A Study On Social Network Information Technology Essay

Published: 2021-07-24 23:40:06
essay essay

Category: Information Technology

Type of paper: Essay

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Hey! We can write a custom essay for you.

All possible types of assignments. Written by academics

Abstract— The explosive growth of the world-wide-web, social networking websites and the emergence of e-commerce has led to
information overload. This lead to the development of recommender systems - a personalized information filtering technology used to identify a set of N items that will be of interest
to a certain user. Recommender systems are traditionally based on users rating of items. The study of social network based top n recommendation began recently. Social networks help in improving the recommendations. In this paper we review various works dealing with recommender systems based on social networks.
Keywords— recommender systems, social network, trust network, top n recommendation, collaborative filtering
We are living in the information age. There are websites that offer many products for sale. The number of options available is overwhelming. Usually, people rely on recommendations or suggestions from other people, especially from people they trust. Recommender systems assist and augment this natural social process to help people sift through available options. There are recommender systems for movies, music, books etc. For e.g. when viewing a product on Amazon.com, the store will recommend additional items based on information about what other shoppers bought along with the currently selected item[1]. News services use recommender systems to identify articles of interest to readers, based on the articles that they have read in the past [17].
A simple recommender system consists of a set of users and a set of items. Each user rates items of his preference with some rating (for e.g. on a scale from 1 to 5). There are two recommendation research problems. First one is predicting the rating for user on target item [9][10]. Second one is producing a ranked list of items that he has not yet rated but is likely to rate highly. This is referred to as top-N item recommendation problem [2][6][7].
Recommender systems can be broadly classified as content-based and collaborative filtering systems [18]. In content-based recommender systems [19], the user will be recommended items similar to the ones the user preferred in the past. Collaborative filtering [3] systems identify customers (neighbors) whose interests are similar to those of a given customer and recommends products that the neighbors of a given customer have liked. There are also hybrid systems which use both content-based and collaborative methods.
Collaborative filtering (CF) [3] is the most commonly used recommendation technique, but has two major limitations – sparsity and scalability. The number of ratings already obtained is very small compared to the number of ratings that need to be predicted because typical collaborative filtering requires explicit non-binary user ratings for similar products. This approach is only effective when users have expressed enough ratings to have common ratings with a good number of other users. It performs poorly for so-called cold start users, i.e. new users who have expressed only very few ratings and are unlikely to have users with similar rating profiles. Algorithms to find the neighborhood usually require very long computation time that grows linearly with both the number of customers and the number of products. With millions of customers and products of real world situations, existing collaborative filtering based recommendations suffer serious scalability problems.
It is natural for us to ask other people for suggestions when we have to make a decision. We normally go to friends who are knowledgeable in the required area or to people who have a good reputation in that area. It is tempting to build a recommender system that mimics this real world recommendation process. Growth of online social networks has helped the development of such systems.
There are trust aware recommender systems and social recommender system. Concept of trust is different from that of friendship [16]. In product review websites like epinion.com users can trust a person whose reviews are very good. They may not know the person in real life. Also the relationship is not mutual. However, in social friendship people know each other personally and relationship is mutual. Traditional recommender systems are based on single data source – user-item matrix. This matrix as shown in Fig. 1 consists of ratings of each user for each item he prefers. Trust aware and social recommender systems have an additional data source – trust network or social network matrix as shown in Fig. 2. This matrix shows the relationship between the users. In a trust or social network based recommender system, recommendations for a user are based on the ratings of the users that are directly or indirectly trusted by him. These recommenders can make
recommendations as long as a new user is connected to a large enough component of the network [4]. Therefore, these methods tend to outperform collaborative filtering methods for cold start users.
Fig. 1 User-Item Matrix
The rest of the paper is organized as follows: Section II deals with collaborative filtering based recommender systems. We review the works done on top-N recommendation problem which use only user-item matrix. Then in section III we discuss different social and trust network based approaches for recommendation. In section IV we conclude the paper.
cf based recommender systems
Collaborative filtering approaches find a neighborhood of similar users and rank the items rated by these users to perform top-N recommendation. Two types of collaborative filtering approaches are widely studied: memory-based and model-based [5]. Memory based algorithms make recommendations based on the entire collection of previously rated items by the users. Model based algorithms use the collection of ratings to learn a model, which is then used to make rating predictions. Collaborating filtering can be user-based or item-based. To produce recommendations for active users, user-based approaches [6] use ratings of their similar users, and item-based approaches [2] use computed information of items similar to those chosen by the active user.
User-based and item based approaches often use the Pearson Correlation Coefficient algorithm [7] and the Vector Space Similarity algorithm [5] as the similarity computation methods. Pearson Correlation Coefficient tries to measure how much two users vary together from their normal votes - that is, the direction/magnitude of each users vote in comparison to their voting average. If they vary in the same way on the items they have rated in common, they will get a positive correlation; otherwise, they will get a negative correlation. In vector similarity measurement, two users are treated as vectors in m-dimensional space, where m is the number of items in the database. Then the angle between the two vectors is measured. If the two vectors generally point in the same direction, they get a positive similarity; if they point in opposite directions, they get a negative similarity.
A user based collaborative filtering approach is presented by McLaughlin et al. [6]. They introduce the Belief Distribution Algorithm that computes the belief (distribution) of rating differences instead of point estimates of the rating as done by existing methods. They estimate the belief difference between each user’s average rating and the estimated rating on the items. The predicted belief difference for the source user and a given item is computed and added to the source user’s average rating to obtain the predicted rating. Finally, the items having one of the top-N predicted ratings are returned as the recommended items. But this method cannot handle cold start users.
Fig. 2 User-User Trust Matrix
A item based approach for top n recommendation is presented in [2]. They present two different methods of computing the item-to-item similarity. One models the items as vectors in the user space, and uses the cosine function to measure the similarity, whereas the other computes the item-to-item similarities using a technique based on the conditional probability between two items. The second measure can differentiate between users with varying amounts of historical information as well as between frequently and infrequently purchased items. However, these methods are also not suitable for cold start users.
Kwon et al [7] proposed a technique that can be used along with existing recommender system. They give techniques to improve accuracy by filtering out recommendations above a minimum rating standard deviation threshold. But it was found to decrease diversity of recommendations. So instead of filtering out high variance items, they give methods that penalize such items, so that a user can control the balance between the accuracy and diversity of recommendations.
cf based social recommender systems
Matrix Factorization and Neighborhood based methods have been proposed for incorporating trust or social network information into recommender systems. In the matrix factorization technique user-item rating matrix and trust matrix is used to learn low rank matrices that captures user latent features and item latent features. These matrices are then used for predicting the rating matrix. Neighborhood based approaches use the ratings in the user-item matrix and trust values directly in the recommendation process.
Matrix Factorisation Based Approaches
A factor analysis method based on the probabilistic matrix factorization is presented in [9]. Their method learn the user latent feature space and item latent feature space by employing a user trust network and a user-item matrix simultaneously and seamlessly. The user latent feature space connects the two data sources as it is shared between the two. Although social network information is integrated into the recommender, the real world recommendation processes are not reflected in the model. Users have two different sets of feature vectors which makes it difficult to understand the model.
The same authors then improved upon their work by proposing a new probabilistic factor analysis framework [20]. Their work is based on the idea that every user has his own taste and at the same time, every user may be influenced by his friends he trusts. The model naturally fuses the users’ tastes and their trusted friends’ favors together. However, in their model, the feature vectors of direct neighbors of a user u affect the ratings of u instead of affecting the feature vector of u. Their model does not handle trust propagation.
Authors of [10] advanced the previous work by incorporating the mechanism of trust propagation into the model. In their model the latent feature vector of a user u is dependent on the latent feature vectors of all his direct neighbors. There is a transitive effect as feature vector of each direct neighbor is dependent on the feature vector of his direct neighbors. This allows model to handle transitivity of trust and trust propagation. The model estimates latent feature vector of a user as the weighted average of the latent feature vectors of his direct neighbors.
All the above matrix factorization techniques are aimed at improving prediction task. Yang et al. [11] extended these techniques for top-n recommendation by including observed and missing ratings in their model. Their work is based on the assumption that user’s have a selection bias which causes the observed feedback (e.g. ratings, purchases, clicks) in the data to be missing not at random(MNAR). The technique of training model on all items outperforms other state of the art approaches [15].
Ma et al [16] proposed a new matrix factorization model for prediction problem which gives importance to social friendship. They give two methods to capture social network information. First method is based on the assumption that every user’s taste depends on the average taste of his friends and similarity between them. However, this approach is insensitive to those users whose friends have diverse tastes. The second method considers each user and his friends individually. It tries to minimize the distance between feature vector of a user and each of his friends based on how similar they are. This method models the propagation of tastes.
[21] extends the previous idea of treating each friend differently. They focus on inferring category-specific social trust circles from available rating data combined with social network data. Only a subset of friends is taken into account when performing rating prediction in a specific category. For e.g. when a person needs suggestions on a electronic item he consults friends who have good taste in that field. To infer the weight of a edge in a circle, they estimate a user's expertise level in a category based on his rating activities as well as of all users trusting him and assign values proportional to their expertise levels. The authors apply this idea on the model proposed in [10]. They create a separate model for each category and use them for making predictions.
Neighbourhood Based Approaches
Tidal Trust [12] performs a modified breadth first search in the trust network to compute a prediction. It finds all raters with the shortest path distance from the source user and aggregates their ratings weighted by the trust between the source user and these raters. To compute the trust value between users u and v who are not directly connected, it aggregates the trust value between u’s direct neighbors and v weighted by the direct trust values of u for its direct neighbors. Since TidalTrust only uses information from raters at the nearest distance, it may lose a lot of valuable ratings from users a little further apart in the network.
Massa et al. [13] proposed a similar method but it considers all raters up to a maximumdepth which is given as an input. maximum-depth is independent of any specific user and item. Also, to compute the trust value between u and v, a backward exploration is performed. It means that the trust value from u to v is the aggregation of trust values between u and users directly trusting v weighted by the direct trust values.
It is difficult to determine how far to go in a trust network. Suggestions from direct neighbors are more trustworthy than those from far neighbors in trust network. Based on this idea, a random walk method is proposed in [14] to combine the trust-based approach and item-based collaborative filtering approach for predicting the rating of single items. The method performs random walks on the trust network to find ratings for the target items or similar items. The stopping criteria for a single random walk at a certain user depends on the similarity of items rated by that user and the target item, and on the current step of the random walk. The same authors extended the method to deal with top-N recommendation [4]. Here stopping criteria for a single random walk is based on similarity of two users being considered and not on the target item. This captures the intuition that users with similar rating patterns are more likely to agree on their top-N items.
Yang et al [11] proposed a nearest neighbour based top-n recommendation method that combines users’ neighborhoods in the trust network with their neighborhoods in the latent feature space. To calculate users latent feature space they used a low rank matrix factorization model which considers all the items instead of only the observed ones [15]. The users are then clustered in the user latent feature space using the Pearson correlation coefficient. Once both the neighborhoods are calculated, they use voting instead of the weighted averaging of the observed ratings, to construct top-n recommendations.
A bayesian inference technique is presented in [22] to deal with prediction problem. They measure the rating similarity between a pair of friends by a set of conditional probabilities derived from their mutual rating history. When a user needs a recommendation he requests his friends. This is modeled as a one level bayesian tree with user as the root and each of his friends as a direct child. They assume that friends’ ratings are independent with each other and conditional on user’s rating. Then they use bayesian rule to calculate conditional probability of user’s rating based on his friends rating. To model real world situation more accurately, they allow a user to propagate his query through the social network and collect ratings from indirect friends who are several social hops away.
CF based Recommender Systems are well studied in academia. Social network based recommender systems are just evolving. In this paper we reviewed various works dealing with recommender systems with emphasize on social networks. First we saw recommender systems which use only user-item matrix. This system suffers from cold start and scalability problem. Then we discussed various works that use social or trust network to deal with these problems. We saw different matrix factorization and neighborhood based approaches used for injecting social or trust network information into the recommender system. Most of the works are aimed at solving prediction problem. There are only few works that use social network information to improve top n recommendation accuracy.

Warning! This essay is not original. Get 100% unique essay within 45 seconds!


We can write your paper just for 11.99$

i want to copy...

This essay has been submitted by a student and contain not unique content

People also read