Traditionally, recommender systems for the Web deal with applications that have two dimensions, users and items. Based on access logs that relate these dimensions, a recommendation model can be built and used to identify a set of N items that will be of interest to a certain user. In this paper we propose a method to complement the information in the access logs with contextual information without changing the recommendation algorithm. The method consists in representing context as virtual items. We empirically test this method with two top-N recommender systems, an item-based collaborative filtering technique and association rules, on three data sets. The results show that our method is able to take advantage of the context (new dimensions) when it is informative.
Most Web sites offer a large number of information resources to their users. Finding relevant content has, thus, become a challenge for users. Recommender systems have emerged in response to this problem. A recommender system for a Web site receives (implicit or explicit) information about users and their behavior and recommends items that are likely to fit his/her needs [12].
Recommender models for Web personalization can be built from the historical record of accesses to a site, where one access is a pair < user id, item >. Each access is interpreted as a rating of 1 given by the user to the item. However, other dimensions, such as time and location, can add contextual information and improve the accuracy of recommendations. For instance, the type of books that a user looks for in Amazon during work hours is probably different from the books searched for during leisure hours.
According to [11], the idea that contextual information is important when predicting customer behavior is not new. Many Web sites are supported by Content Management Systems (CMS), that often store much contextual information. However, this is not true in all cases and, additionally, getting information that is really relevant for recommendation is a hard task in many applications [8]. Adomavicius et al. [1] have investigated the use of context for rating estimation in multidimensional recommender systems. Palmisano et al. [11] have used contextual information to improve the predictive modeling of customer’s behavior. Both authors have developed a special-purpose browser to obtain rich contextual information.
In this paper we exploit how contextual information can be used to improve the accuracy of Top-N Recommender Systems. Existing contextual recommender systems typically use contextual information as a label for segmenting/filtering sessions, using them to build the recommendation model (e.g., [1,11]). We follow an alternative approach, which uses the contextual attribute as a virtual item. This means that it is treated as an ordinary item for building the recommendation model, which has the advantage of allowing the use of existing recommendation algorithms. As our contextual information are obtained from multidimensional data, we have called our approach DaVI (Dimensions as Virtual Items). Instead of a special-purpose browser [1,11], we collect the multidimensional data from Web access logs and from attributes stored in databases of the Web sites. We have empirically tested our approach with two recommendation techniques, item-based collaborative filtering and association rules, to assess the effect of adding context on the accuracy of traditional Web recommender systems. We present results obtained on three data sets.
In the following section, we present the contextual information used in our experiments. Next, we describe the recommendation techniques and the approach proposed. Then, we discuss results and present conclusions and future work.
There are many definitions of context in the literature depending on the field of application and the available customer data [11]. In this paper, context is defined as any information that can be used to characterize the situation of an entity [5]. Here an entity is an access to an item/Web page by a user.
A critical issue is how to obtain the rich contextual information [6]. In some circumstances, context is explicit, such as a person informing a movie recommender system where he/she wants to watch a movie. On the other hand, the contextual information can also be inferred from Web access data. For example, we can observe if a person bought an item, from an e-commerce Web site, on a weekday or a weekend, from the Web access logs.
Besides general contextual information that can be obtained from access logs, we may use domain-specific information, that is typically collected from the CMS. For example, if an item represents an access to a music, the genre of the music can be used as a dimension of contextual information.
In Table 1 we present the dimensions/contextual information considered in the experiments presented in this paper. The first group of contextual information was obtained by pre-processing Web access logs. The second group was collected from the CMS of a Web site of Portuguese Music1 used in this study. The last group refers to a public data set2 that contains a record of user interactions with the Entree Chicago restaurant recommender system. All the information is stored in a data warehouse that was specifically designed for modeling Web sites [7]. The intention of navigation in a restaurant recommendation system (for example, the search for a restaurant cheaper, closer, more traditional, more creative, and so forth). There are 9 different intentions of navigation in our experiments.
A recommender system for the Web typically outputs an ordered list of recommendations, given a trail of recent Web page requests. Historical information about the behavior of the use
This content is AI-processed based on open access ArXiv data.
