TCNSVD: A Temporal and Community-Aware Recommender Approach

Abstract

Recommender systems support users in €nding relevant items in overloaded information spaces. Researchers and practitioners have proposed many di‚erent collaborative €ltering algorithms for different information scenarios, domains and contexts. One of the laŠer, are time-aware recommender methods that consider temporal dynamics in the users' interests in certain items, topics, etc. While there is extensive research on time-aware recommender systems, surprisingly, researchers have paid liŠle aŠention to model temporal community structure dynamics (community dri‰). In consequence, recommender systems seldom exploit explicit and implicit community structures that are present in online systems, where one can see what others have been watching, sharing and or tagging. In this paper, we propose a recommender method that not only considers temporal interest dynamics in online communities, but also exploits the social structure by the means of community detection algorithms. We conducted o„ine experiments on the Netƒix dataset and the latest MovieLens dataset with tag information. Our method outperformed the current state-of-the-art in rating and item-ranking prediction. Œis work contributes to the connection of two separate recommender research directions, in which exploits community structure and temporal e‚ects together in recommender systems.

INTRODUCTION

Since many years, recommender systems based on collaborative €ltering techniques provide recommendations for us by applying speci€c approaches on a huge rating matrix. However, it is expensive to create and maintain such huge rating matrices for online shops and rating websites. From our own experience, we know that we do not update ratings when our judgment has changed due to changes in our taste, nor do we reƒect that our ratings are based on the inƒuence of somebody we know. But, research has shown that we can increase the recommendation accuracy by taking into account temporal e‚ects with computational methods e.g. changes in the user preferences or the item popularity over time. Œis has lead to the development of time-aware recommender systems [ 21 ]. In parallel, the social network research community has investigated the detection and analysis of community structures as implicit inƒuence among members of a social network [ 14 ]. Members of a community are supposed to possess similar properties so that they form dense connections inside communities and sparse connections among them. Correspondingly, more and more recommender systems consider community structures to e.g. improve accuracy [ 9 ]. However, one important property of social network research is still missing in recommender research, namely the temporal dynamics of online community structures [ 1, 22 ]. Temporal community structures - detected from explicit and implicit users’ interactions and item-item networks - provide dynamics of collective information carried by groups of people. Œese communities are dynamic similar to the network, in which this needs to be reƒected in recommender systems. To the best of our knowledge, there is still no other work that explains to what extent temporal dynamics of online communities can be e‚ective in the proposal of a recommender system.

Objective. In this line of research, we propose two recommender models named CNSVD and TCNSVD. CNSVD considers the collective user preferences and item receptions at the same time. TCNSVD, on the other hand, includes temporal dynamics of (overlapping) community structures, which is not yet addressed by the research community. Œese models are extensions to the NSVD and TNSVD models proposed by Koren [ 20, 21 ] - leaning upon neighborhood and factor models of recommendation. Using user-user and item-item networks contributes to the evaluation of (overlapping) community detection algorithms as well as the graph construction methods. Furthermore, we perform extensive studies of the proposed models on two large-scale and popular datasets - MovieLens and Netƒix - and compare them with the state-of-the-art approaches. 2

RELATED WORK

In this paper, the proposed algorithms are related to both timeaware and community-aware recommender models. As such the related work in the area is shortly reviewed.

Time-Aware Recommendation. Some research has been done in the proposal of recommender systems dealing with temporal e‚ects in recent years. Campos et al. [ 5 ] presented a survey and analysis of time-aware recommender systems. Œey claimed that time is one of the most useful contextual dimensions in recommender systems. Koren [ 21 ] applied a model-based collaborative €ltering approach using a combination of neighborhood and factor models. He used models that track temporal shi‰s over several relevant characteristics e.g. user and item biases, covering both long-term and short-term temporal e‚ects. Daneshmand et al. [ 10 ] assumed a hidden item network structure that can be inferred from users’ sequences of selecting items. Œe basic idea is that if two items in the hidden network are related, then a user selecting one of the two items is likely to also select the other item. Baltrunas and Amatriain [ 2 ] proposed a slightly di‚erent approach to time dependency in recommendation, and assumed time-changing but repetitive user preferences. Œey recommended music, which depends on the time of day, week or year using a collaborative €ltering approach. Charlin et al. proposed a dynamic matrix factorization model based on Poisson factorization for recommendation, which considers temporal users’ interests and item popularity [ 8 ].

Community-Aware Recommendation. Œere has also been work on using community detection for recommendation tasks. Dolgikh and Jelinek [ 11 ] proposed an approach to recommend music using community detection. Œey constructed an artist-tag network for each user from the user’s favorite artists and the tags assigned to these artists. Community detection was applied on these networks to determine each user’s interest sub€eld, which were then used for recommending artists to the user. Cao et al. [ 7 ] applied a neighborhood-based collaborative €ltering approach for recommending movies to users. Œey reduced computation time and improved recommendation precision by using community structures. Œey applied a community detection algorithm on the network constructed from similarity among users which was derived from the user-item ratings. Choo et al. [ 9 ] proposed a neighborhood-based collaborative €ltering approach that uses user communities. Œe user network was derived from reviewreply paŠerns, i.e. if one user reviews an item and another user replies to that review then there may be a relation between the two users. User communities were derived from this network and used as a basis for the recommendation process. Bellogin and Parapar [ 3 ] constructed a user graph using Pearson correlation similarity and applied normalized graph cuts to €nd clusters of users. Œese clusters were then used for neighbor selection in user-based collaborative €ltering. Other approaches using community structures alleviated the cold-start problem in collaborative €ltering. [ 25 ] and [ 30 ] both addressed the cold-start problem for new users by taking into account additional user information.

Summary. To the best of our knowledge, there are no approaches that take into account temporal dynamics of community structures to support recommendation. Œus, the literature manifests that 1) our knowledge regarding performance of (overlapping) community structures on recommendation systems is imperceptible. 2) we are not aware about the goodness of graph construction similarity metrics, e.g. Jaccard, Cosine, etc, in time-evolving recommender models. 3) we also know very liŠle regarding the e‚ect of metadata information on graph construction in temporal community-aware recommender systems. 4) we need models to connect temporal dynamics with (overlapping) community structures to improve item ranking and precision accuracy metrics in recommender systems. 3

PROPOSED RECOMMENDER MODELS

In this section, we introduce the proposed recommender models. Koren [ 20 ] employed the neighborhood and factor models to compute the rating of a speci€c user on a particular item. In this model, weights of user or item similarities are interpreted as o‚sets that need to be added to a baseline estimation. In other words, this approach combines three components including a baseline estimation, a neighborhood and a factor model, in which can be wriŠen as follows:

1 X ((ru j μ + bu + bi + jIu j 2 where the €rst block of terms refers to the baseline estimation, in which μ is the average rating over all users and items and bu is the user bias, i.e. the deviation of the average rating given by user u from μ. Besides, bi is the deviation of the average rating given to item i from μ (item bias). the second block of terms refers to the neighborhood model contribution, in which Iu is the set of items rated by user u, wi j relates to explicit rating feedback, which is multiplied by ru j bu j , and ci j is related to implicit feedback and is added whenever user u has given a rating to item j. In addition, to avoid overestimation of the rating of users who provide much feedback, i.e. jIu j is 1 high, the estimation is scaled down by multiplying with jIu j 2 . €nally the last block of terms indicates the factor model contribution, in which qi and pu are vectors characterizing item i and user u, respectively. Moreover, the user preference vector u is complemented by a sum of vectors yj , that represent implicit feedback from each item j 2 Iu . Œis model is named Neighborhood-Integrated SVD (NSVD), in which parameters, i.e. bu , bi , wi j , ci j , yi j , qi and pu are learned by minimizing a squared error function. Koren extended the NSVD model by using temporal information to improve recommendation accuracy. Œe temporal information allows the modeling of user preferences and item characteristics that change over time [ 21 ]. Temporal information was included into each of the three components i.e., baseline estimation, neighborhood model and factor (2) (3) (4) (5) (6) where we sum over all communities that user u belongs to, in other words, we add the corresponding biases bC weighted by the user’s membership level muC regarding each community. b Ci refers to the community bias for item i with Ci as the set of item communities that i belongs to. Similarly, we de€ne the item community bias as follows: b Ci =

X C 2 Ci bC miC ; where bC shows the community bias of community C, in which miC represents the membership level of item i belonging to community C.

Neighborhood Model. To use user community information for the neighborhood model, we extend the original neighborhood model to capture additional implicit feedback from each item that has been rated by a member of one of the user’s communities. Œe extension is as follows:

1 X ((ru j jIu j 2 j 2ICu X

C 2 Ci where the block of terms shows the community contribution in the neighborhood model. Here, I Cu represents the set of items that any user belonging to one of user u’s communities has rated. Œen, di j shows the o‚set that such an item j contributes to the rating for item i.

Latent Factor Model. For the latent factor model, we again consider the community information as implicit feedback, in which each item rated by a user’s community contributes to the user’s preference vector pu . Using the original factor model, we can extend it as follows: (qi + oC miC +)T (pu + oC

muC + + X

C 2 Cu jIu j 12 X yj + jI Cu j 21 X zj ); j 2ICu where the vector zj represents the contribution from item j. Also, to model the user’s communities, we introduce an additional vector oC that represents the preferences or characteristics of community C. Since a user can belong to multiple communities, we de€ne o Cu as the combined community preferences of all communities that user u belongs to. Œis is achieved by summing the community factors oC weighted by the user’s membership level muC to each community. Likewise, to model the item community characteristics, we de€ne o Ci as the combined characteristics of the communities that item i belongs to. Combining baseline estimation, neighborhood model and factor model from equations 5, 8 and 9, we compute the predicted rating rˆui from a user u to an item i. 3.2

Time and Community-Aware NSVD (TCNSVD) Model Œe TCNSVD model is an extension of the TNSVD model that uses temporal dynamics of community structures. To capture user and item community dri‰, i.e. time-changing community structures, models as follows:

1 X e ϕu jt tu j j ((ru j + jIu j 2 j 2Iu μ + bu (t ) + bi (t ) where the €rst block of terms refers to the time-aware baseline estimation. bu (t ) and bi (t ) indicate the time-dependent user and item bias at time t , in which bu (t ) can be computed as with bu + αu devu (t ) + bu;t . Here, bu , αu devu (t ) and bu;t describe time-independent user bias, the linear dri‰ of the user bias and a time-speci€c parameter capturing short-lived e‚ects. Moreover, item characteristics are less complex to be described and thus bi (t ) can be simply computed by bi + bi;Bin(t ) , in which short-lived e‚ects are captured through time by bi;Bin(t ) . the second block of terms describes the time-aware neighborhood model contribution where the function e ϕu jt tj j decays the item contributions wi j and ci j such that ratings that are more distant to time t have less impact on the estimation. the last block of terms indicates the contribution of the timeaware factor model. pu (t ) represents the time-dependent user preferences that can be captured through parameters including time-independent preference, gradual and short-lived e‚ects. Œis model is known as Time-aware Neighborhood-Integrated SVD (TNSVD), in which parameters, i.e. bu , bu;t , bi , bi;Bin(t ) , wi j , ci j , yi j , αu , αuk , ϕu , qi , puk and puk;t are again learned by minimizing a squared error function. In the following subsections, we introduce two models based on NSVD and TNSVD models. 3.1

Community-Aware NSVD Model

Œe community-aware neighborhood-integrated SVD (CNSVD) model is an extension of the NSVD model that use community information to improve rating prediction accuracy. Similar to the NSVD model, it consists of baseline estimation, neighborhood and factor model contributions.

Baseline Estimation. For the baseline estimation, we presume that user and item communities have their own bias. Œe average rating of a user community tends to deviate from the overall average rating and each user’s average rating tends to deviate from the community’s rating, and likewise with item communities. Œis is based on the assumption that users or items in a community have similar preferences or characteristics, which may imply a common bias and thus the baseline estimation bui , is extended as follows: μ + bu + b Cu + bi + b Ci ; , in which b Cu shows the community bias for user u, in which Cu represents the set of user communities that user u belongs to. We model the user community bias as follows: b Cu =

X bC muC ; (7) (9) we compute user and item graphs and their community structures for several time ranges. Here, available time range is divided into community time bins. Œe set of communities of a user u at time t is represented by Cu;CBin(t ) and the corresponding set of item communities is shown by Ci;CBin(t ) , where CBin(t ) indicates a function that returns the community time bin for a given time t . TCNSVD model has three components that are explained in the following.

Baseline Estimation. For the baseline estimation, we extend the Equation 5 to capture time-dependent community biases and a time-dependent linear dri‰ in community biases and thus we write the baseline estimation for TCNSVD model as follows: μ + bu + αu devu (t ) + bu;t + bu;Period(t) + bi + bi;Bin(t ) + bi;Period(t) (bC + bC;t ) muC + αC

devC (t ) muC + X C 2 Cu + vector function, oC (t ) is made up of multiple functions, each representing a component of the vector, i.e. oC (t )T = (oC1 (t ); oC2 (t ); : : : ; oCn (t )). Each component is de€ned as: oCk (t ) = oCk + αCk devC (t ) + oCk;t k = 1; : : : ; n; (13) where oCk is the time-independent part of the community preferences, oCk;t captures short-term e‚ects and αCk devC (t ) models a linear shi‰ of community preferences. Using the user community preferences, we extend the formula for the factor model as follows: rˆui (t ) = (qi +o Ci )T (pu (t )+o Cu (t )+jIu j 21 X yj +jI Cu j 12 X zj ): (14) Koren [ 21 ] does not make the item vector qi time-dependent. He claims that item characteristics are inherent and do not change with time. We expect that this also applies to the item community vector, so we also leave it to be time-independent. Finally, we combine the baseline estimation, the neighborhood model and the factor model by summing their predictions.

For all the models, a squared error function is minimized to learn the parameters of the model e.g. regularization parameters, using stochastic gradient descent. A regularization factor λ is used to penalize high parameter values to avoid over€Šing the training data. An implementation is included in LibRec 1 and we adapt it to our NSVD-based models [ 15 ]. We performed some validations on holdout data and found the optimum learning rate and regularization parameters for the NSVD, TNSVD, CNSVD and TCNSVD models using RMSE error. We selected the learning rate decay strategy from LibRec and used the same combination of learning rate and regularization for parameters of each model. Œe best learning rates and regularization parameters for each model are given in Table 1. j 2ICu 4

EVALUATION

Regarding experiments, we use the popular Netƒix (NF) and the latest MovieLens (ML) dataset to evaluate the proposed recommendation algorithms. Œe basic statistics of the datasets are shown Table 2. In MovieLens both ratings and tags information are available. As for tags-based graph construction, an edge is created between any two users that have used the same tag on any item. Similarly, an edge is created between any two items that have received the same tag from any user [ 16 ]. Regarding MovieLens and Netƒix graph construction based on rating information, we apply k-NN proposed by Park et al. as an approach mainly suitable for information retrieval and recommender algorithms [ 23 ]. As such, to compute the similarities between users and items from the rating information, we use similarity measures such as Pearson Correlation, Cosine Similarity and Jaccard Mean Squared Distance (JMSD) [ 4, 6, 17, 19, 28 ]. 1http://www.librec.net/ in which the block of terms shows the community contributions. In this formula, the time-independent bias of community C is denoted as bC , and community bias based on short-lived e‚ects at time t is denoted as bC;t for user communities and bC;CBin(t ) for item communities. Linear dri‰ in community biases is captured by αC devC (t ). All community biases are summed over the respective set of communities and weighted by the respective user’s or item’s community membership level. For the time-independent biases and the short-lived temporal e‚ects, we use the dynamic community structures. However, for the linear dri‰ we use static community structures so that longer-term temporal e‚ects on each community can be captured. In addition, the periodic user and item biases are also reƒected by bu;Period(t) and bi;Period(t) , where Period(t) represents a function that returns an index showing the week day of time t . To keep our model from geŠing overly complex, we capture only one of these potential recurring temporal e‚ects, namely weekly recurring e‚ects.

Neighborhood Model. For the neighborhood model, we use a decay function e ψu jt tj j on the additional implicit feedback di j , which was added to Equation 8. Œis leads to the following formula: 1 X e ϕu jt tj j ((ru j jIu j 2

bu j )wi j + ci j )+ j 2Iu

Latent Factor Model. For the latent factor model, we change the vector modeling the user community preferences o Cu , which was introduced in Equation 8 to being a function o Cu (t ): o Cu (t ) =

X oC (t ) ; muC ;

C 2 Cu with the community preference vector oC being replaced by the time-dependent vector function oC (t ). As with the user preferences E 1:05 SM 1 R 0:95

For the evaluation of our proposed recommender model, we compute both rating prediction and the item ranking quality as well as accuracy. For measuring the accuracy of rating predictions, we used the Root Mean Squared Error (RMSE). For measuring the accuracy of item rankings, we used the measures Precision at k (Prec@k), Recall at k (Rec@k) and Normalized Discounted Cumulative Gain (NDCG) [ 18, 27 ]. Finally, we use Wilcoxon Rank Sum tests to identify statistical di‚erences of the results generated by the models [ 29 ]. Evaluation of time-aware recommender algorithms is challenging since the time ordering of the ratings needs to be considered and thus the use of cross validation approaches are not suitable. Campos et al. [ 5 ] describe in detail the issues regarding time-aware recommender systems in their 2014 UMUAI paper. Instead of k-fold cross-validation, we applied a sliding-window approach taking snapshots along the timeline [ 12 ]. Taking a snapshot at time t means using the ratings within a certain number of days before t for training and the ratings within a certain number of days a‰er t for testing. In total we took €ve of these snapshots over the timeline in each dataset and report the overall means in the results section. Regarding complexity, the running times of the proposed methods are more than the baselines, in which we used a subsampled version of datasets on a compute cluster. We considered a maximum running time of €ve days, in which TCNSVD and CNSVD models had higher running times compared to NSVD and TNSVD models. As for future works, we plan to perform time complexity analysis of the models, and ignore individual learning parameters to €nd a compromise between accuracy and time complexity.

5 RESULTS

In the following section, we present the results of our o„ine simulations. We did preliminary experiments on the current state-of-theart community detection algorithms regarding time complexity and number of found communities. We chose DMID [ 26 ] and Walktrap [ 24 ] as two alternatives that can identify overlapping and disjoint communities. Œey not only scale well on large amount of data but can also handle weighed and directed networks. To use Walktrap, a step size input parameter needs to be set that was 2 and 5 in our case. Figure 1 presents the results with respect to RMSE, Prec@10 and Rec@10 for the NSVD, TNSVD, CNSVD and TCNSVD models on the MovieLens dataset (for space reasons we omiŠed the Netƒix results here which show similar tendencies). As shown, the RMSE values are quite close to each other for both of the two community detection algorithms investigated. However, as also shown, Walktrap performs slightly beŠer compared to the DMID algorithm. Moreover regarding RMSE, algorithms based on the TCNSVD model - using tags and ratings constructions - achieve higher values than the CNSVD model. Œis trend also holds for precision as well as recall. In general though, the results show that Walktrap (WT5) yields the beŠer performance regarding RMSE, CNSVD-tags TCNSVD-tags CNSVD-ratings TCNSVD-ratings CNSVD-tags TCNSVD-tags CNSVD-ratings TCNSVD-ratings

CNSVD-tags TCNSVD-tags CNSVD-ratings TCNSVD-ratings recall as well as precision. Œe only exception is TCNSVD recall results based on tags and ratings, in which DMID slightly outperforms Walktrap versions. To verify the overall performance of the proposed algorithms, online user studies need to be done.

Figure 2 illustrates how the models perform when using di‚erent similarity metrics using the Netƒix dataset (we omit for space reasons the results of the MovieLens dataset, which though are comparable). k was set to 10 in this case. As shown, there are observable di‚erences with respect to the measures chosen for both models. In general we can observe the following: Pearson achieves in four cases the best results for RMSE, Prec@10 and Rec@10, followed by Cosine and JMSD. Œis paŠern is also emerging when testing with di‚erent ks and the di‚erent similarity metrics at the same time. Table 3 shows the best performing parameters for the ratings-based graph construction regarding RMSE, precision and recall and the best performing parameters for k. To €gure out the practical performance of the models and similarity metrics, we need to deploy them online and study users’ feedback.

To illustrate how the models perform compared to some baselines, a series of experiments have been performed. First, the models were compared with the baseline methods NSVD and TNSVD on the MovieLens (ML) and Netƒix (NF) datasets. Œerea‰er, we compared them to current state-of-the-art item-ranking models such as WRMF and ItemKNN as present in the LibRec library.

Table 4 shows the €rst set of results in this respect. In general we can observe that the TCNSVD model achieves the best results here. For instance in the MovieLens dataset, compared to its baseline (TNSVD), the method increases NDCG, Precision and Recall with 45.45 %, 675 % and 443 % in the best case relying on a ratings-based graph. Similar trends are also observed for the Netƒix dataset. Here, TCNSVD can also improve NDCG, Prec@10 and Rec@10 with 15.67 %, 126.73 % and 115.90 %, compared to it’s baseline method. Œe result “paŠerns” on the other hand for the CNSVD model are not that clear as RMSE and NDCG values are sometimes decreased, while Prec@10 and Rec@10 are not. Œis is actually an interesting behaviour which we need study further in future work.

Figure 3 provides an overview of the computed item ranking and precision metrics. Regarding the MovieLens dataset and the NDCG metric, the TCNSVD model could achieve the best results, that is statistically beŠer than the baselines WRMF, ItemKNN, CNSVD, NSVD and TNSVD. TCNSVD surpasses the other baselines also for Prec@10 and Rec@10 with even higher di‚erences. As for Prec@10, TCNSVD achieves 0.28359, which is higher than 0.18842 and 0.17433 as obtained by WRMF and ItemKNN. Œe results and the ranking of the methods are consistent with other benchmarks run on the MovieLens dataset, although our evaluation protocol is time-based [ 13 ]. As for the Netƒix dataset, again TCNSVD gives us the best results with notable statistically signi€cant di‚erences to the other models.

6 SUMMARY & FUTURE WORK

Œe main €ndings of the paper can be summarized as follows: We proposed a community-aware model named CNSVD based on neighborhood and factor models of recommendation. Furthermore, we introduced TCNSVD as a model that considers temporal community structure and dynamics.

We showed the e‚ect of community detection algorithms on the recommendation performance and found that Walktrap is the beŠer option. 10 00::0034 ceR@ 00::0012 0 0:55 CGD 00:4:55 N 0:4 0:35

ML ML ML

Also it was shown that Pearson correlation as the similarity metric for graph construction achieves the best performance when considering temporal dynamics of community structures in the recommendation task.

Finally, we show that TCNSVD as a temporal and communityaware recommender model performs sign beŠer than CNSVD and compared state-of-the-art baseline recommendation approaches on the MovieLens and Netƒix datasets.

One limitation of the proposed CNSVD and TCNSVD models are their high dimensionality. As such, it is planned for future work to improve the models in such a way that less parameters need to be set, e.g. by omiŠing individual user and item parameters. In our experiments we selected the optimum learning rate and regularization based on RMSE. Future work needs to assess whether further improvements can be found when optimizing the models e.g. with respect to NDCG.

Although temporal dynamics of overlapping community structures are considered by TCNSVD, modelling the e‚ect of community life cycles, i.e. birth, death, atrophy, grow, split and merge, on recommendation systems still need to be studied. Moreover, we plan to study the impact of time bins as well as speed of community changes in the proposal of a recommender system. In addition, more community detection algorithms need to be investigated with our models, to identify best performing ones. Finally, we plan to investigate the e‚ect of explicit community structures as the current ones are based on implicit ones and already achieving remarkable results.

ACKNOWLEDGMENTS

Œis work is in part supported by BIT research school.

References

[1] Backstrom , L. , Huttenlocher , D. , Kleinberg , J. M. , and Lan , X. Group formation in large social networks: Membership, growth, and evolution . In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining - KDD '06 ( 2006 ), ACM Press, pp. 44 - 54 .

[2] Baltrunas , L. , and Amatriain , X. Towards time-dependant recommendation based on implicit feedback . In Workshop on context-aware recommender systems ( 2009 ).

[3] Bellogin , A. , and Parapar , J. Using graph partitioning techniques for neighbour selection in user-based collaborative €ltering . In Proceedings of the 6th ACM conference on recommender systems ( 2012 ), ACM, pp. 213 - 216 .

[4] Bobadilla , J. , Ortega , F. , Hernando , A. , and Gutie´rrez, A. Recommender systems survey . Knowledge-Based Systems 46 ( 2013 ), 109 - 132 .

[5] Campos , P. G. , D´ez, F., and Cantador , I. Time-aware recommender systems: A comprehensive survey and analysis of existing evaluation protocols . User Modeling and User-Adapted Interaction 24 , 1 - 2 ( 2014 ), 67 - 119 .

[6] Candillier , L. , Meyer , F., and Boulle´, M. Comparing state-of-the-art collaborative €ltering systems . In Proceedings of the 5th international conference on machine learning and data mining in paˆern recognition ( 2007 ), Springer, pp. 548 - 562 .

[7] Cao , C. , Ni , Q. , and Zhai , Y. An improved collaborative €ltering recommendation algorithm based on community detection in social networks . In Proceedings of the 2015 annual conference on genetic and evolutionary computation ( 2015 ), ACM, pp. 1 - 8 .

[8] Charlin , L. , Ranganath , R. , McInerney , J. , and Blei , D. M. Dynamic poisson factorization . In Proceedings of the 9th ACM Conference on Recommender Systems (New York, NY, USA, 2015 ), RecSys '15 , ACM, pp. 155 - 162 .

[9] Choo , E. , Yu , T. , Chi , M. , and Sun , Y. Revealing and incorporating implicit communities to improve recommender systems . In Proceedings of the 15th ACM conference on economics and computation ( 2014 ), ACM, pp. 489 - 506 .

[10] Daneshmand , S. M. , Javari , A. , Abtahi , S. E. , and Jalili , M. A time-aware recommender system based on dependency network of items . Še Computer Journal 58 , 9 ( 2015 ), 1955 - 1966 .

[11] Dolgikh , D. , and Jel´nek, I. Graph-based music recommendation approach using social network analysis and community detection method . In Proceedings of the 16th international conference on computer systems and technologies ( 2015 ), ACM, pp. 221 - 227 .

[12] Eberhard , L. , and Trattner , C. Recommending sellers to buyers in virtual marketplaces leveraging social information . In Companion of the 25th international conference on World Wide Web ( 2016 ), ACM, pp. 559 - 564 .

[13] Gantner , Z. , Rendle , S. , Freudenthaler , C. , and Schmidt-Thieme , L. MyMediaLite: A free recommender system library . In Proceedings of the 5th ACM Conference on Recommender Systems (RecSys 2011 ) ( 2011 ).

[14] Gauvin , L. , Panisson , A. , and Cattuto , C. Detecting the community structure and activity paŠerns of temporal networks: A non-negative tensor factorization approach . PLoS ONE 9 , 1 ( 2014 ).

[15] Guo , G. , Zhang, J., Sun , Z. , and Yorke-Smith , N. Librec: A java library for recommender systems . In Posters, demos, late -breaking results and workshop proceedings of the 23rd conference on user modeling, adaptation and personalization ( 2015 ).

[16] Helic , D. , Strohmaier , M. , Trattner , C. , Muhr , M. , and Lerman , K. Pragmatic evaluation of folksonomies . In 20th International World Wide Web Conference (WWW2011) , Hyderabad, India, March 28 - April 1, ACM ( 2011 ).

[17] Herlocker , J. L. , Konstan , J. A. , Borchers , A. , and Riedl , J. An algorithmic framework for performing collaborative €ltering . In Proceedings of the 22nd annual international ACM conference on research and development in information retrieval ( 1999 ), ACM, pp. 230 - 237 .

[18] Herlocker , J. L. , Konstan , J. A. , Terveen , L. G. , and Riedl , J. T. Evaluating collaborative €ltering recommender systems . Transactions on Information Systems 22 , 1 ( 2004 ), 5 - 53 .

[19] Jesu´ s Bobadilla, J. , Serradilla , F. , and Bernal , J. A new collaborative €ltering metric that improves the behavior of recommender systems . Knowledge-Based Systems 23 , 6 ( 2010 ), 520 - 528 .

[20] Koren , Y. Factorization meets the neighborhood: A multifaceted collaborative €ltering model . In Proceedings of the 14th ACM international conference on knowledge discovery and data mining ( 2008 ), ACM, pp. 426 - 434 .

[21] Koren , Y. Collaborative €ltering with temporal dynamics . In Proceedings of the 15th ACM international conference on knowledge discovery and data mining ( 2009 ), ACM, pp. 447 - 456 .

[22] Leskovec , J. , Kleinberg , J. , and Faloutsos , C. Graphs over time: Densi€cation laws, shrinking diameters and possible explanations . In Proceedings of the Eleventh ACM SIGKDD International Conference on Knowledge Discovery in Data Mining (New York, NY, USA, 2005 ), KDD '05 , ACM, pp. 177 - 187 .

[23] Park , Y. , Park , S. , Lee , S.-g., and Jung , W. Greedy €ltering: A scalable algorithm for k-nearest neighbor graph construction . In Proceedings of the 19th international conference on database systems for advanced applications ( 2014 ), Springer, pp. 327 - 341 .

[24] Pons , P. , and Latapy , M. Computing communities in large networks using random walks . In Proceedings of the 20th international symposium on computer and information sciences ( 2005 ), Springer, pp. 284 - 293 .

[25] Sahebi , S. , and Cohen , W. W. Community-based recommendations: A solution to the cold start problem . In Proceedings of the 3rd workshop on recommender systems and the social Web ( 2011 ), ACM, pp. 40 - 44 .

[26] Shahriari , M. , Krott , S. , and Klamma , R. Disassortative degree mixing and information di‚usion for overlapping community detection in social networks (DMID) . In Proceedings of the 24th international conference on World Wide Web ( 2015 ), ACM, pp. 1369 - 1374 .

[27] Shani , G. , and Gunawardana , A. Evaluating recommendation systems . In Recommender Systems Handbook . Springer, 2011 , pp. 257 - 297 .

[28] Su , X. , and Khoshgoftaar , T. M. A survey of collaborative €ltering techniques . Advances in Arti€cial Intelligence ( 2009 ), 421425 : 1 - 421425 : 19 .

[29] Trattner , C. , Kowald , D. , Seitlinger , P. , Ley , T. , and Kopeinik , S. Modeling activation processes in human memory to predict the use of tags in social bookmarking systems . Š e Journal of Web Science 2 , 1 ( 2016 ), 1 - 16 .

[30] Yanxiang , L. , Deke , G. , Fei , C. , and Honghui , C. User-based clustering with topn recommendation on cold-start problem . In Proceedings of the 3rd international conference on intelligent system design and engineering applications ( 2013 ), IEEE, pp. 1585 - 1589 .