parameter learning to learn rules that express probabilistic relationships. The learning problem considered is defined as follows [17]:

Given:

1.
A set of training examples $E$ , consisting of pairs $(x_{i},p_{i})$ , where $x_{i}$ is a ground fact for the unknown target predicate $t$ and $p_{i}$ is a target probability;
2.
A background theory $B$ containing information about the examples in the form of a ProbLog [16] program;
3.
A loss function $\textit{loss}(H,B,E)$ , measuring the loss of a hypothesis (set of clauses) $H$ w.r.t. $B$ and $E$ ;
4.
A space of possible clauses $\mathfrak{L}_{h}$ specified using a declarative bias;

Find:A hypothesis $H\subseteq\mathfrak{L}_{h}$ such that $H=\text{argmin}_{H}\textit{loss}(H,B,E)$ and the loss function aims at minimizing the standard error of the predictions, that is,

$\displaystyle\textit{loss}(H,B,E)=\sum_{(x_{i},p_{i})\in E}|P_{S}(B\cup H% \models x_{i})-p_{i}|$ (6)

Algorithm 3.1 [17] presents ProbFOIL algorithm of which the probabilistic rules are in the form $x::\textit{target}\leftarrow\textit{body}$ . ProbFOIL is initialized with the target predicate in the head and it greedily adds body literals and clauses until no more improvement is observed with respect to a global scoring function. The clause to be added is obtained by the function LearnRule, which greedily searches for the clause with the highest local scoring function, using the refinement operator $\rho$ .

[htbp] The ProbFOIL learning algorithm[1] ProbFOILtarget $H:=\emptyset$ true $\textit{clause}:=\Call{LearnRule}{H,\textit{target}}$ $\Call{GScore}{H}<\Call{GScore}{H\cup\{\textit{clause}\}}$ $H:=H\cup\{\textit{clause}\}$ $H$ LearnRule $H,\textit{target}$ $\textit{candidates}:=\{x::\textit{target}\leftarrow\textit{true}\}$ $\textit{best}:=(x::\textit{target}\leftarrow\textit{true})$ $\textit{candidates}\neq\emptyset$ $\textit{next\_cand}:=\emptyset$ $x::\textit{target}\leftarrow\textit{body}\in\textit{candidates}$ $\textit{refinement}\in\rho(\textit{target}\leftarrow\textit{body})$ not Reject $H,\textit{best},x::\textit{target}\leftarrow\textit{body}$ $\textit{next\_cand}:=\textit{next\_cand}\cup\{x::\textit{target}\leftarrow% \textit{body}\land\textit{refinement}\}$ LScore $H,x::\textit{target}\leftarrow\textit{body}\land\textit{refinement}$ $>$

LScore $H,\textit{best}$ $\textit{best}:=(x::\textit{target}\leftarrow\textit{body}\land\textit{% refinement})$ $\textit{candidates}:=\textit{next\_cand}$ best

The probability $x$ will be determined by the rule learning algorithm. Each call to LScore returns the best score that can be achieved for any value of $x$ . The best rule found will be returned with $x$ which is a probability that yields the highest local score. Equation (3.1) is a sample recommendation rule learned by ProbFOIL:

$\displaystyle 0.86735::\textit{likeMusic}(A,B)\text{ :- }\textit{musicGenre}(B% ,\textit{country}),$ (7) $\displaystyle\textit{likeMovie}(A,\textit{Var}\_1),\textit{movieGenre}(\textit% {Var}\_1,\textit{western}).$

This rule states that person $A$ will like music $B$ with a probability of 0.86735 if he likes the western movie and $B$ is a country music. In other words, the country music $B$ will be recommended to the user with a probability of 0.86735 because of his preference on a western movie. With this rule, it is clear why music $B$ is recommended to the user.

Figure 1.
An example of one reviewer in Amazon product data format.

Figure 2.
A probabilistic logic representation of data in Fig. 1.

3.2 Data preparation

In this section, how the dataset is prepared to be used with ProbFOIL is explained. The dataset for RS typically consists of users’ information and preferences in the form of attributes and ratings supplied by users. Figure 1 is a portion of an example of Amazon product data. However, this data cannot be directly submitted to ProbFOIL since the input data of ProbFOIL is represented in first order logic format and contains the probabilistic facts. Therefore, before the learning can be performed, the dataset must be converted. The attribute will be turned into predicate and the rating provided in the dataset will be considered as a probability. In the movie domain for example, the predicate likeMovie/2 refers to the movie that the user likes. The constants i.e. drama, action, kids and the predicate movieGenre/2 refer to the genre of movie. The rating 0–5 is associated to probability 0–1 to be used in probabilistic logic representation. The example of the Amazon data in Fig. 1 could be represented with a probabilistic logic formula for probFOIL as shown in Fig. 2.

Figure 3.

Our proposed recommendation framework overview.

Figure 4.

The example of ProbFOIL settings.

3.3 Recommendation relational rule learning

We design our framework for RS with all features previously described in Section 3. The overview of our framework is depicted in Fig. 3. ProbFOIL learns the recommendation rules from data and settings. Data is input as facts in Prolog format. The datasets which will be described in the next section are preprocessed and converted into Prolog format using Problog library loader provided in ProbFOIL package. The settings define target predicate, modes, type information for the predicates and others which relate to the data. Target predicate is the predicate that we want to learn while mode refers to the predicates that can be added to the rules in the condition part. The specified predicates are from all interested domains. The example settings are shown in Fig. 4. In this setting, a definition/rule for learn(likeMusic/2) is induced. The predicates from movie and music domains; likeMovie, movieGenre and musicGenre and their type information are specified. The rules learned by ProbFOIL are represented in FOIL as shown in Fig. 5. The first rule states that if the user likes the western movie, then there is 86.735% chance that the user will like the country music. The second rule indicates that the users who like the same genre of movie will also like the same music and the last rule indicates that the users who like two different genres of movie will also like love music. The output learned rules will be ranked based on their probabilities and be used to compute the recommendation item. The top-N items in the list will be recommended to a particular user.

Figure 5.

The portions of learned rules likeMusic/2.

4. Experiments and results

The purpose of this experiment is to answer the two research questions discussed in Section 3. More specifically, we would like to demonstrate that our proposed framework provides more choices of recommendations to users accompanying by understandable explanations. We conduct the experiment on three different domains i.e. music, movie and book, which are selected as representative examples. Given background knowledge, positive example facts, and negative example facts, we use ProbFOIL to learn the recommendation rules. We compare the performance of our algorithms to the three well-known baseline methods using HitRatio@N [20] and SRDP@N [21] which are the quality of the top-N recommendation measurement and the usefulness of unexpectedness measurement.

4.1 Dataset

We use “Amazon product dataset” provided by UCSD [11, 30] to carry out our experiments in this research. The dataset contains product reviews, metadata and links from “Amazon”, including 142.8 million reviews spanning 1996 to 2014. Product reviews include ratings, text and helpfulness votes. Product metadata are descriptions, category information, price, brand, image features. Links are “also viewed” and “also bought graphs”.

It is not unusual to find lots of missing values in the user-item preference (rating) matrix $R$ and the sparsity of $R$ is greater than 90% in “Amazon product dataset”. However, it is true task of the recommender systems to predict ratings for all missing values (unrated items) [1]. There may also be some noise values from the reviewers that users might rate an item highly but they do not like and write negative textual review for that item (e.g. The reviews of “The World Rose” as shown in Fig. 6). In addition, it is noted that the dataset from Amazon and other e-commerce websites may contain fake reviews generated by marketing companies to mislead shoppers and boost their business. The fake review issue is considered as the noise values. To overcome this problem, domain knowledge or domain experts play an important role in the data preprocessing stage to analyze and spot noise reviews.

Figure 6.

The content of top positive review (5-star rating and people found it helpful) for “The World Rose” in Amazon is also negative.

Preprocessing of this dataset includes the following steps: First, create the user preference dataset by extracting the top 1,000 reviewers from product review dataset for each domain, music, movie and book. Second, create the item attribute dataset using Amazon’s Product Advertising API, Last.fm’s API,1

https://www.last.fm/api/.

and OMDb API2

http://www.omdbapi.com/.

to retrieve the attributes of each item listed in the user preference dataset. Third, create the domain-pair dataset by taking the top 1,000 reviewers from one category (domain) in user preference dataset. Then, extract their preference of the other category. For example, the top 1,000 reviewers from music category are taken and their preferences on movie are extracted to form the music-movie dataset.

4.2 Recommendation rule and evaluation

To run the experiment on the three selected domains: music, movie and book, the setting in ProbFOIL is properly specified for each domain-pair. Examples of recommendation rules learned by ProbFOIL are presented in Fig. 7.

Figure 7.

The unexpected recommendation rules for likesMusic/2.

Each rule is represented in relational logic associating with probability. The rule has a condition part (the right-hand side of the rule) and a conclusion part (the left-hand side of the rule). The first rule can be read as the following: If user1 and user2 like the same movie genre thriller, then user2 will like the same music as user1 does with a probability of 0.69523. The second rule can be read as: If user likes the film Twilight then he/she will like the music performed/composed by an artist who started his career in 1996 with a probability of 0.14434. These rules are declarative, that is, easy to understand to human. In addition, this formal language is unambiguous, less redundant and more concise compared to natural language used in other explanation recommendation systems. The experiment results confirm the answer to the first research question: each rule (in relational learning) states clearly why the items are recommended to the particular user. Therefore, it can be concluded that the generated rules provide explanation facilities for our recommender system.

These rules also demonstrate that user’s preference on one domain can be used to predict user’s preference on another domain, for example, user’s movie taste can be used to predict his/her music taste.

The first and second rules recommend different music to the user based on his/her interest in movie genre. The third and the fourth rules recommend particular music to the user due to his/her preference on different domain objects. Ed Sheeran’s music will be recommended to the user whose preference is Harry Potter book while Aerosmith’s music will be recommended to the user who likes Sci-Fi movie. These rules return the recommendation items which cannot normally be found in similarities-based recommended system. These experiment results provide us with the solution to the second research question: the user retrieves the recommendation item which is new and he/she is not familiar with.

We also conducted an experiment on providing the recommendations in one domain using information from other different domains: Music, Movie, and Book. Examples of recommendation rules learned by our framework are presented in Fig. 8. “Back in Black” song will be recommended to the user whose preferences are Stan Lee’s book and movies from Marvel Studios while “Forrest Gump” movie will be recommended to the user who likes “I Really Like You” song and “Catch Me If You Can” book.

Figure 8.

Examples of music recommendation rules.

In this research, we evaluate the performance of our proposed framework in two aspects: the quality of the top-N recommendations (HitRatio@N) and the usefulness of unexpectedness recommendations (SRDP@N).

Deshpande and Karypis [20] suggested that the quality of the top-N recommendations can be measured by the number of hits and their position within the top-N items. For each user $u$ belonging to $U$ , a set of target users, we have arbitrarily chosen item $i$ such that it is tagged by user $u$ previously. If $i$ is in the top-N recommendation list, then hit occurs. The HitRatio@N is defined as a total number of hits divided by a total number of users as follows:

$\displaystyle\textit{HitRatio@N}=\frac{\textit{Number of Hits}}{|U|}$ (8)

Ge and his colleagues [21] suggested that the serendipitous recommendation can be accurately and precisely measured by considering the two essential aspects of serendipity: unexpectedness and usefulness. The serendipity (SRDP) is then defined as follows:

$\displaystyle\textit{SRDP}(u)=\frac{|\textit{UNEXP}(u)\cap\textit{USEFUL}(u)|}% {|\textit{UNEXP}(u)|}$ (9)

Where the unexpected (UNEXP) set contains elements of $R S$ which do not appear in $P M$ , denoted by:

$\displaystyle\textit{UNEXP}=RS\setminus PM$ (10)

$R S$ is a set of recommendations generated by our framework while $P M$ is a set of recommendations generated by a primitive prediction model.

USEFUL is a set of useful items which is determined by user. For instance, the usefulness of an item can be approximated by the user’s feedback. To determine the serendipity of our framework, we consider the average serendipity of all users denoted by SRDP@N as follows:

$\displaystyle\textit{SRDP@N}=\frac{1}{|U|}\sum_{u\in U}\textit{SRDP}(u)$ (11)

4.3 Evaluation results

In this section, our framework is compared to the four following baseline methods in terms of HitRatio@N and SRDP@N:

•
UPCC (User-based PCC): the user-based collaborative filtering algorithm using Pearson correlation coefficient.
•
IPCC (Item-based PCC): the item-based collaborative filtering algorithm using Pearson correlation coefficient.
•
FUSE: the cross-domain recommendation which was proposed in Chen et al. [31].
•
CIT: the cross-domain RS with consistent information transfer which was proposed in Zhang et al. [28].

The four baseline methods directly provide the top-N recommendation items for a particular user. To compare with these four baselines, our generated recommendation rules with their probability were used to find the top-N items. Partitioning technique used was hold-out setting. In this setting, test reviews were sampled and hidden from our dataset without partitioning the users (Fig. 9).

Tables 2 and 3 show the HitRatio@5 and HitRatio@10 for two domain-pairs, Music-Movie and Book-Movie respectively. The quality of the top-5 and top-10 recommendations of our framework is moderate, however, it produced recommendations not found in the primitive single-domain based system. Such recommendations were not expected by the users and some of them were found interesting as shown by SRDP@10 in Table 4. No unexpected recommendation (if any) by UPCC and IPCC on both domain-pairs was found useful to the users.

Table 2
Comparison results for top-N recommendations (@N) on Music-Movie data

UPCC IPCC FUSE CIT Our framework

HitRatio@5 0.34 0.23 0.29 0.33 0.32

HitRatio@10 0.39 0.30 0.48 0.50 0.53

Table 3
Comparison results for top-N recommendations (@N) on Book-Movie data

UPCC IPCC FUSE CIT Our framework

HitRatio@5 0.39 0.27 0.32 0.41 0.38

HitRatio@10 0.41 0.34 0.53 0.56 0.57

Table 4
SRDP@10

Dataset UPCC IPCC FUSE CIT Our framework (single) Our framework

Music-Movie 0 0 0.15 0.16 0 0.21

Book-Movie 0 0 0.10 0.12 0 0.14

Figure 9.
Hold-out data partitioning technique.

Our framework was also implemented on a single domain using only music dataset and there was no serendipitous recommendation found (SRDP@10 is 0 in Table 4). Although unexpected recommendations were suggested, $|\textit{UNEXP}|$ was 2.25 on average out of top 10 recommendations. The output rules basically state that if the user likes music “A”, then there is P% chance that the user will like music “B”. Some music “B”, e.g., “Take Five” and “You Never Know Who Your Friends Are” (Fig. 10), are considered as unexpected to the users since they are not found in a set of recommendations generated by a primitive prediction model. However, since only user preferences are available on the single domain, it is inadequate to predict useful unexpected results which thus make the recommendations unattractive to the users. Hence, this experiment confirms that using user’s preference on one domain to predict user’s preference on another domain provides more chance in recommending unknown items that are interesting to the user.

Figure 10.
The unexpected recommendation rules for likesMusic/2 (single domain).

5. Discussion

	UPCC	IPCC	FUSE	CIT	Our framework
HitRatio@5	0.34	0.23	0.29	0.33	0.32
HitRatio@10	0.39	0.30	0.48	0.50	0.53

	UPCC	IPCC	FUSE	CIT	Our framework
HitRatio@5	0.39	0.27	0.32	0.41	0.38
HitRatio@10	0.41	0.34	0.53	0.56	0.57

Dataset	UPCC	IPCC	FUSE	CIT	Our framework (single)	Our framework
Music-Movie	0	0	0.15	0.16	0	0.21
Book-Movie	0	0	0.10	0.12	0	0.14

We began our research by affirming that relational learning and multiple domain could be combined to the benefit of RS. Relational learning provides us with the ability to make a prediction of possible new relations between users and items. These relations are presented in the form of an if-then statement. Incorporating information or knowledge from various or multiple domains leads to a broader and wider set of recommendations including serendipitous items. The experiment carried out in this research illustrates that our framework generates rules with explanation facilities for our recommender system and some surprising relations found lead to the unexpected but useful recommended items. The experiment results show that our proposed algorithm is very promising, however, there are a few issues worth to be discussed in the following sub-sections.

5.1 Explainability

In addition to simplicity and transparency, effectiveness and efficiency also affect user satisfaction [5]. Thus, an evaluation of the effectiveness and the efficiency with four hypotheses should be conducted to ensure the outcomes as shown in Fig. 11.

Figure 11.

The hypotheses of the effectiveness and the efficiency evaluation.

However, it is surprising that even nowadays, the most common way of evaluating explanations is by means of real user studies and acceptance testing [7]. The offline experiment of a new form of explanation is still lacking a proper evaluation. Developing a reliable and easily usable evaluation will save a lot of efforts for offline evaluation of explainable recommendation systems. A further study on how to evaluate the explainable RS based on historical data (offline experiment) should be considered.

5.2 Novel and unexpected recommendation coverage

To provide more coverage on the recommendation items, we may consider merging our results with the recommendations found by traditional methods. For example, merge our music recommendation result with the result from traditional method on users’ music preference only (single domain) shown as the highlighted section in Fig. 12.

Figure 12.

The results contain unexpected and useful recommendations – shown as the relative complement of A in B (the colored section) – not found in single-domain based system.

The degree of “novelty” can be controlled or increased by incorporating more available user personal property and unexpected item attributes. Some examples of incorporating unexpected data and the output are as follows:

•

We can discover music through samples, cover songs and remixes connections (using data from WhoSampled).3

https://www.whosampled.com/.

Figure 13 shows an unexpected connection between Busta Rhymes and Bernard Herrmann.

Figure 13.

An example of unexpected connection via item attributes.

•

Figure 14 shows examples of association via a user personal property. On the left, the user and “The cup of life” are linked since the user was a soccer fan and the song was a theme song for 1998 FIFA World Cup. On the right, the user is linked to “Candle in the wind” because the user was an admirer of Princess Diana and the song was performed at Princess Diana’s funeral.

Figure 14.

The examples of unexpected connection via user attributes.

5.3 Extendibility

Another important feature of our framework is extendibility. New domains (e.g., TV show, game, sport, and beauty) can be incorporated directly. As mentioned earlier, adding a new domain into our system does not require re-training, only predicate setting is simply specified. New data (e.g., user/item feature and feature from review text) and different contexts (e.g., location, time, and mood) can also be incorporated as different domains. We believe that adding more useful features and contextual knowledge (e.g., user’s activities collected from wearable devices) to the framework can provide a more specific and useful recommendation to the particular user. Example recommendation rule Eq. (5.3) is the output learned rule that is expected (not from the actual experiment) if the contextual knowledge “user is sleeping” is known to the framework. According to Eq. (5.3), music2 is recommended to the user because he/she is sleeping and his/her preference is music1. The sleeping context is considered in suggesting the music.

$\displaystyle 0.80000::\textit{likeMusic}(\textit{User},\textit{music2})\text{% :- }$

(12) $\displaystyle\textit{likeMusic}(\textit{User},\textit{music1}),\textit{% userActivity}(\textit{User},\textit{sleeping}).$
5.4 Scalability

Most ILP methods including ProbFOIL do not scale to very large datasets as the time complexity of inference using grounding is exponential over the size of the dataset [17]. Hence, the execution time of our proposed framework is slower than the baseline methods. However, either a scalability minimization or finding a scalable probabilistic rule learner is a challenge because incorporating additional domains, data, or context can create a scalability problem. Nonetheless, in this research we accept it as a trade-off for a more coverage set of recommendations. A further study should be conducted to address the scalability issue, and it is considered as our future work.

5.5 Cold-start problem

As mentioned earlier, if we can get the recommendation process out of the black box, we will be able to solve some problems. Here, for example, the cold-start problem can simply be handled by generating the preferences for new users using our learned recommendation rules. The item recommendation module integrates the new user generated preferences with existing user preferences dataset, thus the framework has a better ability to deal with the cold-start problem and provides explainable recommendations.

5.6 Negative recommendations

The negative recommendations refer to the items that the system will not suggest to the user. A generated rule can be transformed into an equivalent negative form of which the conclusion part is negative. The expected rule is shown as follows:

$\displaystyle 0.80000::\textit{not}(\textit{likeMusic}(\textit{User},\textit{% music1}))\text{ :- }\textit{not}(\textit{userMood}(\textit{User},\textit{good}% )).$ (13)

The music1 will not be recommended to the user because the user is not feeling good at that moment. The explanation from the rule might let the user know why not to try out or to purchase a certain item, the system can help save time for the user and improve its trustworthiness.

The negative recommendation rule also specifies that the users with particular preferences will not like particular items. These found (by the system) disliked items are located in the tail of the popularity distribution and not previously rated by a user. Thus, all disliked items must be discarded from the unexpected recommendation list to ensure that the un-useful results are excluded, resulting in increasing the serendipity measurement SRDP@N. Therefore, negative rules and disliked items could be used to enhance the system performance.

To our knowledge, research in negative recommendation has not yet been available. A further experimentation with negative rules may be necessary to confirm the capabilities of our approach.

6. Conclusions

We amalgamate relational learning with multi-domain to form an alternative recommendation system framework. With relational learning, our framework constructs general recommendation rules with probabilities. The recommendation items will be made for the users in any domain led by the rule. Each recommendation is accompanied with explanations. We focus on a simple, understandable, precise and useful explanation. The explanation, therefore, describes why the items are chosen for the users in an if-then logical format.

Our framework aggregates all information and knowledge from multiple domains to broaden the search space, thereby significantly increase a chance to discover items which are previously unseen or unexpected to a user resulting in a wide range of recommendations. The explanation and its associated probability help persuade users to try out or purchase the new or unexpected recommended item faster with more confidence.

Furthermore, our recommendation rule can be used to recommend items to a new user who has no preference history provided; to recommend the items which are new to the system; and not to recommend the items that the users dislike. Our framework is also extendible in that additional domains, new data (e.g., features), and context (e.g., time) can easily be incorporated. The quality and degree of novelty can also be adjusted by the numbers of domains used or by incorporating more user personal property and unexpected item attributes available.

Lastly, this research can proceed in many ways, for example, experimenting on context-aware and negative recommendation, performing both offline and online evaluation on explainability, and improving the accuracy of our framework by using deep learning (embedding-based recommendation model) [32] to further enhance the system performance.

Footnotes

Acknowledgments

This work was supported by the JSPS Core-to-Core Program, A. Advanced Research Networks and the Brain Circulation project.

References

Marlin

B.M.

Zemel

R.S.

Roweis

S.T.

and Slaney

, Recommender systems: missing data and statistical model estimation, in: Proceedings of the Twenty-Second International Joint Conference on Artificial Intelligence, 2011, pp. 2686–2691. doi: 10.5591/978-1-57735-516-8/IJCAI11-447.

Billsus

and Pazzani

M.J.

, A personal news agent that talks, learns and explains, in: Proceedings of the Third Annual Conference on Autonomous Agents, AGENTS ’99, ACM, New York, NY, USA, 1999, pp. 268–275. doi: 10.1145/301136.301208.

Cosley

Lam

S.K.

Albert

Konstan

J.A.

and Riedl

, Is seeing believing? How recommender system interfaces affect users’ opinions, in: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI ’03, ACM, New York, NY, USA, 2003, pp. 585–592. doi: 10.1145/642611.642713.

McSherry

, Explanation in recommender systems, Artif. Intell. Rev. 24(2) (2005), 179–197. doi: 10.1007/s10462-005-4612-x.

Gedikli

Jannach

and Ge

, How should i explain? a comparison of different explanation types for recommender systems, International Journal of Human-Computer Studies 72(4) (2014), 367–382. doi: 10.1016/j.ijhcs.2013.12.007.

Cantador

Fernández-Tobías

Berkovsky

and Cremonesi

, Cross-Domain Recommender Systems, Springer US, Boston, MA, 2015, 919–959. doi: 10.1007/978-1-4899-7637-6_27.

Nunes

and Jannach

, A systematic review and taxonomy of explanations in decision support and recommender systems, User Modeling and User-Adapted Interaction 27(3) (2017), 393–444. doi: 10.1007/s11257-017-9195-0.

Konstan

J.A.

McNee

S.M.

Ziegler

C.-N.

Torres

Kapoor

and Riedl

J.T.

, Lessons on applying automated recommender systems to information-seeking tasks, in: Proceedings of the 21st National Conference on Artificial Intelligence – Volume 2, AAAI’06, AAAI Press, 2006, pp. 1630–1633. http://dl.acm.org/citation.cfm?id=1597348.1597458.

Hoxha

and Rettinger

, First-order probabilistic model for hybrid recommendations, in: 12th International Conference on Machine Learning and Applications, Vol. 2, 2013, pp. 133–139. doi: 10.1109/ICMLA.2013.118.

10.

Herlocker

J.L.

Konstan

J.A.

and Riedl

, Explaining collaborative filtering recommendations, in: Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work, CSCW ’00, ACM, New York, NY, USA, 2000, pp. 241–250. doi: 10.1145/358916.358995.

11.

McAuley

Targett

Shi

and van den Hengel

, Image-based recommendations on styles and substitutes, in: Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’15, ACM, New York, NY, USA, 2015, pp. 43–52. doi: 10.1145/2766462.2767755.

12.

Picado

Termehchy

Fern

and Ataei

, Schema independent relational learning, in: Proceedings of the 2017 ACM International Conference on Management of Data, SIGMOD ’17, ACM, New York, NY, USA, 2017, pp. 929–944. doi: 10.1145/3035918.3035923.

13.

Struyf

and Blockeel

, Relational Learning, Springer US, Boston, MA, 2010, pp. 851–857. doi: 10.1007/978-0-387-30164-8_719.

14.

McCarthy

Reilly

McGinty

and Smyth

, Thinking positively-explanatory feedback for conversational recommender systems, in: Proceedings of the European Conference on Case-Based Reasoning (ECCBR-04) Explanation Workshop, 2004, pp. 115–124.

15.

Antanas

Dries

Moreno

and De Raedt

, Relational affordance learning for task-dependent robot grasping, in: Lachiche

Vrain

(Eds.), Inductive Logic Programming, Springer International Publishing, Cham, 2018, pp. 1–15.

16.

De Raedt

Kimmig

and Toivonen

, Problog: A probabilistic prolog and its application in link discovery, in: Proceedings of the 20th International Joint Conference on Artifical Intelligence, IJCAI’07, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 2007, pp. 2468–2473. http://dl.acm.org/citation.cfm?id=1625275.1625673.

17.

De Raedt

Dries

Thon

Van Den Broeck

and Verbeke

, Inducing probabilistic relational rules from probabilistic examples, in: Proceedings of the 24th International Conference on Artificial Intelligence, IJCAI’15, AAAI Press, 2015, pp. 1835–1843. http://dl.acm.org/citation.cfm?id=2832415.2832504.

18.

De Raedt

and Thon

, Probabilistic rule learning, in: Proceedings of the 20th International Conference on Inductive Logic Programming, ILP’10, Springer-Verlag, Berlin, Heidelberg, 2011, pp. 47–58. http://dl.acm.org/citation.cfm?id=2022735.2022746.

19.

Bilgic

and Mooney

R.J.

, Explaining recommendations: Satisfaction vs. promotion, in: Beyond Personalization Workshop, IUI, Vol. 5, 2005, p. 153.

20.

Deshpande

and Karypis

, Item-based top-n recommendation algorithms, ACM Trans. Inf. Syst. 22(1) (2004), 143–177. doi: 10.1145/963770.963776.

21.

Delgado-Battenfeld

and Jannach

, Beyond accuracy: Evaluating recommender systems by coverage and serendipity, in: Proceedings of the Fourth ACM Conference on Recommender Systems, RecSys ’10, ACM, New York, NY, USA, 2010, pp. 257–260. doi: 10.1145/1864708.1864761.

22.

Tintarev

and Masthoff

, A survey of explanations in recommender systems, in: Proceedings of the 2007 IEEE 23rd International Conference on Data Engineering Workshop, ICDEW ’07, IEEE Computer Society, Washington, DC, USA, 2007, pp. 801–810. doi: 10.1109/ICDEW.2007.4401070.

23.

Tintarev

and Masthoff

, Explaining Recommendations: Design and Evaluation, Springer US, Boston, MA, 2015, 353–382. doi: 10.1007/978-1-4899-7637-6_10.

24.

Celma

Ò.

, Music recommendation and discovery in the long tail, Ph.D. thesis, Universitat Pompeu Fabra, Barcelona, 2008.

25.

Adamopoulos

and Tuzhilin

, On unexpectedness in recommender systems: Or how to better expect the unexpected, ACM Trans. Intell. Syst. Technol. 5(4) (2014), 54:1–54:32. doi: 10.1145/2559952.

26.

Kouki

Fakhraei

Foulds

Eirinaki

and Getoor

, Hyper: A flexible and extensible probabilistic framework for hybrid recommender systems, in: Proceedings of the 9th ACM Conference on Recommender Systems, RecSys ’15, ACM, New York, NY, USA, 2015, pp. 99–106. doi: 10.1145/2792838.2800175.

27.

Wärnestål

, User evaluation of a conversational recommender system, in: Proceedings of the 4th IJCAI Workshop on Knowledge and Reasoning in Practical Dialogue Systems, 2005, pp. 32–39.

28.

Zhang

Liu

and Zhang

, A cross-domain recommender system with consistent information transfer, Decision Support Systems 104 (2017), 49–63. doi: 10.1016/j.dss.2017.10.002.

29.

Catherine

and Cohen

, Personalized recommendations using knowledge graphs: A probabilistic logic programming approach, in: Proceedings of the 10th ACM Conference on Recommender Systems, RecSys ’16, ACM, New York, NY, USA, 2016, pp. 325–332. doi: 10.1145/2959100.2959131.

30.

and McAuley

, Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering, in: Proceedings of the 25th International Conference on World Wide Web, WWW ’16, International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, Switzerland, 2016, pp. 507–517. doi: 10.1145/2872427.2883037.

31.

Chen

Hsu

and Lee

M.L.

, Making recommendations from multiple domains, in: Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’13, ACM, New York, NY, USA, 2013, pp. 892–900. doi: 10.1145/2487575.2487638.

32.

Wang

Cao

Liu

and Chua

T.-S.

, Kgat: Knowledge graph attention network for recommendation, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, ACM, New York, NY, USA, 2019, pp. 950–958. doi: 10.1145/3292500.3330989.

33.

Gantner

Drumond

Freudenthaler

Rendle

and Schmidt-Thieme

, Learning attribute-to-feature mappings for cold-start recommendations, in: 2010 IEEE International Conference on Data Mining, 2010, pp. 176–185. doi: 10.1109/ICDM.2010.129.

Explainable and unexpectable recommendations using relational learning on multiple domains

Abstract

Keywords

1. Introduction

2. Related work

2.1 Explanations in RS

Table 1 Examples of explanations in commercial and academic systems

3. Research methodology

4.1 Dataset

5.1 Explainability

(12) 𝑙𝑖𝑘𝑒𝑀𝑢𝑠𝑖𝑐 ⁢ ( 𝑈𝑠𝑒𝑟 , music1 ) , 𝑢𝑠𝑒𝑟𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦 ⁢ ( 𝑈𝑠𝑒𝑟 , 𝑠𝑙𝑒𝑒𝑝𝑖𝑛𝑔 ) . 5.4 Scalability

5.5 Cold-start problem

5.6 Negative recommendations

Footnotes

Acknowledgments

References

Table 1
Examples of explanations in commercial and academic systems

(12) $\displaystyle\textit{likeMusic}(\textit{User},\textit{music1}),\textit{% userActivity}(\textit{User},\textit{sleeping}).$
5.4 Scalability