Modified Conditional Restricted Boltzmann Machines for Query Recommendation in Digital Archives
Round 1
Reviewer 1 Report
1. This paper uses rich metadata in DAs to modify conditionally restricted Boltzmann machines and improve the retrieval function of the DAs so that non-expert users can reach the level of expert users.
2. The research objectives are clear, the essay is well-structured, and the design description is appropriate.
Author Response
Thank you very much for your affirmation of our research, we will continue to improve this research in the future.
Reviewer 2 Report
This paper describes a system for query recommendation in Digital Archives (DAs).
The proposed approach is theoretically sound and it relies on Boltzmann machines; experiments suggest the system is able to generate semantically similar recommendations.
The paper is clear and well written. The authors distribute the code through GitHub which makes their reserach work fully reproducible.
The theoretical contribution is strong and also experiments are sound and convincing. The task of recommending items has a long tradition but the considered problem is novel and interesting. I therefore recommend paper acceptance after few minor modifications.
Specifically:
1) Please check last items of bibliography (from Ref 30) do not include authors.
2) Is the proposed model capable of considering contextual factors in generating recommendations? For instance, trust is a key ingredient to compute reliable re commendations and I would like to kno if the proposed model can incorporate weights which reflect the levl of trust of a user. I would suggest to add a small discussion section and to cite some papers about trust in recommender systems:
Masoud Dadgar, Ali Hamzeh. How to Boost the Performance of Recommender Systems by Social Trust? Studying the Challenges and Proposing a Solution. IEEE Access 10: 13768-13779 (2022)
Pasquale De Meo. Trust Prediction via Matrix Factorisation. ACM Transactions on Internet Technology 19(4): 44:1-44:20 (2019)
Athanasios N. Nikolakopoulos, Xia Ning, Christian Desrosiers, George Karypis. Trust Your Neighbors: A Comprehensive Survey of Neighborhood-Based Methods for Recommender Systems. Recommender Systems Handbook 2022: 39-89
3) Can you briefly comment on the scalability of the proposed system?
Author Response
Thank you very much for your affirmation and suggestions for our manuscript. In response to your revision suggestions, we respond as follows.
1) We apologize for giving wrong references. They are the un-removed reference template. We have removed them.
2) Thank you very much for the articles you recommended. They are of great value for our future research directions. Trust-based recommendation is a very interesting research topic. The proposed model can make more personalized recommendations by adding the trust relationship among users. In the context of query recommendation task in DAs that involved in our research, different users are interested in different topics, so when the same query is input, the obtained query recommendations should be different according to user preferences. Different from simply adding collaborative filtering (CF), which only considers whether users have similar preferences on certain items, trust-based recommendation allows users interested in different topics, which could tolerate different preferences for different items. In the ARC-UPD dataset used in this paper, the users' interest in different topics and trust weights can be extracted from the access log data. Adding these to the current recommendations can enable users to obtain more personalized recommendation results. For DAs associated with online social networks, user relationship information can be extracted more easily. The user relationships can be used as learning data for trust weight to improve personalization.
The discussion is added to Section 7 (line 532-545).
3) It is considered that the proposed model is scalable to other DA datasets. The experiments utilize the Europeana dataset added in the manuscript in this revision round illustrate this point (please refer to Section 5 (line 380-404) and Section 6 (line 514-524) in the manuscript). The experimental results using the Europeana dataset show that the model and training method we propose can effectively learn the query co-occurrence patterns in another DA (digital archive) dataset. However, according to the practical situations of different DAs, the proposed model needs to be adjusted. For example, the Europeana dataset used in this experiment is small and has little conditional information, thus the performance is not as good as that of ARC-UPD dataset. In such situation, other context data can be considered to use to improve the performance of the model.
Reviewer 3 Report
The major contribution of the paper is the proposal modified conditional restricted Boltzmann machines (M-CRBMs) for recommendation in digital archives.
Strengths:
1. Release of code implementing the proposed method.
2. Detailed experimental assessment
3. Well written paper.
Weaknesses:
1. Please explicitly mention the size of all variables.
2. Replicate the experiment on a dataset of Europeana.
3. Have a look on the references 30 -37.
Author Response
Thank you very much for your affirmation and suggestions for our manuscript. In response to your revision suggestions, we respond as follows.
- the size of all variables in the proposed M-CRBMs can be calculated as:
Variable number = size of weights + size of bias
= size(W) + size(D1) + size(D2) + size(D3) + size(r1) + size(r2) + size(r3) + size(v).
size(W) = size(v)*size(h)
size(D1) = size(r1)*size(h)
size(D2) = size(r2)*size(h)
size(D3) = size(r3)*size(h)
In the experiment using ARC-UPD dataset and take three types of conditional information, the number of variables is around 269 million (15803 variables for layer v, 1624 variables for layer r1, 2052 variables for layer r2, 14125 variables for layer r3, 8000 variables for layer h). In the experiment using Europeana dataset, the number of variables is around 9 million (3816 variables for layer v, 5600 variables for layer r1, 1000 variables for layer h). The number of variables are added to Section 5 (line 289-290, line 382-384).
- We add an experiment on Europeana dataset. The description of the experiment is stated in Section 5 (line 380-404) and Section 6 (line 514-524). The experiment shows that the provided algorithm is also available for Europeana dataset. Utilizing a large amount of real query search logs and adding other appropriate conditional information are considered to further improve the results. We will soon update the code and dataset of this experiment to GitHub.
- We apologize for giving wrong references. They are the un-removed reference template. We have removed them.
