Leveraging Social Signal to Improve Item Recommendation for Matrix Factorization

05/17/2018 ∙ by Ze Wang, et al. ∙ 0

Although Recommender Systems have been comprehensively studied in the past decade both in industry and academia, most of current recommender systems suffer from the following issues: 1) The data sparsity of the user-item matrix seriously affect the recommender system quality. As a result, most of traditional recommender system approaches are not able to deal with the users who have rated few items, which is known as cold start problem in recommender system. 2) Traditional recommender systems assume that users are independently and identically distributed and ignore the social relation between users. However, in real life scenario, due to the exponential growth of social networking service, such as facebook and Twitter, social connections between different users play an significant role for recommender system task. In this work, aiming at providing a better recommender systems by incorporating user social network information, we propose a matrix factorization framework with user social connection constraints. Experimental results on the real-life dataset shows that the proposed method performs significantly better than the state-of-the-art approaches in terms of MAE and RMSE, especially for the cold start users.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction and Related Work

As the exponential growth of information generated from the website, the recommender systems have become more and more popular in both academia and industry. Recommender systems are a subclass of information filtering techniques that seek to predict the rating or preference that user would give to an item (movies, books, music, restaurant).

Although recommender systems have been widely used, such as Google, Amazon, Ebay, most of recommender systems suffer from several common drawbacks. First of all, data sparsity is the most serious problems for designing effective and efficient recommender systems. As indicated in  [17, 23, 18], the density of data is less than in most of recommender systems. In other words, in typical user-item rating matrix, the number of nonzero ratings in this matrix is less than . This phenomenon is obvious and for example, customers in real life always purchase very few items in amazon compared to the whole items in amazon database. So most of collaborative filtering based recommender systems such as  [1, 2, 3, 8, 10, 13, 12, 19, 21] cannot handle the users who have rated few items. What is more, traditional recommender systems ignore the social connection among users in both social based network and trust based network. It assumes that all the users are independently and identically distributed and only use the user-item matrix for recommendation task. However, in real life, we always turn to our trusted friend for suggestions for buying books, cloth, and watching movies, etc. So our preference can be easily affected by our friends. Therefore, simply ignoring social connection is not realistic when doing recommendation systems and considering the social connection information in recommender system can efficiently deal with the cold start users [25, 9].

In this paper, aiming at solving the above problems, we propose a matrix factorization framework with social network information as regularization term. The proposed model effectively handles the information from two resources, user-item matrix and user social/trust network. More specifically, the social network information is used for designing the social regularization term to constrain the matrix factorization objective function. The experimental analysis on one real-life dataset shows that our proposed model outperforms several state-of-the-art algorithms.

Social based recommendation system has been studied in [15, 14, 11]. Furthermore, matrix factorization is widely used in the areas of social network analysis [6, 7, 5, 4], and name entity disambiguation [16, 24, 20, 22]. The rest of this paper is organized as follows: The problem definition and preliminary background are presented in section 2. Basic matrix factorization framework is introduced in section 3. The proposed model is described in section 4. Our experiment is reported in section 5. Finally I conclude the paper.

2 Problem Definition and Preliminary Background

In recommender system, we have a set of users , and a set of items . The user-item rating matrix is represented as and is denoted as the rating of user on item . Typically the domain of rating is the real number ranging from to .

For social network, given a directed network , wehre every user , if user is a directed neighbor of user , in other words and is a set of directed neighbors of user . For example, in trust network, if user likes the item review that user writes, then there is a outgoing edge from to , but at the same time doesn’t necessarily link to . Without losing generality, we denote the adjacent matrix of network as . In this project, the edge weight in matrix is a binary value . For instance, if user has an outgoing edge to user , then , otherwise .

Assume that a given user , an item , the recommender task in this project is to predict the missing value given matrix and . In this work I utilize the matrix factorization based model to learn the latent factors of users and items and predict the missing rating value in matrix .

3 Basic Matrix Factorization Model

In recommender system, an efficient approach for predicting missing values in user-item matrix is to employ matrix factorization method. Basically matrix factorization method is to factorize the user-item matrix and use the low ranked user and item factor matrices for missing rating value prediction.

Suppose given user-item matrix , the basic matrix factorization model is as follows:

(1)

where is the user-item rating matrix, is the user factor matrix and is the item factor matrix, and is the number of latent space dimension for users and items in user-item matrix .

The approximation of user ’s rating on item , which is denoted by , is defined as , where is the user factor for user , is the item factor for item . And the objective function of matrix factorization model is as follows:

(2)

where . We are going to minimize the so as to solve both and . In order to solve the optimization problem in equation  2, a gradient descent approach can be used to obtain a local minimum.

4 Social Connection Based Matrix Factorization Model

Traditional recommender systems, like collaborative filtering, only utilize the user-item rating matrix information for recommendation but ignore the social connections among users. Due to the factor that the online social network is becoming more and more popular, incorporating social network information in recommender system becomes more and more important. Particularly by embedding the social connection information in recommender system, it can efficiently solve the cold start problem. In this section, I will incorporate the user social connection information as regularization term into basic matrix factorization framework and use the gradient descent approach to solve the proposed model.

4.1 Social Regularization based Model

We formulate the objective function as the following minimization problem:

(3)

In the formulation above, is the user factor for user (the u-th column in ), is the item factor for item (the i-th column in ). is the similarity function to indicate the similarity between user and user . Also we use to denote user ’s outlink friends and use the notation to represent user ’s inlink friends. In trust network, like Epinion, doesn’t necessarily equal to since trust network is directed network typically.

A local minimum of the objective function given by the objective function can be found by performing the gradient descent with respect to latent vectors

and . The derivation procedure for user and item latent factors are as follows:

One of the advantages to employ this model is that the proposed model can capture the trust propagation of different users. For instance, if user has a outgoing friend , and also has a outgoing friend , but at the same time and are not friend based on the social network connection. Using this model, it can indirectly minimize the feature distance between and .

4.2 Similarity Measurement

in the proposed model, we need to quantify the similarity between users in social network. Given the knowledge of user-item matrix, we can model the user user similarity based on purchased items and corresponding ratings. In order to achieve this goal, Pearson Correlation Coefficient (PCC) and Vector Space Similarity (VSS) are proposed to define similarity between different users. The equations are as follows:

where represents the average rate of user and belongs to the subset of items which user and user both rated. is the rate user on item .

From the above definitions, we can see that and . The larger values for both VSS and PCC, the more similar between different users. Also in order to constrain the range of measurement into , I use a simple mapping function to bound its similarity range from to .

5 Experiments and Results

5.1 Epinions Trust Dataset

The dataset we employ for this work is Epinions111http://www.epinions.com. In Epinions, every user can read the reviews about a variety of items and also users can write a review for particular items. For the social network viewpoint, each member in Epinions has a trust list of other members to indicate if I trust your review or not. From my study, the dataset contains different users and different items. And the total number of rating is .

In order to verify the correctness of similarity measurement in my model, the first study is to validate my assumption that in trust network such as Epinions, social friends have similar tastes. We utilize Vector Space Similarity (VSS) as the metric to evaluate the similarity between user and user .

The analysis we conduct is to understand how does the social friends similarity compare with random peer similarity? The detailed analysis is as follows:

1. For each user in trust network, we calculate the average social friends’ similarity as follows:

(4)

where is the trusted list of user which also means user ’s outgoing friends.

2. For each user , we also calculate the average random peer similarity as follows:

(5)

where represents the random peer list of user , which has same size with and .

The motivation of carrying out this experiment is that we want to confirm that the social peer relation in trust network has strong positive correlation with user interest similarity. In order to reduce the noise of dataset, we only consider the users who has more than five outgoing social peers in Epinions. For those users who has less than five outgoing social peers, we simply ignore them in the evaluation.

In order to quantify the correlation between social relations and user interest similarity, we would like to measure the proportion of users whose social similarities are greater than their random similarities followed by the equation . In the Epinions dataset, based on my study, there are users whose social similarities are greater than their random similarities. So we can safely make conclusion that friends have similar taste in Epinions.

We also test PCC based similarity measurement. We observe the similar result so for simplicity, therefore we only report the result using VSS similarity function.

5.2 Evaluation Metrics

The quality of the results is measured by the the Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE). The MAE and RMSE are defined as follows:

where denotes the rating user gave to item , is represents the rating user gave to item which is predicted by my proposed method, and is the total number of nonzero ratings in the test dataset.

From the definition, we can see that the smaller MAE and RMSE, the better performance of our proposed method.

5.3 Comparison Methods

Several of state-of-art methods are used for comparison with the proposed approach in this work. The detailed descriptions of comparison methods are explained below:

UserMean: this method uses the mean values of every user to predict the missing values

ItemMean: the method utilizes the mean value of every item to predict the missing values.

BasicMF: Traditional Matrix Factorization and the method uses classic matrix factorization formulation without incorporating user trust network information.

Proposed Method: This method uses classic matrix factorization formulation with trust network information as regularization term which is the method proposed in this work.

The results we obtain are shown in the tables  1 and  2. In the whole process, we use PCC for user user similarity measurement. During the experiment, parameters is set to a trivial value 3.0 and in our model is set to 0.01. For the Epinions dataset, we use 90% and 80% of the original data as the training data settings and the remaining as the test set. The random selection is carried out times independently, and I report the average MAE and RMSE values in table  1 and table  2.

As we can see our proposed method performs much better than baseline methods on Epinion dataset. For instance, traditional matrix factorization method achieves and on MAE and RMSE under training setting respectively, whereas our proposed method achieves and

on MAE and RMSE respectively. And the performance of UserMean and ItemMean methods are even worse than basic matrix factorization method. Cross-validation t-test shows that our proposed method is significantly better (p-value 0.0042 for MAE and 0.0036 for RMSE) than traditional matrix factorization. Similar results are obtained compared with other two baseline methods and also

training setting.

Method MAE RMSE
UserMean 0.9415 1.2361
ItemMean 1.2236 1.7985
MF 0.8641 1.1071
Proposed Method 0.8324 1.0756
Table 1: Performance Comparisons (Dimensionality = 10 and 90% training setting)
Method MAE RMSE
UserMean 0.9545 1.2489
ItemMean 1.2663 1.8575
MF 0.8722 1.1254
Proposed Method 0.8462 1.0964
Table 2: Performance Comparisons (Dimensionality = 10 and 80% training setting)

5.4 Study of Parameter Sensitivity

In our proposed method, the parameter plays the important role for how much our method should incorporate the social network information. Large value of indicates more impact of social connection information in the propsoed model and smaller value of indicates less impact of social connection information and zero value makes the proposed model close to basic matrix factorizaton method. In this experiment, we can see how the performance of model changes as we vary the value of parameter. The result of this experiment is shown in Figure  1(a) and  1(b). As we can see that the performance in terms of both MAE and RMSE degrades for the choice of .

(a) Epinions (MAE)
(b) Epinions (RMSE)
Figure 1: Parameter Sensitivity of with Dimensionality = 10

5.5 Impact of Similarity Function

The similarity function measures how similar of user and in the social network. In the proposed method, two similarity functions PCC and VSS are used and in this section, in order to measure how similarity functions contribute the proposed model, besides PCC and VSS, two other similarity frameworks will be used for comparison:

1. Setting all the similarities between users as 1.

2. Assigning a random similarity between and to any pair of friendship.

The result is shown in the following tables  3 and  4.

Similarity MAE RMSE
Sim=1 0.8415 1.1061
Sim=Random 0.8636 1.1085
Sim=VSS 0.8351 1.0789
Sim=PCC 0.8324 1.0756
Table 3: Similarity Analysis (Dimensionality = 10 and 90% training setting)
Similarity MAE RMSE
Sim=1 0.8515 1.1121
Sim=Random 0.8536 1.1145
Sim=VSS 0.8491 1.0989
Sim=PCC 0.8462 1.0964
Table 4: Similarity Analysis (Dimensionality = 10 and 80% training setting)

As we can see from the result that the performance of unweighted similarity and random similarity measurements are worse than VSS and PCC. Among these four similarity functions, PCC performs best. So this observation demonstrates that the similarity function plays important role in my proposed model.

5.6 Performance on cold start users

In this section we study how performance for our proposed model for the cold start users in recommender systems. Cold start users are users who have rated or purchased few items and traditional collaborative filtering methods are not able to deal with this kind of users since cold start users don’t have enough rating/purchase history. In this experiment, we consider users who have rated less than 5 items as cold start users. In Epinions, more than 55% of users are cold start users. So proposing effective recommender systems for coping with cold start users are becoming more and more important. For conducting this experiment, we only test on the cold start users and put one of their rated items into test set and other rated items into training set. The result is shown in table  5.

Method MAE RMSE
UserMean 1.071 1.502
ItemMean 1.082 1.582
MF 0.982 1.261
Proposed Method 0.892 1.121
Table 5: Performance on cold start users (Dimensionality = 10 and )

As we can see from the result, our proposed method could deal with the cold start users and the performance is much better than other three baseline methods.

6 Conclusion

In this paper, we leverage the social network information in the matrix factorization framework to obtain a better recommender system, especially for the cold start users. The experimental analysis on the real-life dataset shows that our proposed approach performs significantly better than several state-of-the-art methods.

References

  • [1]
  • [2] J. S. Breese, D. Heckerman, and C. Kadie. Empirical analysis of predictive algorithms for collaborative filtering. In

    Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence

    , UAI’98, pages 43–52, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc.
  • [3] J. Canny. Collaborative filtering with privacy via factor analysis. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’02, pages 238–245, New York, NY, USA, 2002. ACM.
  • [4] P.-Y. Chen, B. Zhang, and M. A. Hasan. Incremental eigenpair computation for graph laplacian matrices: theory and applications. Social Network Analysis and Mining, 8(1):4, 2017.
  • [5] P.-Y. Chen, B. Zhang, M. A. Hasan, and A. O. Hero.

    Incremental method for spectral clustering of increasing orders.

    In KDD Workshop on Mining and Learning with Graphs, 2016.
  • [6] S. Choudhury, K. Agarwal, S. Purohit, B. Zhang, M. Pirrung, W. Smith, and M. Thomas.

    NOUS: construction and querying of dynamic knowledge graphs.

    In Proceedings of 33rd IEEE International Conference on Data Engineering, pages 1563–1565, 2017.
  • [7] V. S. Dave, M. A. Hasan, B. Zhang, and C. K. Reddy. Predicting interval time for reciprocal link creation using survival analysis. Social Network Analysis and Mining, 8(1):16, 2018.
  • [8] M. Deshpande and G. Karypis. Item-based top-n recommendation algorithms. ACM Trans. Inf. Syst., 22(1):143–177, Jan. 2004.
  • [9] M. Dundar, Q. Kou, B. Zhang, Y. He, and B. Rajwa.

    Simplicity of kmeans versus deepness of deep learning: A case of unsupervised feature learning with limited data.

    In

    Proceedings of 14th IEEE International Conference on Machine Learning and Applications

    , pages 883–888, 2015.
  • [10] J. L. Herlocker, J. A. Konstan, A. Borchers, and J. Riedl. An algorithmic framework for performing collaborative filtering. In Proceedings of the 22Nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’99, pages 230–237, New York, NY, USA, 1999. ACM.
  • [11] M. Jamali and M. Ester. A matrix factorization technique with trust propagation for recommendation in social networks. In Proceedings of the Fourth ACM Conference on Recommender Systems, RecSys ’10, pages 135–142, New York, NY, USA, 2010. ACM.
  • [12] S. Kabbur, X. Ning, and G. Karypis. Fism: Factored item similarity models for top-n recommender systems. In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’13, pages 659–667, New York, NY, USA, 2013. ACM.
  • [13] H. Ma, I. King, and M. R. Lyu. Effective missing data prediction for collaborative filtering. In Proceedings of the 30th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’07, pages 39–46, New York, NY, USA, 2007. ACM.
  • [14] H. Ma, I. King, and M. R. Lyu. Learning to recommend with social trust ensemble. In Proceedings of the 32Nd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’09, pages 203–210, New York, NY, USA, 2009. ACM.
  • [15] H. Ma, D. Zhou, C. Liu, M. R. Lyu, and I. King. Recommender systems with social regularization. In Proceedings of the Fourth ACM International Conference on Web Search and Data Mining, WSDM ’11, pages 287–296, New York, NY, USA, 2011. ACM.
  • [16] T. K. Saha, B. Zhang, and M. Al Hasan. Name disambiguation from link data in a collaboration graph using temporal and topological features. Social Network Analysis Mining, 5(1):11:1–11:14, 2015.
  • [17] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl. Item-based collaborative filtering recommendation algorithms. In Proceedings of the 10th International Conference on World Wide Web, WWW ’01, pages 285–295, New York, NY, USA, 2001. ACM.
  • [18] B. Zhang. Towards Name Disambiguation: Relational, Streaming, and Privacy-Preserving Text Data. PhD thesis, Purdue University, West Lafayette, 2017.
  • [19] B. Zhang, S. Choudhury, M. A. Hasan, X. Ning, K. Agarwal, S. Purohit, and P. G. P. Cabrera. Trust from the past: Bayesian personalized ranking based link prediction in knowledge graphs. In SDM Workshop on Mining Networks and Graphs, 2016.
  • [20] B. Zhang, M. Dundar, and M. A. Hasan. Bayesian non-exhaustive classification a case study: Online name disambiguation using temporal record streams. In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management, pages 1341–1350. ACM, 2016.
  • [21] B. Zhang, M. Dundar, and M. A. Hasan. Bayesian non-exhaustive classification for active online name disambiguation. arXiv preprint arXiv:1702.02287, 2017.
  • [22] B. Zhang and M. A. Hasan. Name disambiguation in anonymized graphs using network embedding. In Proceedings of the 26th ACM International on Conference on Information and Knowledge Management, 2017.
  • [23] B. Zhang, N. Mohammed, V. S. Dave, and M. A. Hasan. Feature selection for classification under anonymity constraint. Transactions on Data Privacy, 10(1):1–25, 2017.
  • [24] B. Zhang, T. K. Saha, and M. Al Hasan. Name disambiguation from link data in a collaboration graph. In IEEE/ACM 5th International Conference on Advances in Social Networks Analysis and Mining, pages 81–84.
  • [25] M. Zhang, J. Tang, X. Zhang, and X. Xue. Addressing cold start in recommender systems: A semi-supervised co-training algorithm. In Proceedings of the 37th International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR ’14, pages 73–82, New York, NY, USA, 2014. ACM.