p.p1 space and improve the scalability of the

p.p1 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px ‘Times New Roman’; color: #000000; -webkit-text-stroke: #000000} p.p2 {margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px ‘Times New Roman’; color: #000000; -webkit-text-stroke: #000000; min-height: 16.0px} span.s1 {font-kerning: none}

One of the first recommendation systems built using collaborative filtering approach was Tapestry. The Tapestry system required the users to manually identify similar users in the user space based on their preferences. This approach is not feasible as the user space contains several thousands of users and it becomes a tedious task for the users to find similar users manually. Later, the development began for more automated recommendation systems that find similar users and items in real time. The GroupLens research team built a collaborative filtering system that recommends Usenet news and movies to the users. Video Recommender and Ringo are web based recommender systems that recommend movies and music respectively.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

Several different techniques have been applied to recommendation systems recently. These techniques include using Bayesian Networks, Clustering and Horting. Bayesian Networks use a decision tree structure where the nodes and edges represent information related to the users and items. This tree is built using the previous data available. The decision tree that represents the Bayesian Network is built based on the users’ preferences and because of this reason Bayesian Networks are suitable in an environment where the preferences of the users do not change constantly. In an environment where the users’ preferences change frequently, learning a model for the Bayesian Network in real time becomes computationally expensive and difficult.

Clustering techniques use the users’ preferences to identify similar users and combine them together to form user groups. To generate recommendations for a user, the ratings of the users present in the respective user group are aggregated. Some clustering techniques assign users to multiple groups based on the level of participation. This level of participation is taken into account while aggregating the ratings from the user group. Clustering techniques reduce the size of the user space and improve the scalability of the recommendation systems but the recommendations generated are not very accurate as they are an approximation of the users present in the cluster. In many cases, clustering is used as an initial technique to narrow down the user space while identifying the nearest neighbors.

Another implementation of recommendation systems is Horting, which uses a graph based approach. In Horting, the users are represented using nodes and the edges represent the level of similarity between the users. Once this graph is built using the previous data, the graph can be traversed across the users to aggregate ratings of the nearby users to generate recommendations. Horting also uses feedback learning while generating the recommendations i.e. after providing recommendations to the users, the algorithm asks the users what rating would they have actually given for the recommended item. The main advantage of Horting over traditional nearest neighbor algorithms is that it also takes into account the users’ preferences who have not rated the target item yet.

As the traditional collaborative filtering algorithms are not scalable, Amazon.com uses an item-based collaborative filtering which is scaled to their massive catalogue and produces accurate recommendations in real-time. The algorithm matches the user’s purchased and rated items to a list of similar items and combines this list to form recommendations. The similarity between the items is determined by using an item similarity matrix that keeps track of the items that are frequently bought together by customers. The key point of this algorithm that makes it scalable is that the similarity matrix of items is computed offline. The online component of the algorithm then searches through this similarity matrix to find the items most similar to the items purchased by the customer. This computation is quick and is independent of the catalog size and the user space and only depends on the number of items actually purchased or rated by the user.

Most of the technological enthusiasts assume that a recommendation system must be completely automated providing ephemeral recommendations. But performs an investigative study on different kinds of recommendation systems usually used in E-commerce. This study concludes that the recommendation systems which include some amount of input from the users form a personal relationship with the users. This kind of relationship helps in increasing the quality of recommendations and also the loyalty of customers towards the website. All of the recommendation systems mentioned so far still do not handle the problem of sparsity and high dimensionality of the input data. investigates a dimensionality reduction approach for recommendation systems to handle the problem of sparsity and improve the quality of recommendations. This study used Singular Value Decomposition (SVD) which drastically reduced the dimensions of the ratings matrix. However, the results of this study indicated that the performance of SVD based recommender systems for very sparse data was worse than the performance of usual collaborative filtering algorithms. But for denser datasets the performance of SVD was better than the performance of collaborative filtering algorithms.

The experiment compared user-based collaborative filtering with item-based collaborative filtering to determine which approach is more scalable and accurate. The experiment tested item-based collaborative filtering with different similarity metrics and neighborhood size and compared it with a benchmark user-based model. The results indicated that ‘adjusted cosine-similarity’ produced accurate results when compared to other similarity measures. The size of the neighborhood also had an impact on the performance of the recommendation system. Finally, the results of the experiment concluded that item-based collaborative filtering is more scalable and produces accurate results when compared to user-based collaborative filtering.