Recommender systems are software tools and techniques that offer personalized suggestions for items a user may like to purchase or view based on such user’s interests and preferences (Ricci et al., 2011) and (Burke, 2007). According to (Ricci et al., 2011), recommender systems exist in domains with larger sets of items where the users are interested in just a small portion of these items. The suggestions could be books to buy, music or news to listen to, financial products to get, etc. (HERLOCKER et al., 2004) defined tasks that a Recommender System can assist a user in implementing as follows:
- Recommend good items as a ranked list with predictions on how much the user would like them.
- Find all items that can match the users’ preferences.
- Annotate or emphasize the list of items worth buying depending on the user’s preference.
- Recommend a sequence of items rather than a single recommendation.
- Recommend a bundle of items that the customer will like.
- Browse items that may fall within the scope of the user’s preferences.
- Incorporate functions that will assist potential users in ascertaining whether the system is credible.
- Improve the profile where the user can provide explicit information about their likes or dislikes.
- Express oneself so users can give feedback even if they don’t use the recommendation.
- Help others by providing evaluations or ratings on products so the community will benefit from the contribution.
- Influence others in buying or penalizing certain items.
A recommender system without explanations has limitations such as the inability of users to justify reasons for the recommendation and the black box problem, which is the difficulty of the system to explain how it arrived at a recommendation resulting in a system with little or no transparency. Limited transparency can lead to increased vulnerabilities and flaws in the system. Recommender systems with explanations or Explainable Recommender systems are being built to address these limitations. Explainable Recommender Systems are systems that offer both personalized recommendations results and clarification on why items were suggested (Y. Zhang & Chen, 2020)
2.1 THE NEED AND OBJECTIVES OF EXPLAINABLE RECOMMEDER SYSTEMS
2.1.1 THE NEED FOR EXPLAINABLE RECOMMEDER SYSTEMS
Explainable Artificial Intelligence (XAI)- a field in Artificial Intelligence (AI) that promotes the improvement of trust and transparency of AI systems is needed for users to understand, trust and manage AI systems. The Explainable Recommender Systems (XRS) is a subset of Explainable Artificial Intelligence (AI). (Adadi & Berrada, 2018) proposed reasons for the need for explainability as follows:
- Justification to clarify why a user made recommendations rather than a description of the inner working of the system.
- Control to detect and correct any errors, vulnerabilities, and flaws that may have been detected.
- Improvement to enable users to know how the systems can produce better results.
- Discovery, to gain knowledge and insights on suggestions made.
These needs being met help build trustworthy and transparent Explainable Recommender Systems.
2.1.2 THE OBJECTIVES OF EXPLAINABLE RECOMMENDER SYSTEMS
The objectives of recommender systems as identified by (Tintarev, 2009), (Tintarev, 2007), (Jannach & Friedrich, 2013), (Detmar Jannach et al., 2011) are transparency, validity, scrutability, trustworthiness, persuasiveness, effectiveness, efficiency, satisfaction, relevance, comprehensibility, and education. It could also be referred to as explanation attributes, as shown in Figure 2.
(i) Transparency: Explanations provide the reason behind a recommendation, such as “Customer who bought item A also bought item B.” (Tintarev, 2009). Transparency explains how a user selected a recommendation, while justification shows why a recommendation was made and is more applicable where some algorithms are not easy to explain for the protection of business secrets and flexibility in designing explanations (Vig et al.,2008).
(ii) Validity: Explanations could enable customers to validate the quality of recommendations by comparing the required and offered product features (Detmar Jannach et al.,2011).
(iii) Scrutability: Explanations could enable users to correct wrong assumptions, reasoning, or steps in the recommendation process, implying that users can tell if the system is wrong and would be able to correct the errors (Tintarev, 2009).
(iv)Trustworthiness: Explanations build users’ confidence in a system. Good explanations of recommendations can influence a user’s decision on whether to buy an item or not.
(v) Persuasiveness: Explanations may convince users to try or buy (Tintarev & Masthoff, 2007)), (Cosley et al., 2003)). Explanations may affect user evaluation positively such that a user rates a product differently from how it has been rated previously (Herlocker et al., 2000), (Tintarev, 2009)).
(vi) Effectiveness: Explanations rather than just persuading users to buy could assist them in making good decisions (Tintarev, 2009). Effectiveness is majorly dependent on the accuracy of the recommendation process because users cannot make a good buying decision if the recommendation algorithm is not accurate. Effectiveness can be measured by the significant difference between liking an item before and after consumption (Tintarev, 2009). For instance, the recommender system can be effective if a user liked a recommended Mastercard before requesting it and still expressed a similar rating after using it. Similarly, one can evaluate the effectiveness of a system by comparing users’ responses to the system with and without an explanation facility to see if the users that used the system with an explanation facility will choose items more suited to their preferences (Cramer et al., 2009).
(vii) Efficiency: Explanations could help users or customers make decisions faster, which means customers can choose a product quicker(Tintarev, 2009). This metric or criteria is usually applied in conversational recommender systems to measure the total number of interactions a user would engage in to find the relevant products(Tintarev & Masthoff, 2011).
(viii) Satisfaction Explanations could make users comfortable and more satisfied with the system. Explanations can increase user satisfaction and tolerance toward the system, but it is not a substitute for accurate recommendations (Tintarev & Masthoff, 2011). In other words, satisfaction is connected with accuracy metrics, just like other explanation criteria. Satisfaction can be measured by
directly sampling users’ opinions on their likeness to the system or indirectly by measuring user loyalty(Tintarev & Masthoff, 2007).
(ix) Relevance: Explanations can be provided to justify why additional information is needed from customers, especially in conversational recommenders (Detmar Jannach et al., 2011)
(x) Comprehensibility Recommender systems are not certain about users' knowledge; hence relate the user's known concepts to the concepts adopted by the Recommender system (Detmar Jannach et al., 2011).
(xi) Education: Explanations can teach users about the product domain to equip them with the requisite knowledge to make better purchasing decisions (Detmar Jannach et al., 2011).
The suitability of explanation criteria or objective depends on the system goal, which has to be defined at the design stage of the system (Tintarev & Masthoff, 2007), (Detmar Jannach et al., 2011)).
2.3 INFORMATION SOURCES FOR RECOMMENDATIONS AND EXPLANATIONS
An explanation is an information available to users for various purposes, such as justification on reason for an item recommendation, transparency, trust and education. The information sources for explanations are numerous and have been categorized by different authors. According to (Y. Zhang & Chen, 2020), the information sources are classified as feature-based explanations, textual sentence explanations, explanations based on relevant users or items, visual explanations, and social explanations. The information sources could be user feedback, user profile, user-generated texts, product images, social networks, and the integrating of one or more sources (hybrid).
The following subsections describe the various information sources for recommendations and explanations:
- Users’ feedbacks: Information sources on recommendations and explanations could come from explicit or implicit feedback. Explicit feedback requires the active user's involvement, possibly through ratings where the user selects their like or dislike for a product(Rich, 1979). Implicit feedback does not involve the user's direct involvement but analyzing users' activities, for example, transaction data. According to (Gemmis et al., 2015), users' implicit feedback can be generated by assigning a score to users' tasks on actions such as liking, bookmarking, saving, and so on through the possibility of bias exists.
- Feature extraction from user’s profile: According to (Kanoje et al., 2016), the user profile is used to define precisely the user. It describes the users’ characteristics or attributes that aid in proffering a personalized product which can be obtained implicitly by extracting users’ details from different social networks or monitoring users’ behaviours from transaction data. It can be gotten explicitly when the user fills out an online form. Though the explicit form is simple, the downside of it is the tendency of users filling fictitious or false information. Users’ demographics, such as age, gender, marital status, and occupation, could be obtained through any of the approaches above to form part of the user profile. A recommendation can be provided in a content-based system by matching the user profile with the item features. Explainable Recommender systems display item features and explain why it is relevant to the user.
- User Generated Texts: User-generated content generated from social media posts, reviews, tags, tweets, and comments can be an essential source of information for user profiling and recommendation (Y. Zhang & Chen, 2020), (Esparza et al., 2011)). Applying sentiment analysis techniques can extract useful information from user-generated content(Esparza et al., 2011). According to (Y. Zhang & Chen, 2020), the information source considers users' sentiments in product reviews. They further classify explanations from user-generated contexts into aspect-level and sentence-level approaches according to how these explanations are displayed. The aspect-level approach presents aspects of a product, such as a shape and colour, as explanations to users. In contrast, the sentence-level approach directly explains to users why the Recommender system chooses products.
- Images: Recommendation Systems could leverage images or a combination of audio, visual, and text(multimedia) for recommendations and explanations. A survey carried out by (Deldjoo et al., 2020) showed that a Recommender system could use multimedia content and images to suggest non-media items like clothes, foot wears, foods, and so on. Photos and images present an information source that reflects users’ interests. A user clicking images of certain types of shoes or searching search engines with images could show their interest in what the image stands for. Similarly, Explainable Recommender Systems can use images to highlight areas of users’ interests. (Y. Zhang & Chen, 2020) proposed a novel, visually Explainable Recommender model highlighting parts of fashion images that users will be interested in. The proposed model offered visual explanations of recommended items highlighting parts or regions in a personalized way that may attract the user.
- Social Networks: Apart from user-generated texts in social networks being used for recommendations, the system can use the user’s trust network comprising friends and even friends of friends to make recommendations (Rastogi & Vijendra, 2016). Social explanations could also be generated using social networks, especially when informing a user that his friend on a social network likes a certain item. The system could also combine User-generated content with trust information from social networks to make recommendations (W. Zhang et al., 2019). (Park et al., 2017) proposed a model for leveraging social network data and observed to recommend items and offer persuasive explanations for recommendations. The system can offer explanations based on the target user’s friends exhibiting similar preferences.
- User Contextual data
Recommender systems could model contextual information such as time, location, and any device for better recommendations(Adomavicius & Tuzhilin, 2015).
(Woerndl & Schlichter, 2005) proposed a gas station recommender that selects gas stations within the current fuel range along a route such that the current location of the car and the fuel level are retrieved on Network-on-wheels interfaces (infrastructure for the exchange of information between cards and a hotspot). Contextual information enriches the recommendation process since accurate predictions of customers' preferences depend on the recommendation's context (Adomavicius & Tuzhilin, 2015). The system can obtain contextual information explicitly by asking users direct questions and obtaining the information from other means. It could be obtained implicitly, such as a change in location detected by a Mobile Phone Service Provider or timestamp of a transaction(Adomavicius & Tuzhilin, 2015). (Misztal & Indurkhya, 2015) developed CARE (Context-Aware Recommender with Explanations) which provides textual explanations for the Context-Aware Recommendation that has been developed.
- Domain Knowledge: (Detmar Jannach et al., 2011) describes domain knowledge as a more detailed component of content knowledge that does not only consider the item features but the item features appropriate for the goal that the user has in mind which they called “Means-end Knowledge”. The Recommender System can deploy an ontology over the item features such that there is a representation, definition of concepts and relationship between features at a deeper relationship. Constraints can also be introduced into the Recommender System for specificity and to avoid inappropriate choice. A typical example was implemented by (Felfernig & Kiener, 2005) in recommending financial products for customers considering some properties like age, willingness to take risk(risk appetite) and so on. Domain Knowledge is usually obtained from experts in that particular domain.
- Hybrid data: This information source involves the collection of data from the combination of the above-mentioned information sources. (Putri et al., 2015) proposed a Public Facilities Recommender System using four data channels, three of which were from social media channels- Twitter, Instagram, and Foursquare and the other channel was an Internal data channel. (Hsu et al., 2007) proposed a personalized TV Recommendation System that uses demographic information and user profile data to recommend Personalized TV programs to users. The information was collected via a user profile questionnaire, and users’ viewing diary kept for a period of time. (Gao et al., 2018) developed a deep network representation learning for item recommendations using heterogeneous information sources, which are item structure, textual content, and tag information. The experimental results were better in terms of precision and Mean Reciprocal Rank (MRR) metrics when compared to baseline methods.
2.4 INFORMATION SOURCES FOR RECOMMENDATIONS AND EXPLANATIONS
(Tintarev, 2009) elaborates on the commonly adopted explanations and further explains that an explanation style may or may not follow the underlying algorithm used in the computation (see Figure III).
For instance, a content-based explanation style can be adopted for a recommender system designed with a collaborative filtering technique. The most commonly adopted explanation styles are explained in (Tintarev (2009), Tintarev & Masthoff (2011)). Burke (2002) uses the notation U to denote a set of users with known preferences, where uÎU for users whose recommendations need to be generated. I is set of items that can be recommended and iÎI is an item to predict u’s preferences. The explanation styles are as follows:
i) Collaborative-based style explanations
Recommender systems use a user u’s rating of an item in I as an input in order to find users that have similar ratings to u. The similar users, often referred to as “neighbours” are used to compute similarity (Tintarev & Masthoff, 2011). This type of explanation style could be referred to as Neighbour style Explanation (NSE) (Al-Taie, 2014). Popular usages of this type of explanation styles are the ones applied by Amazon.com, such as “Customers who bought this item also bought” (see Figure IV).
(ii) Content-based style explanation
This explanation style considers similarities between items. If a user, u, had previously rated items in I, a classifier is generated that fits u’s rating behaviour on i. A prediction of the item I will now depend on how well it fits into the classifier. An example of this explanation style is “A Mastercard was recommended for you because you already have a Visa card.” Tag style and keyword style explanations are kinds of explanation styles under this category that uses tag relevance (relationship between keywords and items) and tag preference (relationship between tags and users) to make recommendations and offer some forms of explanations [(Al-Taie, 2014); (Tintarev & Masthoff, 2011)].
(iii) Case-based reasoning style explanations
The items used to make recommendations are the ones that best fit customers’ preferences and are considered cases for comparison. A similarity between the recommended items could be computed and the similar items used as a justification for a top recommendation, as shown in a study conducted by Tintarev and Masthoff (2012). A hypothetical example could be a customer interested in loans of a certain amount and a certain number of years, putting constraints on the product attributes. A recommended set will comprise all cases, which in this instance are the product types, annual salary/income, amount, and loan tenure. The recommendation set that fulfils the constraint is presented to the customer. The explanation could state, “A salary loan advance is recommended because the tenor is short term and the amount is within the range requested.”
(iv) Knowledge and Utility-based style explanations
The input to the recommender system is the descriptions of the user, u’s needs, and preferences. The system matches these u’s needs with item I and explains recommendations based on the inference (Tintarev, 2009). Like the system described by McSherry (2005), users can specify and modify their preferences until a top recommendation arrives. An example of this explanation style is: “The mid-term loan is the best recommendation you can have although there is a difference in the recommended loan amount. This is based on your annual income.”
(v) Demographic style explanation
The input to the recommender system is demographic data about the user, u. The recommender item, i, is based on how users with similar demographic information, u have previously rated the item (Tintarev, 2009). This explanation style is least used probably due to the sensitivity of demographic data (Tintarev, 2009). An example could be: “We recommend product A because you are female between the ages of 18 and 45.”