Paper 23 (Research track)

Transfer Learning for Item Recommendations and Knowledge Graph Completion in Item Related Domains via A Co-Factorization Model

Author(s): Guangyuan Piao, John G. Breslin

Full text: submitted version

camera ready version

Decision: accept

Abstract: With the popularity of Knowledge Graphs (KGs) in recent years, there have been many studies leveraging the abundant background knowledge available in KGs for the task of item recommendations. However, little attention has been paid to the incompleteness of KGs when leveraging knowledge from them. In addition, previous studies have mainly focused on exploiting knowledge from a KG for item recommendations, and it is unclear whether we can exploit the knowledge in the other way, i.e, whether user-item interaction histories can be used for improving the performance of completing the KG with regard to the domain of items.

In this paper, we investigate the effect of knowledge transfer between two tasks: (1) item recommendations, and (2) KG completion, via a co-factorization model (CoFM) which can be seen as a transfer learning model. We evaluate CoFM by comparing it to three competitive baseline methods for each task. Results indicate that considering the incompleteness of a KG outperforms other compared methods, including a state-of-the-art factorization method leveraging existing knowledge from the KG. In addition, the results show that exploiting user-item interaction histories also improves the performance of completing the KG with regard to the domain of items, which has not been studied before.

Keywords: Recommender Systems; Knowledge Graph; Transfer Learning; Factorization Models

 

Review 1 (by Vito Bellini)

 

(RELEVANCE TO ESWC) This work provides sufficient in-depth discussion of the transfer learning between item recommendations an KG completion. I think this work can be considered an advancement for this task and it is relevant to the Machine Learning track.
(NOVELTY OF THE PROPOSED SOLUTION) The paper mainly presents a novel idea of knowledge transfer between two tasks:
1. item recommendations
2. KG completation
using FM models that are a state-of-the-art factorization models.
(CORRECTNESS AND COMPLETENESS OF THE PROPOSED SOLUTION) Proposed solution is well explained and all details are covered.
(EVALUATION OF THE STATE-OF-THE-ART) They have compared their approach with several baselines to to prove the strength of their work.
(DEMONSTRATION AND DISCUSSION OF THE PROPERTIES OF THE PROPOSED APPROACH) The authors also present some experimental results on two datasets to show that the
methodology is effective and outperforms baselines.
(REPRODUCIBILITY AND GENERALITY OF THE EXPERIMENTAL STUDY) Experiments are reproducible, they have used publicly datasets, evaluation protocol is described and evaluation metrics are appropriate for the task of top-N recommendation.
(OVERALL SCORE) In this work authors have investigated about transfer learning between the task of item recommendations and knowledge-graph completion.
They used the item recommendation as the target task and KG completion as the source task, exploting user-item interaction histories and its results in an improvement in performances of completing the KG.
Strong Points:
1. publicy datasets
2. reproducible experiment
3. paper well written

 

Review 2 (by Vito Walter Anelli)

 

(RELEVANCE TO ESWC) The paper is well-written and easy to follow, the topic is interesting and relevant to ESWC.
(NOVELTY OF THE PROPOSED SOLUTION) The proposed idea is novel and it could be useful for many researchers.
(CORRECTNESS AND COMPLETENESS OF THE PROPOSED SOLUTION) The paper is well-written and easy to follow, The authors propose two different ways to model the relationship between the latent representations of items, CoFMA and CoFMR. The explanation of the method is clear.
(EVALUATION OF THE STATE-OF-THE-ART) The related works are well-written even though the section related to factorization models and Linked Open Data features could be extended with other works in single domain and cross-domain scenarios.
(DEMONSTRATION AND DISCUSSION OF THE PROPERTIES OF THE PROPOSED APPROACH) The results discussion about the two different champions seems to me to be quite short. It would be interesting to read some explanation by the authors about that. Could a third dataset help in this explanation?
(REPRODUCIBILITY AND GENERALITY OF THE EXPERIMENTAL STUDY) The authors propose two different ways to model the relationship between the latent representations of items, CoFMA and CoFMR, and both variants perform well on the two different datasets, but each of them is champion on a dataset. 
Experiments are well conducted but the results discussion about the two different champions seems to me to be quite short. It would be interesting to read some explanation by the authors about that. Could a third dataset help in this explanation?
Datasets used are publicly available and this is a strong point for the work but the lack of a public implementation of the proposed method makes uneasy to reproduce the experiments.
(OVERALL SCORE) In this paper a transfer learning approach between top-N recommendation task (RS) and KG completion task (KGC) is proposed.
The transfer learning is achieved via a coFactorization model in which the learning is devoted to the optimization of the target task.
An evaluation of the method is proposed for both target tasks RS-KGC and KGC-RS.
Strong points: 
- Novel and interesting idea 
- clear explanation of the method
- publicly available datasets.
Weak points: 
- lack of a detailed discussion about the champions
- lack of reference to publicly available implementation of the method.
- some recent missing references in related work

 

Review 3 (by anonymous reviewer)

 

(RELEVANCE TO ESWC) The work is about link prediction for RDF graphs and recommender systems.
(NOVELTY OF THE PROPOSED SOLUTION) The research problem is about transferring knowledge graph completion to item recommendation and vice versa. It looked the approach has some novelty.
However, the approach just assembles some existing techniques.
(CORRECTNESS AND COMPLETENESS OF THE PROPOSED SOLUTION) I was unable to check the experimental results as the paper does not provide a link to their system. 
I got two concerns:
1. The experimental results look too good to be true in the sense that every value for the proposed approach is just a little bit better than the best value of all other systems. This could be true but further evidences are required.
2. My another major concern is that as a knowledge graph completion problem, the two benchmarks used in the paper are too small.
(EVALUATION OF THE STATE-OF-THE-ART) As mentioned, essentially, no new algorithm is proposed. The proposed approach is an integration of some existing techniques, while this is ok for some cases. Experimental results are presented for comparing the proposed approach with some existing ones but they are not very convincing.
(DEMONSTRATION AND DISCUSSION OF THE PROPERTIES OF THE PROPOSED APPROACH) See above
(REPRODUCIBILITY AND GENERALITY OF THE EXPERIMENTAL STUDY) See my above comments.
(OVERALL SCORE) Positives: The paper addresses an interesting problem; they proposed an approach by integrating some existing techniques; experimental results are reported.
Negatives: I have some concerns about the experimental results:
1. The experimental results look too good to be true in the sense that every value for the proposed approach is just a little bit better than the best value of all other systems. This could be true but further evidences are required.
2. My another major concern is that as a knowledge graph completion problem, the two benchmarks used in the paper are too small.

 

Review 4 (by Tommaso Di Noia)

 

(RELEVANCE TO ESWC) The proposed approach is coherent with the track (Machine Learning)
(NOVELTY OF THE PROPOSED SOLUTION) The proposed approach is original
(CORRECTNESS AND COMPLETENESS OF THE PROPOSED SOLUTION) The proposed solution is well explained and almost every choice is motivated
(EVALUATION OF THE STATE-OF-THE-ART) Description of related works demonstrates the authors' extensive knowledge on the topic. Moreover, the authors compared with several baselines.
(DEMONSTRATION AND DISCUSSION OF THE PROPERTIES OF THE PROPOSED APPROACH) The authors used two publicly available datasets in two different domains in order to strengthen the quality of their approach
(REPRODUCIBILITY AND GENERALITY OF THE EXPERIMENTAL STUDY) public datasets, a good description of the evaluation protocol and metrics make this work highly reproducible
(OVERALL SCORE) ---------------------------Summary of the paper---------------------------
This paper presents a very interesting approach where a co-factorization model is used in order to investigate transfer learning between the Item Recommendations and Knowledge Graphs(KG) completion tasks. First, the authors show that incorporating the "incompleteness" of a KG by using KG completion as source task and item recommendations as target task, outperforms state-of-the-art baselines.
Second, they investigate the performances of transferring knowledge from item recommendations(i.e. user-item interactions) to the KG completion(target task) with respect to the specific domain of items.
Moreover, two strategies have been proposed to model the different representations of the latent factors of an item(in item recommendations task) and of a subject(in the KG completion task): 
Shared latent space (CoFM_{a}): they assume that an item and a subject have the same latent representation.
Via latent space regularization(CoFM_{R}): the two latent factors representations are regularized in order to make them not so different from each other. 
Two different datasets (Movies and Books domain) have been used: Movielens1M and DBbook. 
---------------------------Strong Points---------------------------
1) The paper is nicely executed, well written and very easy to follow.
2) The study was carried out in a solid way and with a novel approach.
3) Description of Related works demonstrates the authors'' extensive knowledge. 
---------------------------Weak Points---------------------------
1) Parameters such as the "transfer" parameter have been optimized by using Stochastic Gradient Descent. What about the baselines(i.e. BPRMF) parameters optimization? Maybe it would have been better to show the results without the tuning of the transfer parameter w.r.t. baselines too(not only tuning vs without tuning). 
2) Why in Movielens dataset is CoFM_{a} the absolute champion while in DBbook we have CoFM_{R}? Is there any motivation?
3) Maybe some notation could be misleading and not intuitive i.e. beta0 to denote the bias and bold beta to denote the feature weights

 

Metareview by Andreas Hotho

 

This paper deals with the problem of transfer learning in two settings, item recommendation and knowledge graph completion. Both task are used as source and target. The paper is well written and interesting to the community. The used method is well know but the setting is quite interesting. The authors state in the rebuttal that the data and code will be published which is a big plus. The most controversial weakness is the size of the dataset, which is rather small compared to a standard recommender setting. This rises the question if the results can be generalized or if they are an artefact of the dataset. But the given significance information show, that there is a clear signal which can't be ignored. Ultimately, no reviewer argued against acceptance even with the small dataset. Given the nice idea and the significant results, we recommend to accept it.

 

Share on

Leave a Reply

Your email address will not be published. Required fields are marked *