The adaptation of the information system success model in recommender systems: the validation of dual-coding theory

Introduction

Recommender systems are filters which suggest items or information that might be interesting to users. These systems analyse the past behaviour of a user, build a user profile that stores information about user interests, and then analyse that profile to find potentially useful items (de Gemmis et al., 2015). Some of these systems provide the end user with a personalised item list and an explanation that describes why a specific item is recommended and why the system supposes that the user will like it (Gedikli et al., 2014). In recent years, these recommender systems have played an important role in reducing information overload on websites where users can vote on a series of articles or services (Bobadilla et al., 2012). For example, Netflix, Google, and Amazon utilise self-learning algorithms to keep improving their personalised recommendations to users (Alimamy and Gnoth, 2022). Collaborative filtering predicts the rating a user may give to an item based on the user’s own or other users’ rating behaviour. When other users’ behaviour is involved, a Collaborative filtering algorithm assumes the automatic collaboration of multiple users, and relies on group behaviour and preferences similar to those of that particular user (Moreno et al., 2016). Content-based filtering involves a user’s comments on items, materials read or additional information such as the user’s blogs. This content can be manually defined as needed (Cao, 2015; Lu et al, 2015).

Several survey papers on recommender systems have been published in the last few years. However, these papers focus on either recommendation techniques and approaches or a specific domain of recommender system development (Lu et al., 2015). There have been many recent studies of system effectiveness and evaluation criteria from users’ perspectives (Pu et al., 2012; Konstan and Riedl, 2012; Knijnenburg et al., 2012). Product Recommendation Agents and other web-based decision aids are deployed extensively to provide online shoppers with virtual advice services. That recommendation agents can substantially improve consumers’ decision making is well understood; but the influence of specific design attributes of the Product Recommendation Agents interface on decision making and other outcome measures are not understood so well (Xu, 2014). Konstan and Riedl (2012) reviewed the key advances in collaborative filtering recommender systems, focusing on the evolution from research that concentrates purely on algorithms to research that concentrates on the rich set of questions around the user experience with the recommender. As recommender technology is now becoming widely accepted, the ability to characterise user experience and users’ affective attitudes toward this technology has become extremely important (Pu et al., 2011).

To understand and improve user experience with recommender systems, it is necessary to conduct empirical evaluations that consider the entire process which defines the user experience (Knijnenburg et al., 2012). This can provide a better understanding of why and how certain aspects of the system result in better user experience while others do not, thereby helping to further user-centric research and the development of recommender systems (Knijnenburg et al., 2012). Pu et al. (2011) tried to understand the criteria for measuring the success of a recommender system from users' point view. And they have found that even though existing work has suggested a wide range of criteria, the consistency and validity of the combined criteria have not been tested. Research on recommender systems typically focuses on the accuracy of prediction algorithms (Knijnenburg et al., 2012; Zhu et al., 2015). However, there is a growing consensus that small changes in mean absolute error are not the path to significant improvements in user experience (Swearingen and Sinha, 2001; Konstan and Riedl, 2012).

In addition, researchers believe that good visual design is one way to improve user satisfaction. Effective design of a graphical user interface would allow more users to navigate a recommender system (Gedikli et al., 2014). Numerous popular social networking sites, such as Flickr, have found that the use of relevant visual technology can improve user satisfaction (Gedikli et al., 2014). Visualisation technology in recommender systems has now also been widely used in various e-commerce sites, such as Amazon and IMDb.com (Al-Taie and Kadry, 2014). To date, little research has been carried out to enable user interaction with such systems as a basis to support exploration and control by end users (Verbert et al., 2013). Previous studies have pointed out this research gap and suggested that a more comprehensive set of facets must be used for assessment (Gedikli et al., 2014).

Dual-coding theory has always been a very useful and complete theory in the field of human learning effectiveness. It has also been applied to cognitive phenomena in the study of educational psychology, such as problem solving, memory, language learning and concept learning (Clark and Paivio, 1991). The dual-coding theory by Paivio (1971) helps explain the differential effects of visual and verbal stimuli. This theory views cognition activities as resulting from two mental subsystems: a verbal system (processing verbal events) and an imaginal system (processing nonverbal events). These two subsystems are separate but interconnected components of human cognition. The verbal system facilitates sequential processing, while the imaginal system facilitates parallel processing. The dual coding theory suggests that pictures enhance memory for verbal (written) information because humans process written information and pictures via two independent cognitive subsystems: one devoted to verbal information and one to imagery-based information (Paivio, 2014). Dual-coding theories have been the underlying theories guiding many of the recommended practices, though they have rarely been mentioned explicitly. There has been a steady increase in attention to critical literacy since 2010 (Yang et al., 2018), such as, brand influence on consumers (Herz and Diamantopoulos, 2013), use of smart phones (Huang and Chen, 2014), influence of second screens on brands (Jensen et al, 2015), online learning (Koc-Januchta et al, 2017), and improve reading comprehension (Wang and Li, 2019). However, few research papers have applied dual-coding theory to recommender systems.

Since a recommender system can help the user to solve the problem of information overload, it is also important that the recommender system be successfully applied in this field. In the evaluation of recommender systems, it is necessary to rely on some basic assessment aspects (such as: perceived usefulness, user satisfaction, etc.) to understand whether the information system is successful. The success of the recommender system depends on whether the user is willing to continue using it. Therefore, this study intends to construct a success model for a recommender system to explore the relevance between user satisfaction and continued use intention. Also, the “dual-coding theory” argument is introduced to examine the model.

Literature review

Visual design

Display designers and content producers are currently soliciting users’ attention in a time when people have tended to develop a degree of immunity against the steady information overload (Liang et al., 2021). The main goal for designers is to find the essential features that can be used in implementing visualisation so as to attract users’ attention (Meyer, 2010). Abstracting and implementing the visual presentation of views, as products of very complex system models, is nearly as important to the effectiveness of these efforts to inform decision-making as the technical competency and completeness of those models. However, the information visualisation of data in complex system models is often considered secondary to technical considerations (Sindiy et al., 2013). Visualisation is very important in any simulation activity and it is important to use the most appropriate representation techniques to match users’ real needs (Boton et al., 2013). Existing network analysis tools, however, often lack intuitive interfaces to support the exploration of large scale data. Verbert et al. (2013) presented a visual recommender system to help guide users during the navigation of network data. Collaborative filtering, similarity metrics, and relative importance are used to generate recommendations of potentially significant nodes for users to explore. In addition, graphic layout and node visibility are adjusted in real-time to accommodate recommendation display and to reduce visual clutter.

Parra (2012) aimed to understand how different visualisations and certain personal traits might influence users’ assessment of recommended items, particularly in domains where multidimensional data or contextual constraints are involved. One study presented a human cognition framework for information visualisation. This framework emphasises how top-down cognitive processing enables the induction of insight, reasoning, and understanding, which are key goals of the visual analytics community (Patterson et al., 2014). Al-Taie and Kadry (2014) focused on the visualisation of explanations in recommender systems and learned what modalities (e.g. text, graphs, tables, and images) can better present explanations to users through the review of a selection of papers in the literature over the last few years. The results show that explanations with simple graphs and descriptions can better present explanations. Gedikli et al. (2014) found that content-based tag cloud explanations are particularly helpful to increase the user-perceived level of transparency and to increase user satisfaction, even though they demand higher cognitive effort from the user.

According to Paivio (1986), different encoding systems of information in memory are activated as a function of the type of stimuli. For example, a visual stimulus activates an imaginal coding, meaning that information is encoded in a pictorial form. On the other hand, a verbal stimulus like text will activate verbal coding, encoding information in a verbal form. Dual-coding theory further explains three distinctive levels of processing of encoding; representational, referential and associative processing. As the most basic level of processing, representational processing is the direct correspondence to the type of incoming stimuli; a visual stimulus is coded using an imaginal system, whereas a verbal stimulus is coded using a verbal system. The next level of processing, called referential processing, makes connections between the verbal and imaginal systems. This occurs when a visual stimulus is named or an image is created for a verbal stimulus. The most complex associative processing occurs when incoming visual and verbal stimuli become connected with other verbal and imaginal codes stored in memory. Dual-coding theory supports the picture superiority effect (Kim, 2019). Therefore, using two independent coding systems to help memory is more effective than using only one coding system (Paivio, 1986). Since coding, decoding, comprehension, response, information retrieval and recalling of dual-coding theory can provide rich teaching cueing, many scholars have applied dual-coding theory in the field of education (Herz and Diamantopoulos, 2013; Jensen et al, 2015; Daniels et al, 2017; Wang and Li, 2019).

DeLone and McLean’s information system success model

DeLone and McLean (1992) comprehensively reviewed information system success measures, and concluded with a model of interrelationships between six information system success variable categories: (1) system quality, (2) information quality, (3) information system use, (4) user satisfaction, (5) individual impact, and (6) organisation impact (Figure 1). This model makes two important contributions to the understanding of information system success. First, it provides a scheme for categorising the multitude of information system success measures which have been used in the research literature. Second, it suggests a model of temporal and causal interdependencies between the categories (Seddon, 1997). Based on prior studies, DeLone and McLean (2003) propose an updated model of information system success by adding a “service quality” measure as a new dimension of the information system success model, and by grouping all the “impact” measures into a single impact or benefit category called “net benefit” (Wang and Liao, 2008). This is the only construct that is ‘case specific’ and depends entirely on the type of information system (Balaban et al., 2013). Discrepancies may exist between actual use and perceived use because most information system are not used voluntarily (Torkzadeh and Doll 1994). Seddon (1997) recommends using perceived usefulness as a measure of information system performance rather than use. Users are typically required to use digital content identifier systems to comply with organisational policies. Perceived usefulness is more appropriate than use to measure system performance in mandatory use contexts and has been employed in recent research (Koo et al., 2013; Gorla and Somers, 2014; Choi and Park, 2015).

The Information System Success model has now been widely applied by many scholars (Tseng, 2015). For example, Zheng et al. (2013) integrated information system post-adoption research and the information system success model, and proposed a research framework to investigate the continued intention of virtual community users from a quality perspective. Balaban et al. (2013) developed an instrument for assessing electronic portfolio (ePortfolio) success and to build a corresponding ePortfolio success model using DeLone and McLean’s information systems success model as the theoretical framework. The impacts of information technology outsourcing on information systems’ success was assessed (Gorla and Somers, 2014). This research found significant direct and indirect effects (through the service quality) of outsourcing on information systems’ perceived usefulness and their users’ satisfaction. Choi and Park (2015) determined the structural influencing relationship between the quality of CyberAirport and the usage intention by integrating the measurement variables of the information systems success model.

The information system success model has also been applied to evaluate the effectiveness of digital learning systems, and system quality and information quality have been proven to be the main drivers of user interest in using digital learning intent and satisfaction (Mohammadi, 2015). In addition, the information system success model has been used by many studies to assess the implementation of health information technology (Bossen et al., 2013; Yu et al., 2013; Salahuddin and Ismail, 2015). Stefanovic et al. (2016) further evaluated the effectiveness of the government information system from the perspective of government employees, and found that the perceived usefulness and user satisfaction have a significant direct impact. Ramírez-Correa et al. (2017) used the information system success model to explore the regulatory effects of learning styles and assess the success of the learning management system from a student perspective. Aldholay et al. (2018) used the mediating role of transformational leadership in the information system success model to prove the impact of online learning. Martins et al. (2019) proposed an education management information system model based on the information system success model to evaluate the satisfaction of students. Al-Fraihat et al. (2020) used information system success model to evaluate the success of digital learning.

Research method

Research model and hypotheses

This study applied the information system success model into the context of a recommender system. The information system success model is adapted as the basis for modifications. In a recent study, Shaikhc and Karjaluoto (2015) reviewed relevant literature to the use of information technology and systems, and found that most studies have tried to identify the human behavioural intent and the leading factors of use behaviour and the subsequent driving forces. Nearly half of the studies (43%) used satisfaction and perceived usefulness as key intrinsic factors, and empirical experience used established the influence of lead factors on continuous behaviour and use intentions. However, most previous studies explored the factors of information technology use and actual use behaviour to predict user acceptance (Tseng, 2015). Few studies have considered explicit characteristics such as perceived usefulness and quality at the same time to explore the user’s continued intention. Therefore, this study intends to construct an information system success model for a recommender system.

Seddon claimed that information system usage is a behaviour rather than a successful measurement method, and recommended using perceived usefulness as a measure of information system performance rather than use. Perceived usefulness is more appropriate than use for measuring system performance in mandatory use contexts and has been employed in recent research (Koo et al., 2013; Gorla and Somers, 2014; Choi and Park, 2015). Therefore, this paper also uses perceived usefulness as a measure. Since the quality of service is included in a successful model of information systems, many studies have begun to test its impact on use and user satisfaction. Several studies have found that the service quality construct is ineffective and included it in as a sub-item of system quality (Koo et al., 2013; Dwivedi et al., 2013; Zheng et al., 2013; Chen et al., 2020). Therefore, this study also removed this aspect. This study also introduces “dual-coding theory” to examine the model. The model consists of six dimensions: information quality, system quality, perceived usefulness, user satisfaction, continued intention to use and continued intention to provide. The research model is shown in Figure 2.

Romero et al. (2013) verified that the design of a tag cloud generation module can help users to more completely comprehend the retrieved information. Through a tag cloud presentation, users can easily search the database (Torres-Parejo et al., 2013). According to dual-coding theory, simultaneous presentation of text and non-text information can help users to understand information (Paivio, 2013). Huang and Chen (2014) have shown that the environmental design of smart phones can achieve better interaction with users and increase satisfaction when conforming to dual-coding theory. Jensen et al. (2015) believe that simultaneously presenting text and non-text information can help consumers recall the brand. Daniels et al. (2017) have shown how systems that conform to dual-coding theory can help people make decisions. Wang and Li (2019) have shown that both teachers and their students can benefit from its novel features based on the Dual-Coding. The teacher could efficiently develop reading comprehension lessons, analyse students’ learning outcomes, and evaluate students’ learning needs, while the students can work on their comprehension and retelling of the target passage/story in an engaging circumstance where the gestalt image of the story is generated through interactive drawings and diversified reading activities. Therefore, we propose the following hypothesis:

H1a: A presentation type using dual-coding theory in Recommender systems has higher perceived usefulness than a presentation type that has no use.
H1b: A presentation type using dual-coding theory in Recommender systems has higher user satisfaction than a presentation type that has no use.

The information system success model uses net benefits to evaluate system performance in improving task performance and effectiveness. Users can participate in recommender system in two ways: information consumption and information provision. Users can consume information by browsing or posting questions for help. They can also contribute information by replying to questions, and initiating or participating in topic discussions. Researchers have agreed that information consumption and provision are indispensable parts of social community (Ahuja and Galvin, 2003), as they are both desirable social behaviour (Ridings et al., 2006). In the Recommender systems context, the meaning of net benefits should be broadened as recommender systems users can obtain different types of net benefits. In an information-sharing recommender system, a user's primary goal is to access information either by purely browsing information or by participating in a discussion with others. Moreover, voluntary behaviour of helping others will occur when users enjoy contributing to a recommender system to benefit others. When users can benefit from information sharing, social support and voluntary contribution, their sense of community will be enhanced (Blanchard and Markus, 2004). As a result, this will greatly motivate users' involvement and active participation in the recommender systems in the future. In addition, when users can obtain useful information and social support from others in a virtual community, they may be motivated to help others in return for what they have received. On the other hand, users who provide information may also expect to receive help and access to superior resources from others (Zheng et al., 2013). It has been found that active users who provide information or participate in discussions feel discouraged if they cannot get useful information and responses from the community (Lin, 2008). Thus, it is expected that individual benefits will motivate two-way participation. Therefore, we define net benefits of a recommender system as two types of continuance intention: continued intention to use and continued intention to provide. The former represents the extent to which a user intends to continue browsing or seeking information in a recommender system, while the latter refers to the extent to which a user intends to continue contributing his/her knowledge.

The construct of information quality includes the desirable characteristics of system outputs. The quality of information the system produces, primarily in the form of a report or a Web page, is measured (Balaban et al., 2013). According to Petter et al. (2008), the information quality construct has proven difficult to capture and measure as it is not often distinguished as a unique construct. While some researchers have used the existing generic scales of information quality (Petter et al., 2008), others have developed their own scales. Some categories of information quality that can be measured are relevance, understandability, accuracy, completeness, usability, and importance (Balaban et al., 2013). Recently, the majority of related studies of information system success models have found that the information quality construct affects perceived usefulness and user satisfaction (Dwivedi et al., 2013; Tseng, 2015; Ramírez-Correa et al., 2017; Huang, 2018; Xu and Du, 2018; Martins et al., 2019; Al-Fraihat et al., 2020). Thus we present the following hypotheses:

H2: Information quality has a significant positive impact on perceived usefulness.
H3: Information quality has a significant positive impact on user satisfaction.

System quality measures the desirable characteristics of an information system. Since this dimension captures the system itself, it is oriented toward technical specifications like data processing capabilities, response time, ease of use, system reliability, and sophistication (Balaban et al., 2013). According to DeLone and McLean (2003), the system quality construct should measure technical success that Shannon and Weaver (1949) defined as the accuracy and efficiency of the communication system that produces information. The most common measure of system quality is the perceived ease of use related to the technology acceptance model. However, many researchers, including DeLone and McLean, have suggested that the perceived ease of use does not capture the construct as a whole (Petter et al., 2008). Therefore, researchers have created their own indices of system quality based on literature reviews or DeLone and McLean’s recommendations (Wang and Wang, 2009). Recently, the majority of related studies of Information System Success models have found that the system quality construct affects perceived usefulness and user satisfaction (Dwivedi et al., 2013; Tseng, 2015; Stefanovic et al., 2016; Ramírez-Correa et al., 2017; Wu, 2018; Shim and Jo, 2020). Thus we present the following hypotheses:

H4: System quality has a significant positive impact on perceived usefulness H5: System quality has a significant positive impact on user satisfaction

Perceived usefulness is defined as the user's subjective perception of probability that using a particular information technology system will enhance his/her performance (Davis et al., 1989). This in turn influences the user's intention to continue using the system. Kim and Ong (2005) noted that perceived usability and satisfaction are highly correlated. Wang and Liao (2008) provided the first empirical test for an adaptation of DeLone and McLean's information system success model in the context of government to citizens eGovernment. The hypothesised relationship between use and user satisfaction is significantly supported by the data. Liaw and Huang (2013) investigated learner self-regulation in e-learning environments. Their statistical results showed that satisfaction with e-learning is influenced by perceived usefulness. Bonsón et al. (2014) has found that the perceived usefulness has a significant effect on user satisfaction. Stefanovic et al. (2016) further evaluated the effectiveness of the Government Information System from the perspective of government employees, and found that perceived usefulness and user satisfaction have a significant, direct relationship. Ramírez-Correa et al. (2017) used the information system success model to explore the success of a learning management system. The study confirmed that perceived usefulness and satisfaction are highly correlated. Numerous research studies have provided empirical support for the positive association between usefulness and information technology use intention (Boakye et al., 2014; Sun and Mouakket, 2015; Tan, 2015; Hadji and Degoulet, 2016; Hadji et al., 2016; Apostolou et al., 2017; Joo et al., 2017; Demoulin and Coussement, 2020; Al-Fraihat et al., 2020; Bölen, 2020; Liang et al., 2021). Multiple studies have also attempted to identify the antecedents and drivers of post-adoption human behavioural intention and usage behaviour. A few significant antecedents are ‘consumer satisfaction’, ‘perceived usefulness’, ‘perceived ease of use’, ‘subjective norms’ and ‘perceived enjoyment’. Nearly half of the studies (43%) used ‘satisfaction’ and ‘perceived usefulness’ as key intrinsic factors to empirically establish the influence of these antecedents on continuous behavioural intention and usage (Shaikh and Karjaluoto, 2015). Thus we present the following hypotheses:

H6: Perceived usefulness has a significant positive impact on user satisfaction. H7: Perceived usefulness has a significant positive impact on continued intention to use. H8: Perceived usefulness has a significant positive impact on continued intention to provide.

The topic of user satisfaction has held a significant position in the marketing literature since satisfied customers can generate long-term benefits for companies, including customer loyalty and sustained profitability (Homburg et al., 2006). Customer satisfaction is highly related to behavioural intentions and is regarded as one of the key antecedents of post-purchase behavioural intentions because customer satisfaction has a positive effect on the customer's attitude towards the product or service and it can reinforce the customer's conscious effort to purchase the product or service again in the future (Oliver, 1980; Liu and Jang, 2009). In the mobile banking context, continued use intention is found to be solely dependent on the satisfaction of customers (Reji and Ravindran, 2012). In another empirical study, students’ continued use intention to e-texts are directly and meaningfully influenced by their satisfaction and perceived usefulness (Stone and Baker-Eveleth, 2013); user satisfaction with Web 2.0 applications (Facebook, Plurk, Twitter, and YouTube) and online knowledge groups significantly affect electronic word-of-mouth, which in turn significantly influences their intention to continue (Chen, 2012). Similarly, satisfaction and perceived usefulness were found to play a significant role in the continuance intention and Internet-based learning technologies (Limayem and Cheung, 2011). Most research has proved that user satisfaction has a significant positive impact on continuance intention (Zheng et al., 2013; Hsu et al., 2014; Sun and Mouakket, 2015; Hadji and Degoulet, 2016; Apostolou et al., 2017; Bae, 2018; Li and Fang, 2019; Alalwan, 2020; Shiau et al., 2020; Liang et al., 2021; Ambalov, 2021). This leads to the following hypotheses:

H9: User satisfaction has a significant positive impact on continued intention to use. H10: User satisfaction has a significant positive impact on continued intention to provide.

Experimental design

The experimental participants are readers who love books and are currently looking for favorite books. He/she uses the recommender system to find his/her favorite books. Experimental participants were randomly assigned to two different groups according to the experimental design for subsequent evaluation. A practical example was used to illustrate the process of this experiment. Through web page presentation, the experimental participants input the preference of four books categories and received the recommendation list (top 20 items). Then participants browsed the results and filled out the questionnaire. The experiment process took about 5-10 minutes. Roscoe (2004) noted that as a rule of thumb, between 30 and 500 is an appropriate sample size. The participants in this research were all graduate students from an information management department, and all had a basic understanding of the concepts of recommender systems. A total of 120 participants were randomly divided equally into two groups, A and B. Random assignment can effectively control the interference factor, reducing the interference effect that may occur during research and analysis, and avoid the loss of value of research results.

This study uses laboratory experimentation (Cooper and Schindler, 2013) and adopts C# as the programming language to construct the experiment platform. Microsoft Visual Studio Express 2013 is the development environment and Access2010 is the database. The algorithm that is used in recommender systems is adapted from a centering-bunching based clustering algorithm that is used for hybrid personalised recommender systems (Shinde and Kulkarni, 2012). All participants (both Group A and Group B) used the same recommender systems interface and algorithm to search and browse. The only difference between Group A and Group B was the presentation type of search results. Group A used the type of presentation of using dual-coding theory (text+image) (Figure. 3 left) and Group B did not use the type of presentation of using dual- coding theory (text only) (Figure 3 right). The experiential process is shown in Figure 4.

Figure 3: Text+image presentation (left) and text only presentation (right)

Measurement development

To ensure the content validity of the scales used in a study, the items selected for the constructs should represent the concepts about which generalisations are to be made. Hence the items selected for the constructs in this study were mainly adapted from prior studies to ensure content validity. Three items, selected from Gorla and Somers (2014) were adapted and used in this study to measure information quality. Three items, selected from Seddon (1997) and Gorla and Somers (2014) were adapted and used in this study to measure system quality. Three items, selected from Calli et al. (2013) and Boakye et al. (2014) were adapted and used in this study to measure perceived usefulness. Two items, selected from Belanche et al. (2012) and Aversano and Tortorella (2013) were adapted and used in this study to measure user satisfaction. Three items, selected from Bhattacherjee (2001) and Boakye et al. (2014) were adapted and used in this study to measure continued intention to use. Two items, selected from Yoon and Kim (2007) and Zheng et al. (2013) were adapted and used in this study to measure continued intention to provide. A 5-point Likert-type scale ranging from 1 (strongly disagree) to 5 (strongly agree) was used to quantify the questionnaire items. To rectify ambiguous wording and mitigate the length of the instrument, a pretest was used with three experts familiar with this field. After retesting the measures, these items were modified to fit the recommender systems context studied here (Table 1).

Data analysis

Measurement model

Reliability analysis

These questionnaires are considered to be reliable because their Cronbach α values are within a range of 0.764-0.922 (Table 2) and thus are greater than the 0.70 threshold, which indicates that the scales are reliable (Nunnally, 1978). The Cronbach αvalues are between 0.868 and 0.951 and all exceed 0.70, showing that the scales also have good reliability (Fornell and Larcker, 1981).

Convergent validity analysis

The questionnaire was developed based from the literature and then amended by experts to ensure it has content validity. Finally, this study uses factorial analysis to assess construct validity. Convergent validity is evaluated by factor loading and average variance extracted. Convergent validity is established if individual factor loadings are greater than 0.7 and average variance extracted exceeds 0.5 (Hair et al, 1992). Table 3 presents the factor loading values of individual measurement items and the average variance extracted of each construct. All the factor loading values are greater than 0.7 and the average variance extracted values exceed 0.6. The measurement scales exhibit convergent validity.

Discriminant validity analysis

Discriminant validity is determined through an examination of a cross loading table. A cross loading table must indicate that the measurement items load highly on their theoretically assigned factors and not highly on other factors (Gefen and Ridings, 2003). Table 4 shows that the loadings of individual items on their assigned constructs are higher than with other constructs. The square root of average variance extracted can also be used to determine discriminate validity, which is confirmed if the square root of average variance extracted is greater than the off-diagonal elements of a correlation matrix (Fornell and Larcker, 1981). Table 5 shows that all constructs satisfy this criterion and demonstrate adequate discriminant validity.

Structural model

This study compared the results of the SPSS analysis from the two groups, A and B, to test the hypotheses examining the effect of dual-coding theory (H1a-H1b). The analysis of SPSS is shown in Table 6. These results clearly indicate that Group A (text+image) was superior to Group B (text only), since both constructs all reached significant levels of 0.001. Thus, hypotheses H1a and H1b are supported.

This study used SmartPLS 3.0 to analyse the path coefficients of the main effects. The result is shown in Table 7. All path relationships (main effects) are significant. Therefore, hypotheses H2, H3, H4, H5, H6, H7, H8, H9, and H10 of this study are confirmed.

Finally, the predictive power estimate for this model is determined using the R-squared value in SmartPLS 3.0. The results account for 76.2% of the variance in perceived usefulness (PU), 70.6% of the variance in user satisfaction (US), 77.5% of the variance in continued intention to use (CIU), and 48.7% of the variance in continued intention to provide (CIP). Thus, the results indicate that the research model significantly explains the successful adoption of recommender systems.

Discussion

To start with the analysis of related data, we found a significant difference between the two groups (A, B) (P-value < 0.001) in both constructs, which means that dual-coding theory is effective on recommender system. The analysis results show that experimenters using dual-coding theory have a significantly higher perceived usefulness and user satisfaction than those not using. Both visual (image) and verbal (text) stimuli elicit not only imagery processing but also discursive processing in this study. This suggests that visual and verbal stimuli may have evoked referential or associative processing, as proposed in dual-coding theory. Dual-coding theory explains human behaviour and experience in terms of dynamic associative processes that operate on a rich network of modality-specific verbal and nonverbal representations (Clark and Paivio, 1991). Dual-coding theory assumes that cognition occurs in two independent but connected codes: a verbal code for language and a nonverbal code for mental imagery (Sadoski, 2005). Dual-coding theory has been validated in this study and the findings are consistent with those previous studies (Huang and Chen, 2014; Jensen et al, 2015; Daniels et al, 2017; Koc-Januchta et al, 2017).

In the main effect path coefficient of the research model, SmartPLS 3.0 was used for analysis, and the results showed that all hypotheses were significant. This study has four dependent variables: perceived usefulness, user satisfaction, continued intention to use and continued intention to provide. First, both information quality and system quality have a significant influence on perceived usefulness, but the path coefficient shows that system quality has a greater effect on perceived usefulness than does information quality. Therefore, we found that the system quality is more useful than the information quality. In other words, when we feel that the system quality is more stable and safe, our perceived usefulness will be more positive.

Three constructs affect user satisfaction, namely, information quality, system quality and perceived usefulness. It can be seen that information quality affects user satisfaction more than the other constructs do. In other words, we feel that the information quality is good when it affects subsequent satisfaction. The constructs that influence the continued intention to use are perceived usefulness and user satisfaction. Under comparison, user satisfaction is stronger. Obviously, the higher the user satisfaction, the higher the continued intention to use. User satisfaction is the degree of preference that a system generates after use. Therefore, user satisfaction has more effect on the continued intention to use. This study added continued intention to the model. Both perceived usefulness and user satisfaction have a significant influence on the continued intention to provide. And the effects almost are the same.

In an information-sharing recommender systems, a user is expected to access information either by purely browsing through information or by participating in a discussion with others. When users can benefit from information sharing, social support and voluntary contribution, their sense of community will be enhanced (Blanchard and Markus, 2004). In fact, users benefit from using a recommender system only when a recommender systems provides information valued by them. High-quality feedback and discussion help users have a better understanding of the topic, feel support from others and make a better decision (Zhang and Watts, 2008). High-quality information feedback benefits not only users who want to obtain useful information and get a recommendation on a particular topic, but also users who provide information. Thus, we believe that continued intention to provide plays an important role in creating various benefits to users. Therefore, continued intention to provide is a very important construct for evaluating the success of a recommender system. This finding is worthy of further study.

Implications

The argument in this study is that users presented with a recommender system using two or more media (such as voice, animation, pictures and text) have a high degree of comprehension. Thus, assuming they have higher perceived usefulness and user satisfaction, this argument echoes dual-coding theory. The two codes are believed to mutually support each other, and integration of multiple modalities should enhance the comprehensibility of the language. However, few studies have combined the information system success model with dual-coding theory. This study puts forward various hypotheses through the viewpoint of dual-coding theory, while also applying the information system success model. The findings are in line with the expected results of this study.

Theoretically, as dual-coding theory postulates, comprehensibility of a text is enhanced when the information is presented in a combination of verbal and non-verbal formats (Sadoski et al., 2012). Thus, when users retrieve information through a recommender system, users approach the interface and learn the recommended contents through more diverse channels, such as visuals, audio, and animation. These non-verbal cues, together with the verbal information, increase and enrich user' experiences and thus support user comprehension of a recommender system. Therefore, developers are encouraged to maximise comprehension by providing their users with multisensory searching/retrieving functions.

In this study, perceived usefulness and user satisfaction have a significant positive relationship both with continued intention to use and continued intention to provide. This result points to the usefulness of the user's cognition and the user's feelings of satisfaction when using a recommender system. A bookseller can introduce innovative or distinctive systems, improve the customer's awareness of the benefits of the recommender system, and increase the customer's interests and value. With more of these accumulated benefits, the higher the perceived value, the higher the satisfaction will be. In addition, activities or advertisements can be used to improve the customer's user satisfaction. With positive user satisfaction, customers naturally have good willingness to use and provide.

The R-squared value of the variance in continued intention to use, is 77.5% while that of the variance in continued intention to provide is only 48.7%. Although perceived usefulness and user satisfaction have a high degree of explanatory power for continued intention to use and continued intention to provide, the ability to explain continued intention to provide is relatively low. A possible explanation for this is that users are not accustomed to using a recommender system and providing a lot of information or feedback. That is, users do not know that feedback of users is a critical factor in the success of recommender systems. System developers can provide some incentives to encourage users to give feedback and provide information.

We show not only that dual-coding theory provides a unified explanation in recommender system, but also that its mechanistic framework accommodates theories cast in terms of strategies. Although much additional research needs to be done, the concrete models that dual-coding theory offers for the behaviour and experience of recommender system users further our understanding of recommender system and strengthen related practices.

Conclusion

The proposed research model was developed and validated utilising the extended information success model adapted to the context of recommender systems. This study examined the impact that factors relevant to recommender systems have on continuance intention to use and continuance intention to provide. System quality, information quality, perceived usefulness, and user satisfaction were used to measure continuance intention. The experiment's results have shown that the recommender system designed on the basis of the dual-coding theory is better than the traditional recommender system on all aspects of the facet. From the results of this study, we can see that in order to improve the continuance intention of users, it is necessary to improve the perceived usefulness, system quality and information quality of a site. If we focus only on slight changes to the algorithm, users may not be able to perceive a substantial difference. However, direct improvements to the user interface may greatly improve user willingness to continue using a recommender system. Future research into the following issues is warranted:

• Due to geographical constraints, the subjects in this study are all Asians. Due to cultural relations, the subjects’ willingness to provide is not high. This research can be extended to experiment in other regions in the future.
• The user’s opinions and ratings are important data needed by the recommender system before it generates recommendations. If the recommender system can continue to receive ratings and data from users, the system can continue to improve itself and keep up with the trend of the times. Therefore, how to make users willing to provide comments and ratings should be very important for the future.

Acknowledgements

This research has been partially supported by funds from the Taiwan Ministry of Science and Technology (MOST 107-2410-H-260-014-MY3; MOST 110-2410-H-260-016-MY2; MOST 109-2410-H-018-015).

About the authors

Wen-Yau Liang is a Professor of Information Management at National Changhua University of Education, Taiwan. His research interests are object-oriented design, artificial intelligence, intelligent agent and electronic commerce. His contact address is wyliang@cc.ncue.edu.tw
Chun-Che Huang, Corresponding author. He is a Professor in the Department of Information Management at National Chi Nan University, Taiwan and directs the Laboratory of Intelligent Systems and Knowledge Management (the ISKM Lab.). His contact address is cchuang@ncnu.edu.tw.
Bo-Ren Shih received his Master's degree in Information Management from the National Changhua University of Education, Taiwan. His contact address is vulbp739@gmail.com.

How to cite this paper