JOURNAL OF L A T E X CLASS FILES, VOL. XX, NO. XX, AUGUST 20XX 1 Implementation and Evaluation of a Collaborative Conversational Recommender in a 3D Virtual World David Contreras, Maria Salam´ o, Inmaculada Rodr´ıguez, and Anna Puig Abstract—A 3D Virtual World facilitates users’ interaction as they feel immersed and engaged in a shared virtual space. This type of interface may be specially useful when consumers employ home electronics for accessing online personalized services. In a previous research we focused on a Collaborative Conversational Recommender framework, where a synchronous and online 3D interface for multiple consumers integrates with a recommender. In this paper we go further and define a state-based model of user-recommender interaction that allows users move from different states of interaction (i.e., individual and collaborative) among users. Then, it is evaluated with users and compared with an individual approach. Our results demonstrate that the collaborative capacities proposed in the framework improves user experience and significantly increases the performance of the recommendation process, i.e users take less time in achieving the desired service. Index Terms—Application/Implementation < Entertainment & Services Technology, Human-Computer Interface < Human- Device Interaction, Interactive Technology < Human-Device Interaction. I. I NTRODUCTION Nowadays, users need to personalize services when they use electronic equipment in their everyday. For example, to personalize what content to watch in a smart TV, what music to listen in an audio device, or what video game to play in a videogame console [1]. Either from entertainment or from home-office devices, users are more than ever searching for products to consume among a large volume of content. The wide range of products and their specific and varied characteristics make it difficult for a user to search for an item. Recommender Systems [2] assist users in this search and provide them with suggestions of items to consume or to buy, taking into account their requirements. Recommender systems based on collaborative filtering [3] techniques have been used to suggest highly rated items for a target user based on the products that similar users have experienced in the past. However, for high-risk product domains 1 , where users are likely to search and buy products for the first time, the recommender cannot establish a meaningful profile for many of its recommendation seekers [5]. To overcome such cold start problems, Conversational Recommenders Systems [6] have been broadly recognized as an effective preference- based search and recommender technology. Conversational Recommender Systems use product’s features to help users to navigate through a product space, making alternatively product suggestions and eliciting user feedback [7]. 1 In high-risk product domains (e.g., domains where the products are very expensive), the task of locating a desired choice among a large set of options is indeed becoming intimidating for the average customer [4]. Nevertheless, most of these recommender systems lack of online collaboration to enable the user to be aware of and interact with other users that are simultaneously searching for a product. To address this issue, in a previous research we proposed a Collaborative Conversational Recommender (CCR) that integrates a conversational recommendation process in a 3D collaborative environment [8]. In this framework, users have the possibility of engaging in a joined search of a desired product. This framework consists of two main layers in charge of providing the User Interface and the Recommender Sys- tems, and a third layer which is responsible for communicating both of them [9], [10], [11]. In this paper, based on aforementioned research work and encouraged by the results obtained there with a simulator and with a preliminary evaluation with users, we propose a model that defines users’ states and the transitions during user- recommender interaction. This state-based model distinguishes individual and collaborative states that allows a more detailed analysis of the efficiency and efficacy 2 of the CCR algorithm with real users in a 3D Virtual World implementation. Ac- cordingly, we compare an Individual Critiquing [13] algorithm (IC), which has no transitions as the user does not collaborate with anyone, with a collaborative approach (where users’ states change over time) that enables interaction among users. The evaluation shows good results concerning usability as well as a significant improvement on efficiency and efficacy of our framework with respect to a non-collaborative one. II. RELATED WORK In this section, we analyze the main approaches that in- tegrate recommenders within virtual environments according different points of view: application domains, recommenders visualization and interaction platforms, recommendation meth- ods, and recommender collaboration capabilities. Regarding the application domain, most of the previous studies have been focused on implementing shopping assis- tants. For example, a solution of a virtual shopping mall on the Internet [14] or the recommendation of virtual objects inside a virtual reality interface [15]. Alternatively, others have focused on recommending locations inside the virtual world [16]. In the cultural domain, a recommender has been used to help users in navigating in 3D spaces [17] (i.e., both museums and galleries). Although the application example of our CCR framework focuses on an e-commerce domain, it is applicable to other domains. 2 In this work, the efficiency is measured through the number of recommen- dation cycles to reach a desired product and the efficacy is measured through the Decision Accuracy measure using the same methodology described in [12].