EXPLORE: EXPLainable item-tag CO-REcommendation
Introduction
Recommender systems have been studied and deployed in various domains. They become indispensable because they help manage information overload problems and provide customized recommendations based on users’ preferences. Collaborative filtering (CF) is one of the most successful approaches to recommendations [14], [30]. The core idea of CF is that users with similar tastes share similar rating distributions toward the same items. Popular methods combine CF with auxiliary information to make recommendations.
Recently, social tagging systems (e.g., Delicious, CiteULike, and Flickr) have become prevalent and accumulated abundant tags. These tags contain rich information and provide effective ways for users to organize, manage, share, and search various kinds of items [7]. Researchers have exploited tags to improve recommendations with various methods. We can categorize the existing studies into two groups: (1) tag-based item recommendation, and (2) tag recommendation. Tag-based item recommendations use tags as auxiliary information to improve recommendation accuracy. Some researchers used tags to calculate the similarity of users [40] or items [41], and applied these similarities into the CF procedure. Other researchers mined the semantic meanings of tags to model users’ interests and items’ features [7], [33], [39]. These studies have confirmed that the abundance and semantic meanings of tags could help improve the performance of recommender systems. Tag recommendation refers to the automated process of suggesting useful and informative tags to an emerging object based on historical information [29]. Researchers explored the co-occurrences of tags [28], the interactions between items and tags [10], [20], [21], [36], and the semantic meanings of tags [11], [18] to recommend tags. For both item recommendation and tag recommendation, CF-based techniques have been well studied and have shown promising results.
However, most existing studies have only focused on improving the accuracy of either item recommendation or tag recommendations. In reality, a social tagging system requires both types of recommendations. When viewing the recommendations from a unified perspective, we can utilize more coherent features and mine the relations between the two recommendations. We can consider not only the accuracy but also other factors that influence recommendation (e.g., recommendation novelty, explanation, and metric design). Aside from accuracy, in this study, we mainly consider two issues that have been overlooked in single recommendation scenarios.
One issue is that recommendations lack explainability, which reveals why users might like the items that a recommender system has recommended thus helping users make appropriate decisions. For example, as depicted in Fig. 1, we recommend two papers about CF to users. The tags serve as the explanations to the papers. Two users with different interests look through the recommended list to decide which paper to read. Alice is new to recommender systems and wants to learn about CF. She finds that the first paper “Item-based Collaborative Filtering Recommendation Algorithms” is explained with “Classic” and “Beginner” tags and thus decides to read this paper. Bob is an engineer in distributed systems and has some knowledge about CF. He wants to implement CF in a distributed style and he chooses the second paper “Fast Item-based Collaborative Filtering” when he sees the explanations are “Parallel,” “Clustering,” and “Hashing.” Both Alice and Bob make the right decisions with the tags serving as the explanations to the recommended papers. By contrast, one may not choose the right paper without any explanations because the titles of the two papers are too similar, and users, especially those who lack preliminary knowledge, may take time to tell the difference between them. Thus, the explanations play an important role in the decision making of users. Compared with other forms of explanations (e.g., keywords of items [5], similar users’ choices [4], and items’ features [32]), tags contain both factual and subjective information that are more friendly and acceptable to users. However, tags are annotated by human, and not all items can be sufficiently tagged for explanation. The other issue is the metric design for tag recommendation. A proper metric is vital to evaluate the effectiveness of a recommender system. Existing works use all the items to compute the evaluation metrics for tag recommendations (e.g., precision and recall), assuming that users will look through each item with the same probability and treat items equally in the evaluation procedure. In real scenarios, items have different levels of exposures to users. Some popular items are frequently searched or recommended, and the accuracy of theses items’ recommended tags is important for a social tagging system to improve user experience. Other unpopular items draw less attention from users, and the recommended tags of these items have less influence on user experience. Ignoring this issue may lead to biased evaluations. An ideal item subset could be based on item popularity or item exposure to users, which can be difficult to compute. A reasonable alternative is therefore needed.
We propose the idea of item–tag co-recommendation to address the two main issues identified in single-recommendation scenarios. This scheme indicates that the recommender system can recommend an item and the corresponding tags to users simultaneously. For the explainability issue, we use recommended tags to ensure that each item can be explained. In this manner, the problem of unevenly distributed tags is alleviated, and the item–tag co-recommendation system can improve the explanations. For the metric issue, we design a more meaningful metric when the recommended items and recommended tags are available at the same time. The recommended items can be viewed as user exposures. We test the tag recommendation performance by only using users’ recommended items instead of the whole set of items. The proposed technique utilizes is a novel metric that provides a new perspective for observing the effectiveness of a recommender system. We further realize this idea by proposing a novel model that combines the training phases of item recommendation and tag recommendation, thus enabling the simultaneous recommendation of items and their corresponding tags.
Item-tag co-recommendation also has two additional advantages. First, the training phases are unified under the same framework, which means that item-tag co-recommendation can model interrelation and mutual effects by simultaneously updating the common factors used in item recommendation and tag recommendation. Considering that all the parameters are tuned together, we can also explore the tradeoff between item recommendation and tag recommendation. Second, the different aspects of auxiliary information can be thoroughly exploited to improve system performance. The existing item recommendation or tag recommendation methods often utilize content from a single source (e.g., user profiles, product descriptions, reviews, or tags) to capture latent factors that exploit only one aspect of items or users. In our item–tag co-recommendation, multiple aspects of content information are involved, and we can utilize all of them rather than one aspect only to capture the latent factors of items. Here, we leverage the implicit feedback of users and the descriptions and tags of items to capture the items comprehensively. This method also helps alleviate the sparsity problem, which means that the recorded data is extremely sparse compared with that of the user-product matrix.
In this study, we propose a novel unified framework named “EXPLainable item-tag CO-REcommendation” (EXPLORE) to item and tag recommendations for providing explanations to recommendation results, thus boosting the prediction performance of items and tags. The main contributions of this study are listed as follows:
- •
We propose a novel co-recommendation framework that jointly recommends items and corresponding tags simultaneously. To the best of our knowledge, this it is the first work to use content CF-based techniques to perform co-recommendation.
- •
The recommended tags serve as explanations to the recommended items, thus making the recommendation much more convincing and interpretable. We also introduce a new metric to evaluate tag recommendation.
- •
We capture the latent factors of an item by considering user’s implicit feedback, item’s content, and the tags people annotated. Contrary to most existing studies that utilize a single source of content, we take advantage of the multiple aspects of content to capture items comprehensively.
- •
Experiments on real-world datasets show that EXPLORE achieves higher performance in both item and tag recommendations than the state-of-the-art methods. The co-recommendation proves to be a mutual promotion for both recommendations.
The remainder of this paper is organized as follows. In Section 2, we introduce the related work, including item recommendation, tag recommendation, topic-based CF techniques, and recommendation explainability. In Section 3, we describe the details of our EXPLORE framework. In Section 4, we present the experiments and discussions. In Section 5, we discuss the conclusions and future works.
Section snippets
Related work
In this section, we review three groups of work related to this study, including (1) CF-based recommendation, (2) recommendation explainability, and (3) Collaborative Topic Regression (CTR).
EXPLORE: An Explainable item-tag CO-Recommendation framework
We now focus on describing the details of our proposed framework named EXPLORE. We first formulate the problem of item-tag co-recommendation. Then we introduce the model formulation of EXPLORE by explicating the parameter estimation algorithm and the prediction method. Some important notations are explained in Table 1.
Experiments
We conducted comprehensive experiments on three datasets and compared EXPLORE with several baseline algorithms to show the superiority of our framework. The questions we aim to answer are as follows:
- 1.
How does EXPLORE outperform the baseline methods, in terms of item recommendation and tag recommendation?
- 2.
How is the prediction performance affected by the parameters? What roles do the two confidence parameters λuv and λvt play in recommendations?
- 3.
To what extent do the recommended tags provide
Conclusion and future work
In this study, we developed a novel hierarchical framework named EXPLORE that jointly and synchronously recommended items and corresponding tags for explanations. We explored multi-aspects of content information to capture the item comprehensively by using the user’s implicit feedback and the item’s contents and tags. The comprehensive experimental results of the three real-world datasets show that the proposed EXPLORE can outperform both item recommendation methods and tag recommendation
Acknowledgement
This work was supported in part by the National Natural Science Foundation of China (No.U1509221), the National Key Technology R&D Program (2015BAH07F01), the Zhejiang Province key R&D program (No.2017C03044).
References (41)
- et al.
Trust-based recommendation systems: an axiomatic approach
Proceedings of the Seventeenth International Conference on World Wide Web
(2008) - et al.
The social bookmark and publication management system bibsonomy
VLDB J. Int. J. Very Large Data Bases
(2010) Nonlinear programming
J. Oper. Res. Soc.
(1997)- et al.
Explaining recommendations: satisfaction vs. promotion
Proceedings of Beyond Personalization 2005: A Workshop on the Next Stage of Recommender Systems Research at the 2005 International Conference on Intelligent User Interfaces
(2005) - et al.
A personal news agent that talks, learns and explains
Proceedings of the Third Annual Conference on Autonomous Agents
(1999) - et al.
Latent Dirichlet allocation
J. Mach. Learn. Res.
(2003) - et al.
Capturing semantic correlation for item recommendation in tagging systems
Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence
(2016) - et al.
Context-aware collaborative topic regression with social matrix factorization for recommender systems
Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence
(2014) - et al.
Deep neural networks for youtube recommendations
Proceedings of the Tenth ACM Conference on Recommender Systems
(2016) - et al.
Personalized tag recommendation through nonlinear tensor factorization using Gaussian kernel
Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence
(2015)
Survey on social tagging techniques
SIGKDD Explor.
Explaining collaborative filtering recommendations
Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work
A synthetic approach for recommendation: combining ratings, social relations, and reviews
Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence
A comparison of collaborative-filtering recommendation algorithms for e-commerce
Intell. Syst. IEEE
Trustwalker: a random walk model for combining trust-based and item-based recommendation
Proceedings of the Fifteenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
Tag recommendations in folksonomies
Proceedings of the European Conference on Principles of Data Mining and Knowledge Discovery
Matrix factorization techniques for recommender systems
IEEE Comput.
Latent dirichlet allocation for tag recommendation
Proceedings of the Third ACM Conference on Recommender Systems
Deep collaborative filtering via marginalized denoising auto-encoder
Proceedings of the Twenty-Fourth ACM International on Conference on Information and Knowledge Management
Tag sources for recommendation in collaborative tagging systems
Proceedings of ECML PKDD (The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases) Discovery Challenge
Cited by (25)
Example-based explanations for streaming fraud detection on graphs
2023, Information SciencesCitation Excerpt :Diversity: Explanations shall be fair in the sense that they are not biased to one fraud structure. Information redundancy shall be avoided to provide comprehensive insights [25]. Conciseness: Explanations shall be as simple as possible, following the principle of Occam’s razor.
Similarity attributed knowledge graph embedding enhancement for item recommendation
2022, Information SciencesCitation Excerpt :Therefore, various approaches have incorporated user-item side information into recommendations to overcome the breaches. For instance, [4] used plain-text-based user-item information, [5] extracted entity information from social networks, [6] exploited location-information, and [7,8] used contextual information. KG-based side information is deemed as the most feasible type of side information that exhibits facts and figures (i.e., data) through entities and relations among them.
TAERT: Triple-Attentional Explainable Recommendation with Temporal Convolutional Network
2021, Information SciencesIR-Rec: An interpretive rules-guided recommendation over knowledge graph
2021, Information SciencesO<sup>3</sup>ERS: An explainable recommendation system with online learning, online recommendation, and online explanation
2021, Information SciencesCitation Excerpt :Among the aforementioned explainable recommendation models, Bart has a different reward assumption from our proposed models, whereas EFM and EXPLORE have the same reward function as ours. In general, we adopt two widely used metrics, namely, Cumulative Regret and Cumulative Reward, to evaluate the performance of the online recommendation task [8,37,38] and three metrics, namely, Cumulative Recall and Cumulative Precision, and Cumulative NDCG, to evaluate the performance of the online explanation task in the ERSs [5,26,41,45,49]. We provide a detailed description of each metric as follows.
Tag-informed collaborative topic modeling for cross domain recommendations
2020, Knowledge-Based SystemsCitation Excerpt :TRCF (Tag and Rating based CF) [3] first mines the semantic of tags for users and items, then incorporates the information into matrix factorization. EXPLORE (EXPLainable item-tag CO-REcommendation) [5] jointly recommends items and the corresponding tags to improve the performance of both approaches. CTR-SR [4] is a novel hierarchical Bayesian model with social regularization to integrate the item-tag matrix, item content information and social networks between items into the same principled model.