الگوریتم چند معیاری کلونی مورچه ها برای ایجاد لیست پخش موسیقی

In this paper we address the problem of music playlist generation based on the user-personalized specification of context information. We propose a generic semantic multicriteria ant colony algorithm capable of dealing with domain-specific problems by the use of ontologies. It also employs any associated metadata defined in the search space to feed its solution-building process and considers any restrictions the user may have specified. An example is given of the use of the algorithm for the problem of automatic generation of music playlists, some experimental results are presented and the behavior of the approach is explained in different situations.

We live in the so-called informationsociety, which is a natural consequence of the general advance in technology and especially of the development and growth of the Internet since its birth in the early 90s. Users are now faced with the challenges involved in learning to apply the sources of knowledge generated by the new model of society and to evolve according to its requirements. In dealing with these challenges, they have to handle an information overload that makes it difficult to select the right information and are forced to perform the cumbersome task of exploring excessively dense spaces, an impossible task without the assistance of intuitive and efficient searching tools. In order to overcome this problem, recommendersystems were developed in the 90s to facilitate the automatic and personalized selection of the products that best match individual user preferences. They have become an important research area (Adomavicius & Tuzhilin, 2005) since they are problem-rich, have abundant practical applications in helping users to deal with the information overload and provide them with personalized recommendations, content, and services. A good example of these applications is Amazon.com (Linden, Smith, & York, 2003), where customers receive recommendations on books, CDs and other products. Recommender systems follow two main recommendation strategies: content-based and collaborativefiltering. Content-based filtering relies on descriptions of the items that are being recommended (Melville, Mooney, & Nagarajan, 2001) while collaborative filtering is based on similar past tastes and preferences (Sarwar, Karypis, Konstan, & Reidl, 2001). There are also hybrid approaches that combine collaborative and content-based methods (Adomavicius & Tuzhilin, 2005). However, all these approaches have questionable issues or limitations. When it comes to evaluating the relevance of a set of available products, traditional recommender systems analyze user preferences, which must be properly modeled and stored in a user profile by means of more or less sophisticated mechanisms, none of them free from deficiencies. Generally, the most straightforward approaches simply perform syntactic comparisons against a pre-established set of keywords and hence suffer from the usual shortcomings of mechanisms that do not use semantics to consider the meaning of terms such as synonyms, multiple meanings, etc. Some proposals, based on automatic classifiers, evaluate the relevance of a certain product for a certain user through the use of occurrence patterns of the product attributes or characteristics on a pre-defined training set. Others disregard all product descriptions as a criterion and only consider the associated rating (level of interest) previously defined by the user or by other users of the system with similar preferences. For an extended review of common issues and limitations of recommender systems (see Adomavicius & Tuzhilin, 2005). As new technological breakthroughs become mainstream, the amount of information available to the information society just keeps on growing. This has been estimated as 34 gigabytes for an average person on an average day (Bohn & Short, 2009) only in the USA. In other words, the information overload issue is not getting any better. For example, the amount of recorded music of all types available at any online music store (e.g. iTunes, Amazon.com, etc.) exceeds the average life expectancy. However, as all the music has been annotated with metadata (i.e., valuable knowledge) and the amount of music available will go on growing, we find two requirements that any recommender system should address in this domain: firstly, it has to make use of the available metadata that describe the items and, secondly, it has to be able to deal with large collections of items, as users tend to add more items to their collections as storage capacity increases. There is also a clear requirement for an effective recommender system: it must be capable of automatically creating a list of recommendations that match the criteria or preferences specified by the user. In this paper we propose an antcolonyoptimization (ACO) algorithm for use as a recommender system for the automatic generation of music playlists that fulfils the user’s predefined set of criteria. It has been designed to cover the two requirements specified above: i.e. it supports multi-criteria recommendation and guarantees good performance when dealing with large sets of items. The paper is organized as follows: in Section 2 we review similar studies. Section 3 gives an overview of ACO, provides a general definition of the problem and describes the functioning of the proposed algorithm. Section 4 presents the domain-specific problem used to test the algorithm, explains the terms of the ontology used and discusses the results obtained from the execution of the algorithm with different sets of criteria. In Section 5 we present our conclusions and describe future work.

In this paper we have dealt with one of the biggest problems that inhabitants of the so-called information society must address. Due to the vast amount of available information (information overload), we are forced to explore excessively dense spaces, turning the selection of interesting information into a cumbersome task, extremely hard to carry out without the assistance of intuitive and efficient searching tools. We believe recommender systems are one of the most significant advances in dealing with this problem. In this study we focused on the specific context of digital music, where users are forced to resort to electronic devices to access their ever-growing digital music collections. In this particular scenario there is a clear requirement for personalization of information and assistance in generating playlists to make the most of digital music collections. This requirement can be defined in terms of the general optimization problem known as the OrienteeringProblem, which belongs to the category of NP-Hard problems. Obtaining optimal playlists automatically with a large number of albums cannot be done effectively with traditional computing strategies. To overcome this problem and achieve the required high level of personalization, we proposed the utilization of a modified version of an ACO algorithm that is able to tackle both the optimization problem and the personalization requirement. The algorithm presented here can be applied not only to generating music playlists but also, by the use of the appropriate ontology and the definition of evaluation functions, to many other domains where information tagged with descriptive metadata is retrieved while satisfying a specified number of preferences. Other possible uses of the algorithm include generating guided visits to tourist attractions, shopping centers, museums, art galleries, etc. Our plans for future work include evaluating our algorithm on other domain-specific problems, with larger repositories of items, increasing expressivity of restriction sets by allowing the use of logic operators such as or, and or not, and a test with users to assess the usefulness and acceptance of the system proposed here.