PaperlayerView exact revision on arXiv

Source-linked AI summary

Hete-CF: Social-Based Collaborative Filtering Recommendation using Heterogeneous Relations

Chen Luo, Wei Pang, Zhe Wang

arXiv:1412.7610v1cs.IR

TL;DR

Collaborative filtering faces data sparsity and cold-start problems, and many social recommenders use only part of the relations in a social network. Hete-CF models user-user, item-item, and user-item relations together; experiments on DBLP and Meetup demonstrate its effectiveness and efficiency.

  • Problem

    Collaborative filtering suffers from data sparsity and cold start, while many social recommendation methods use only user-user or item-item information.

    Source: §I · p. 1

  • Method

    Hete-CF models user-user, user-item, and item-item relations in a unified recommendation model and weights the introduced social information.

    Source: §I · p. 2

  • Results

    Experiments on DBLP and Meetup demonstrate the effectiveness and efficiency of Hete-CF.

    Source: C. Result and Analysis · p. 7, §VI · p. 10

  • Takeaways & Limitations

    Hete-CF is presented as a general approach for heterogeneous social networks, including LBSNs and EBSNs.

    Source: §I · p. 2, §VI · p. 9

Abstract

from arXiv · show

Collaborative filtering algorithms haven been widely used in recommender systems. However, they often suffer from the data sparsity and cold start problems. With the increasing popularity of social media, these problems may be solved by using social-based recommendation. Social-based recommendation, as an emerging research area, uses social information to help mitigate the data sparsity and cold start problems, and it has been demonstrated that the social-based recommendation algorithms can efficiently improve the recommendation performance. However, few of the existing algorithms have considered using multiple types of relations within one social network. In this paper, we investigate the social-based recommendation algorithms on heterogeneous social networks and proposed Hete-CF, a Social Collaborative Filtering algorithm using heterogeneous relations. Distinct from the exiting methods, Hete-CF can effectively utilize multiple types of relations in a heterogeneous social network. In addition, Hete-CF is a general approach and can be used in arbitrary social networks, including event based social networks, location based social networks, and any other types of heterogeneous information networks associated with social information. The experimental results on two real-world data sets, DBLP (a typical heterogeneous information network) and Meetup (a typical event based social network) show the effectiveness and efficiency of our algorithm.

I. INTRODUCTION

Collaborative filtering faces data sparsity and cold-start problems, while existing social recommenders often use only some network relations. Hete-CF combines user-user, user-item, and item-item relations in a unified model for heterogeneous social networks.

I. INTRODUCTION

  • Collaborative filtering suffers from data sparsity and cold start, and many social recommenders use only user-user or item-item information.The paper motivates using more of the available social information to improve recommendation performance.Source: §I · p. 1
  • Hete-CF combines user-user, item-item, and user-item relations in one model, with a weighting method for the user-item relation.The method evaluates the weight of social information because it may bias recommendations when historical ratings are not very sparse.Source: §I · p. 2

A. Heterogeneous Social Network

  • Heterogeneous social networks include social-information-bearing information networks, such as bibliographic networks, EBSNs, and LBSNs.The paper uses meta-paths to represent relations; examples include Author−Paper−Author for co-authorship and Author−Paper−Conf.−Paper−Author for co-conference relations.Source: §I · pp. 1–2, A. Heterogeneous Social Network · p. 2

B. Collaborative Filtering Model

Collaborative filtering uses historical user-item ratings, representing users and items with lower-dimensional vectors learned by minimizing a loss function. The predicted rating is calculated from the user and item vectors.

B. Collaborative Filtering Model

  • Collaborative filtering uses historical user-item ratings, represented in an n-by-m matrix whose entry records a user's rating for an item.The paper maps users and items to lower-dimensional feature vectors and learns them by minimizing a loss function.Source: B. Collaborative Filtering Model · p. 3
  • The rating prediction is the inner product of a user's feature vector and an item's feature vector.The vectors are the low-rank representations of users and items.Source: B. Collaborative Filtering Model · p. 3
  • Several algorithms have been proposed to improve the learning of the user and item vectors.The paper notes that these algorithms are discussed in its related-work section.Source: B. Collaborative Filtering Model · p. 3

III. THE HETE-CF MODEL

Hete-CF combines user-user, user-item, and item-item relations in a unified recommendation model. It represents relation information through meta-paths and weights their contributions.

A. The Recommendation Model

B. The Learning Algorithm

The learning algorithm alternates between fitting rating vectors and updating the weights for relation meta-paths. Each update step holds the other parameter group fixed.

B. The Learning Algorithm

C. The Complete Algorithm

The complete algorithm takes a heterogeneous information network, user-item ratings, and meta-path sets as input, and outputs predicted ratings. Its social-information weight is adjusted according to rating-matrix sparsity.

C. The Complete Algorithm

  • When the rating matrix is sparse, a larger µ can improve recommendations; when it is not sparse, a larger µ can bias them.The authors state that µ can be calculated from the proportion of non-zero entries in the rating matrix.Source: 1) Parameter Settings · p. 5
  • The algorithm takes a heterogeneous information network, three sets of meta-paths, and rating matrix R as input, and outputs predicted rating matrix R-hat.It initializes U, V, A, B, and W, then iteratively updates U and V and the relation weights.Source: 1) Parameter Settings · p. 6
  • The inner iteration has O(td(mn)) complexity, giving overall complexity O(Ttd(mn)) for T outer iterations.The authors state that this scales linearly with the dimensions of the user-item matrix.Source: 2) Time Complexity Analysis · p. 6

IV. EXPERIMENTAL RESULTS

The experimental section covers evaluation on real datasets, efficiency testing on synthetic datasets, and analysis of parameter values and learned meta-path weights.

IV. EXPERIMENTAL RESULTS

  • The authors report experiments evaluating the algorithm on two real datasets.The supplied section overview does not identify the datasets or report outcome values.Source: §IV · p. 6
  • The section tests algorithm efficiency using synthetic datasets.The overview does not state the efficiency results.Source: §IV · p. 6
  • The authors study the algorithm's parameter and analyze learned weights for selected meta-paths.The supplied overview does not give the findings of these analyses.Source: §IV · p. 6

A. Datasets

The experiments use DBLP and Meetup as heterogeneous social-network datasets, with different recommendation targets and network settings.

A. Datasets

  • Both datasets are described as typical heterogeneous social networks with sparse links, making them settings for testing cold-start and sparsity mitigation.The DBLP schema includes Term objects extracted from paper titles.Source: A. Datasets · p. 6
  • DBLP is a subdataset spanning 261 computer journals and 313 conferences, used to recommend conferences to authors.The author-conference rating is calculated with PathSim on the Author−Paper−Conference meta-path.Source: A. Datasets · p. 6
  • The Meetup subnetwork covers New York City and Los Angeles events, users, and groups, and supports group recommendations to users.The selected meta-paths include all object types while using as few paths as possible.Source: A. Datasets · p. 6

B. Experimental Setup

The experiments average results over repeated train-test splits and compare Hete-CF with baseline and state-of-the-art recommenders using rating-error metrics.

B. Experimental Setup

  • Learning and testing use five-fold cross-validation, with 40% or 60% of data for training, and results averaged over ten trials.The remaining 60% or 40% is used for testing.Source: 1) Cross-validation · p. 6
  • Table II lists selected meta-paths for DBLP and Meetup, which provide the heterogeneous relations used in the experiments.The table caption identifies the two datasets but does not provide additional path details.Source: Table II · p. 7
  • The comparison includes UserMean, ItemMean, NMF, Trust-MF, and Hete-MF; the three baselines use only the target recommendation relation.Trust-MF uses user relations, while Hete-MF considers item relations.Source: 2) Comparison Methods · p. 7
  • Performance is evaluated with MAE and RMSE, using ground-truth and predicted ratings to assess rating error.The MAE denominator is the number of ratings, T = mn.Source: 3) Evaluation Methods · p. 7

C. Result and Analysis

Hete-CF performs better than the compared algorithms in the reported experiments, with performance also varying by training-data amount and feature dimension.

C. Result and Analysis

  • Hete-CF performs better than the other algorithms across the reported comparisons on DBLP and Meetup.Increasing training data improves its performance; on DBLP, d = 5 performs better than other tested feature dimensions.Source: C. Result and Analysis · p. 7

D. Efficiency Study

The efficiency study reports linear runtime growth with feature dimension and dataset size; parameter and weight analyses examine how heterogeneous relations contribute across the two datasets.

D. Efficiency Study

  • Runtime is almost linear in feature dimension, and CPU time increases linearly with dataset size.The feature-dimension study fixes dataset size; the dataset-size study fixes d = 10 and links 20% of objects.Source: D. Efficiency Study · pp. 7–8

E. Parameter Study

  • Extreme μ values yield poorer results, while the calculated values—μ = 0.7 for DBLP and μ = 0.4 for Meetup—coincide with the best results.The authors report that this calculation addresses parameter pre-assignment.Source: E. Parameter Study · p. 8

F. Case Study for Meta-path Weight

  • In DBLP, co-author, conference-through-author, and citation-related meta-paths receive substantial weights; in Meetup, friendship, group-user-group, and event-group paths are prominent.The event-group path weight is interpreted as reflecting the possibility that event participants become group members.Source: F. Case Study for Meta-path Weight · p. 8

V. RELATED WORK

Prior work combines collaborative filtering with social information, but often uses only part of a heterogeneous network’s relations. Hete-CF incorporates user-user, item-item, and user-item relations, and the authors report effective and efficient results on DBLP and Meetup.

A. Collaborative Filtering and Social-based Recommendation

  • Earlier work spans social regularization, event-based group recommendation, online/offline recommendations, and location recommendation using friend and geographic information.Other cited approaches use knowledge from other domains or only part of heterogeneous relationships.Source: A. Collaborative Filtering and… · p. 9

B. Mining Heterogeneous Social Network

  • The paper defines a heterogeneous social network as a heterogeneous information network with social information, including location-based and event-based networks.The proposed method is intended for these network types and other heterogeneous information networks with social information.Source: B. Mining Heterogeneous Social Network · p. 9, §VI · p. 9

A. Collaborative Filtering and Social-based Recommendation

VI. CONCLUSION AND FUTURE WORK

  • Hete-CF is reported effective and efficient on two real-world heterogeneous social networks, DBLP and Meetup.Source: §VI · p. 10
  • The authors identify applying Hete-CF to more real-world recommendation problems and exploring massive social-media data as future work.Source: §VI · p. 10
Loading 1412.7610v1…