A survey of transfer learning for collaborative recommendation with auxiliary data

doi:10.1016/j.neucom.2015.11.059

Neurocomputing

Volume 177, 12 February 2016, Pages 447-453

https://doi.org/10.1016/j.neucom.2015.11.059 Get rights and content

Abstract

Intelligent recommendation technology has been playing an increasingly important role in various industry applications such as e-commerce product promotion and Internet advertisement display. Besides user feedbacks (e.g., numerical ratings) on items as usually exploited by some typical recommendation algorithms, there are often some additional data such as users׳ social circles and other behaviors. Such auxiliary data are usually related to user preferences on items behind numerical ratings. Collaborative recommendation with auxiliary data (CRAD) aims to leverage such additional information so as to improve personalized services. It has received much attention from both researchers and practitioners.

Transfer learning (TL) is proposed to extract and transfer knowledge from some auxiliary data in order to assist the learning task on the target data. In this survey, we consider the CRAD problem from a transfer learning view, especially on how to enable knowledge transfer from some auxiliary data, and discuss the representative transfer learning techniques. Firstly, we give a formal definition of transfer learning for CRAD (TL-CRAD). Secondly, we extend the existing categorization of TL techniques with three knowledge transfer strategies. Thirdly, we propose a novel and generic knowledge transfer framework for TL-CRAD. Fourthly, we describe some representative works of each specific knowledge transfer strategy in detail, which are expected to inspire further works. Finally, we conclude the survey with some summarized discussions and several future directions.

Introduction

Intelligent recommendation technology [1], [4], [18], [31], [45], [48] has been a standard component embedded in many Internet systems such as e-commerce and advertisement systems to provide personalized services. There are two main approaches widely used in personalized recommendation for an active user, i.e., content-based recommendation [3] and collaborative recommendation [14]. Content-based methods promote an item based on the relevance between a candidate item and the active user׳s consumed items, while collaborative recommendation techniques focus on collective intelligence and exploit the community׳s data so as to recommend preferred items from users with similar tastes. However, both methods are limited to users׳ feedbacks of explicit scores or implicit examinations, which may result in a challenging problem, data sparsity, due to the lack of users׳ behaviors.

Fortunately, there are often some additionally available data besides the users׳ feedbacks (e.g., numerical ratings) in a recommender system. There are at least four types of auxiliary data as shown in Table 1, such as content information [52], [56], time contextual information [23], [36], social or information networks [21], [49], [54] and additional feedbacks [19], [29], [39]. These auxiliary data have the potential to help relieve the aforementioned sparsity problem and thus improve the recommendation performance. In this survey, we study on how to exploit different types of auxiliary data in collaborative recommendation, which is coined as collaborative recommendation with auxiliary data (CRAD).

Specifically, we study the CRAD problem from an inductive transfer learning [37] view (instead of unsupervised or transductive transfer learning views [2]), in which we consider the users׳ feedback data as our target data or supervised information, and all the other additional information as our auxiliary data. In particular, we focus on how to enable knowledge transfer from some auxiliary data to the target data in order to address the aforementioned sparsity challenge. We discuss some representative transfer learning techniques, aiming to answer the fundamental question of transfer learning [37], i.e., “how to transfer”. With this focus in our survey, we extend previous categorization of transfer learning techniques in collaborative filtering [38], [43], and answer the above question from two dimensions, including knowledge transfer algorithm styles (i.e., adaptive, collective and integrative knowledge transfer) and knowledge transfer strategies (i.e., prediction rule, regularization and constraint). Then, we propose a novel and generic knowledge transfer framework and describe some representative works in each category to answer the “how to transfer” question in detail, in particular the main idea that may be generalized to other applications. Finally, we conclude the survey with some summarized discussions and several exciting future directions.

Section snippets

Problem definition

We have a target data set and an auxiliary data set. In the target data set, we have some feedbacks from n users and m items, which is usually represented as a rating matrix $R = {[r_{ui}]}^{n \times m}$ and an indicator matrix $Y \in {0, 1}^{n \times m}$ , where $y_{ui} = 1$ means that the feedback r_ui is observed. In the auxiliary data set, we have some additional data such as content, context, network and feedback information as shown in Table 1. Our goal is to predict the unobserved feedbacks in R by transferring knowledge from the

Adaptive knowledge transfer

Adaptive knowledge transfer aims to adapt the knowledge extracted from an auxiliary data domain to a target data domain. This is a directed knowledge transfer approach similar to traditional domain adaptation methods. In this section, we describe two adaptive knowledge transfer strategies as instantiated from Eq. (1), including (i) transfer via regularization, $\min_{Θ} E (Θ | R) + R (Θ | K)$ , and (ii) transfer via constraint, $\min_{Θ} E (Θ | R), s.t. Θ \in C (K)$ .

Collective knowledge transfer

Collective knowledge transfer usually jointly learns the shared knowledge and unshared effect of the target data and the auxiliary data simultaneously, which is a bi-directed knowledge transfer approach with richer interactions similar to multi-task learning algorithms. We describe some representative works of collective knowledge transfer via constraint on model parameters, $\min_{Θ, K} E (Θ | R) + R (Θ) + E (K | A) + R (K), s.t. Θ \in C (K)$ , which is also an instantiation of Eq. (1). Note that the model parameter Θ and

Integrative knowledge transfer

Integrative knowledge transfer incorporates the raw auxiliary data as known knowledge into the learning task on the target data. It can be considered as an embedded knowledge transfer approach similar to feature engineering, information fusion and data integration methods. Mathematically speaking, we can instantiate the generic framework in Eq. (1), and have (i) transfer via prediction rule, $\min_{Θ} E (Θ | R, A) + R (Θ)$ , (ii) transfer via regularization, $\min_{Θ} E (Θ | R) + R (Θ | A)$ , and (iii) transfer via

Discussions

We summarize some representative works of transfer learning for collaborative recommendation with auxiliary data (TL-CRAD) in Table 2. We can see that integrative knowledge transfer via prediction rule and collective knowledge transfer via constraint have recently received more attention, which are also the state-of-the-art TL-CRAD algorithms w.r.t. recommendation accuracy in corresponding problem settings. The interaction between auxiliary data and target data usually becomes richer from

Acknowledgment

I would like to thank Prof. Qiang Yang for advice and comments, Dr. Bin Li for linguistic improvement and helpful discussions, the editors and reviewers for constructive suggestions, and the support of Natural Science Foundation of Guangdong Province No. 2014A030310268, National Natural Science Foundation of China Nos. 61502307, 61170077, 61472258 and Natural Science Foundation of SZU No. 201436.

Weike Pan received the Ph.D. degree in Computer Science and Engineering from the Hong Kong University of Science and Technology, Kowloon, Hong Kong, China, in 2012. He is currently a Lecturer (research oriented) with the College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China. His research interests include transfer learning, recommender systems, and statistical machine learning.

References (61)

J. Bobadilla et al.
Recommender systems survey
Knowl.-Based Syst.
(2013)
Jing Pan et al.
Robust probabilistic tensor analysis for time-variant collaborative filtering
Neurocomputing
(2013)
Weike Pan et al.
Compressed knowledge transfer via factorization machine for heterogeneous collaborative recommendation
Knowl.-Based Syst.
(2015)
Weike Pan et al.
Adaptive bayesian personalized ranking for heterogeneous implicit feedbacks
Knowl.-Based Syst.
(2015)
Deuk Hee Park et al.
A literature review and classification of recommender systems research
Expert Syst. Appl.
(2012)
Xiwang Yang et al.
A survey of collaborative filtering based social recommender systems
Comput. Commun.
(2014)
Liyan Zhang et al.
A random-walk based recommendation algorithm considering item categories
Neurocomputing
(2013)
Ya Zhang et al.
Collaborative filtering with social regularization for tv program recommendation
Knowl.-Based Syst.
(2013)
Gediminas Adomavicius et al.
Toward the next generation of recommender systemsa survey of the state-of-the-art and possible extensions
IEEE Trans. Knowl. Data Eng.
(2005)
Andrew Arnold, Ramesh Nallapati, William W. Cohen, A comparative study of methods for transductive transfer learning,...

Chumki Basu, Haym Hirsh, William Cohen, Recommendation as classification: using social and content-based information in...

Stephen Boyd et al.

Convex Optimization

(2004)

Bin Cao, Nathan Nan Liu, Qiang Yang, Transfer learning for collective link prediction in multiple heterogenous domains,...

Sotirios Chatzis, Nonparametric Bayesian multitask collaborative filtering, in: Proceedings of the 22nd ACM...

Liang Chen et al.

Analysis and detection of fake views in online video services

ACM Trans. Multimed. Comput. Commun. Appl.

(2015)

Liang Chen et al.

Smart streaming for online video services

IEEE Trans. Multimed.

(2015)

Tianqi Chen et al.

Svdfeaturea toolkit for feature-based collaborative filtering

J. Mach. Learn. Res.

(2012)

Wei Chen, Wynne Hsu, Mong Li Lee, Making recommendations from multiple domains, in: Proceedings of the 19th ACM SIGKDD...

Ignacio Fernandez-Tobias, Ivan Cantador, Marius Kaminskas, Francesco Ricci, Cross-domain recommender systems: a survey...

Sheng Gao, Hao Luo, Da Chen, Shantao Li, Patrick Gallinari, Jun Guo, Cross-domain recommendation via cluster-level...

David Goldberg et al.

Using collaborative filtering to weave an information tapestry

Commun. ACM

(1992)

Robert E. Haskell, Transfer of Learning: Cognition, Instruction, and Reasoning, Educational Psychology Series. Academic...

Niklas Jakob, Stefan Hagen Weber, Mark Christoph Müller, Iryna Gurevych, Beyond the stars: exploiting free-text user...

Mohsen Jamali, Martin Ester, A matrix factorization technique with trust propagation for recommendation in social...

Dietmar Jannach et al.

Recommender Systems: An Introduction

(2010)

Gawesh Jawaheer et al.

Modeling user preferences in recommender systemsa classification framework for explicit and implicit user feedback

ACM Trans. Interact. Intell. Syst.

(2014)

Meng Jiang et al.

Social recommendation with cross-domain transferable knowledge

IEEE Trans. Knowl. Data Eng.

(2015)

Meng Jiang, Peng Cui, Fei Wang, Qiang Yang, Wenwu Zhu, Shiqiang Yang, Social recommendation across multiple relational...

Wolf Kienzle, Kumar Chellapilla, Personalized handwriting recognition via biased regularization, in: Proceedings of the...

Yehuda Koren, Collaborative filtering with temporal dynamics, in: Proceedings of the 15th ACM SIGKDD International...

Cited by (119)

CNNRec: Convolutional Neural Network based recommender systems - A survey
2024, Engineering Applications of Artificial Intelligence
Easy internet access and technological advancements have resulted in information overload and a plethora of options, making decision-making extremely difficult. Recommender System (RS) is a potential solution for assisting users in making decisions by recommending or predicting product ratings. Three fundamental forms of RS that use implicit or explicit feedback for recommendation are collaborative, content-based, and hybrid filtering. Ratings are the most common form of feedback, but product descriptions, reviews, images, audios, and videos are also important and can help improve the performance of the traditional RS. These additional variables can have a significant impact on RS’s performance. Traditional RSs used approaches based on the nearest neighbor or other machine learning models, but thanks to recent advances in artificial intelligence and deep learning, RSs are now being developed using Convolutional Neural Networks (CNN), which can efficiently exploit auxiliary information. In addition to comparing CNN-based RSs on common grounds, this article provides a full examination of CNN-based RSs and how they might use various types of auxiliary information. The study also discusses data characteristics, data statistics, and auxiliary information in a variety of publicly available datasets. Different evaluation measures for RSs are also discussed, and readers are provided with interesting challenges and open research issues.
Extracting latently overlapping users by graph neural network for non-overlapping cross-domain recommendation
2024, Knowledge-Based Systems
Cross-domain collaborative filtering (CDCF) is an effective solution to alleviate the data sparsity problem. Most of existing CDCF methods rely on overlapping data, such as users, items or both. But in some realistic scenes, detection and accessibility of overlapping data are difficult or even impossible, which poses a pressing demand for researches on cross-domain recommendation without overlapping data. There actually have been some attempts on addressing this problem by sharing cluster-level rating pattern across source and target domains. But these solutions require explicit ratings, which makes them not suitable for more common implicit feedback recommendation. To address this problem, we propose a novel CDCF model for non-overlapping data scenarios, which adaptively extracts latently overlapping users of source and target domains from all users to build an implicit bridge for knowledge transfer. Specifically, we first design a self-supervised classifier guided by inter-domain contrastive learning to divide domain users into distinct groups based on their preference differences. Then, we perform graph convolution operations on the subgraph formed by such group users and their interactive items to explicitly mine the higher-order collaborative relationships between users and items. Finally, we construct sparse and reasonable implicit bridges between domains by designing flow-aware similarity measures for selective knowledge transfer among the extracted latently overlapping users. Extensive experiments on four public datasets demonstrate the superior performance of our proposed model over several state-of-the-art graph-based single- and cross-domain models.
Transfer learning for collaborative recommendation with biased and unbiased data
2023, Artificial Intelligence
In a recommender system, a user's interaction is often biased by the items' displaying positions and popularity, as well as the user's self-selection. Most existing recommendation models are built using such a biased user-system interaction data alone. In this paper, we introduce an additional specially collected unbiased data, and then have a new problem called collaborative recommendation with biased and unbiased data.
We first formalize the studied problem and list three challenges, including the bias challenge, the heterogeneity challenge and the unbalance challenge. Then we propose a novel transfer learning-based AI solution, i.e., transfer via joint reconstruction (TJR), to achieve knowledge transfer and sharing between the biased data and unbiased data. Specifically, in our TJR, we use two different models to extract the users' preferences and bias information, and then refine the prediction via the latent features containing the bias information in order to obtain a more accurate and unbiased recommendation. We further integrate the two data by reconstructing their interaction in a joint learning manner. Moreover, in order to better address the unbalance challenge, we introduce a bias regularization term and integrate bidirectional knowledge distillation. Finally, we adopt four representative methods, i.e., variational autoencoders, matrix factorization, neural collaborative filtering and graph convolution network, as the backbone models of our TJR and conduct extensive empirical studies on three public datasets, showcasing the effectiveness of our transfer learning solution over some very competitive baselines.
AGRE: A knowledge graph recommendation algorithm based on multiple paths embeddings RNN encoder
2023, Knowledge-Based Systems
More and more researches have focused on the use of knowledge graphs (KG) to solve the sparsity problem of traditional collaborative filtering recommendation systems. Most KG based recommendation algorithms focus on independent paths connecting users and items, or iteratively propagate user preferences in KG. However, the current approachs that focus on indedpent paths ignore the association between paths. Therefore, in this study, we propose a knowledge graph recommendation system algorithm for the multiple paths RNN encoder (AGRE), which fully considers the association between paths. Specifically, the paths between the user and the item are coded by a specified RNN (MRNN) to accurately learn the user’s preferences. Traditional RNNs can encode multiple paths without considering the association between paths, but our RNN can encode multiple paths with considering the association between paths. We have compared AGRE with other state-of-the-art algorithms on three real-world datasets, and achieved good results in terms of AUC and Precision@K. This indicates that AGRE could solve the problem of sparse interaction between users and items, and could make full use of the knowledge graph for recommendation.
BAR: Behavior-aware recommendation for sequential heterogeneous one-class collaborative filtering
2022, Information Sciences
Citation Excerpt :
Multi-feedback Bayesian personalized ranking (MF-BPR) [31] considers that different feedback reflect users’ preferences from different levels, and proposes a non-uniform sampler which places emphasis on high-level items to sample some positive items, while the negative items are sampled from a lower level than that of the positive items. To tackle the sparsity of the target feedback in OCCF, another way to make use of the dense examination data is to use transfer learning [32], which shares some knowledge from the auxiliary feedback with the task of modeling the target feedback. TJSL [33] introduces a new term that learns a similarity between a candidate item and an examined item in the prediction rule of FISM [15] to alleviate the sparsity of the target feedback.
In our daily life, we are often greatly assisted with recommendation engines in finding the required information efficiently and accurately. In this paper, we focus on an emerging and important recommendation problem in enormous real-world applications, i.e., sequential heterogeneous one-class collaborative filtering (SHOCCF). In the studied problem, we have some users’ sequential and heterogeneous one-class feedback, i.e., sequences of (item, behavior) pairs, where the behaviors can be of different types such as examinations and purchases. We propose a generic solution called behavior-aware recommendation (BAR), which is able to adapt an existing RNN-, CNN-, attention- or GNN-based sequential recommendation method to SHOCCF. The main idea of our BAR is to provide the behavior information to the input and output of a representation module that models the item sequence. Specifically, we design a behavior attention layer that uses the behavior at the next timestamp in order to digest the behaviors at different positions in the sequence and obtain more accurate attention scores that will be fed to the input of the representation module. Moreover, we further design a task-specific layer to fuse the real behavior at the next timestamp with the sequential feature generated by the representation module to distinguish different prediction tasks w.r.t. the behavior types. We then conduct extensive empirical studies on four public datasets and find that our BAR is able to significantly improve the performance of a certain sequential recommendation method when it is adapted to SHOCCF.
A hinge-loss based codebook transfer for cross-domain recommendation with non-overlapping data
2022, Information Systems
Recommender systems, especially collaborative filtering (CF) based recommender systems, has been playing an important role in many e-commerce applications. As the information being searched over the internet is rapidly increasing, users often face the difficulty of finding items of his/her own interest and recommender systems often provides help in such tasks. Recent studies show that, as the item space increases, and the number of items rated by the users become very less, issues like sparsity arise. To mitigate the sparsity problem, transfer learning techniques are being used wherein the data from dense domain (source) is considered in order to predict the missing entries in the sparse domain (target). In this paper, we propose a novel transfer learning approach called Transfer of Codebook via Hinge loss or TCH for cross-domain recommendation when both domains have no overlap of users and items. In our approach constructing the codebook and transferring the same knowledge from source to target domain is done in a novel way. We employ a similar formulation of co-clustering technique to obtain the codebook (cluster-level rating pattern) of source domain. By making use of hinge loss function we transfer the learnt codebook of the source domain to target. The use of hinge loss as a loss function is novel and has not been tried before in transfer learning. We demonstrate that our technique improves the approximation of the target matrix on benchmark datasets.

View all citing articles on Scopus

View full text

A survey of transfer learning for collaborative recommendation with auxiliary data

Abstract

Introduction

Section snippets

Problem definition

Adaptive knowledge transfer

Collective knowledge transfer

Integrative knowledge transfer

Discussions

Acknowledgment

Knowl.-Based Syst.

Neurocomputing

Knowl.-Based Syst.

Knowl.-Based Syst.

Expert Syst. Appl.

Comput. Commun.

Neurocomputing

Knowl.-Based Syst.

Toward the next generation of recommender systemsa survey of the state-of-the-art and possible extensions

IEEE Trans. Knowl. Data Eng.

Convex Optimization

Analysis and detection of fake views in online video services

ACM Trans. Multimed. Comput. Commun. Appl.

Smart streaming for online video services

IEEE Trans. Multimed.

Svdfeaturea toolkit for feature-based collaborative filtering

J. Mach. Learn. Res.

Using collaborative filtering to weave an information tapestry

Commun. ACM

Recommender Systems: An Introduction

Modeling user preferences in recommender systemsa classification framework for explicit and implicit user feedback

ACM Trans. Interact. Intell. Syst.

Social recommendation with cross-domain transferable knowledge

IEEE Trans. Knowl. Data Eng.