----------------------- REVIEW 1 ---------------------
PAPER: 64
TITLE: Translation-based Recommendation
AUTHORS: Ruining He, Wang-Cheng Kang and Julian Mcauley
Relevance to RecSys: 5 (excellent)
Novelty: 5 (excellent)
Technical Quality: 5 (excellent)
Evaluation: 5 (excellent)
Presentation and Readability: 5 (excellent)
Overall evaluation: 5 (strong accept)
Reviewer's confidence/expertise: 4 ((high))
----------- Overall evaluation -----------
An important problem in recommender systems is to take into account
recent sequences of events. For example, if one has seen the first
two movies "Matrix", one should maybe get a recommendation for
the third movie (though maybe not). This is sometimes called
"next-basket recommendation".
We can solve such a problem by modeling pairwise relationships between
items. It seems that most of the techniques proposed to solve the
next-basket recommendation problem rely on pairwise modeling. The
authors believe we can do better to model third-order relationships.
To achieve the desired result, they use a particular embedding
where users are represented by a vector with a nearest-neighbor search.
The paper is well written, well presented. There is a thorough
and credible experimental validation that seems to be positive
for the proposed technique.
The authors have an extensive dataset that is valuable in itself (4 million users!),
and they promise it will be made available upon publication of their paper.
All in all, I think it is clear that this work should be accepted.
I have some minor nitpicking comments:
1. The authors number all of their equations... unnecessarily.
2. English words in an equation should not be italicized (e.g.,
"aggregation").
3. Though this is up for debate, I would put a comma after "e.g.".
4. Please go easy on the significant digits (Table 4). Something
like "32.54%" makes no sense scientifically in this context.
----------------------- REVIEW 2 ---------------------
PAPER: 64
TITLE: Translation-based Recommendation
AUTHORS: Ruining He, Wang-Cheng Kang and Julian Mcauley
Relevance to RecSys: 5 (excellent)
Novelty: 4 (good)
Technical Quality: 3 (fair)
Evaluation: 3 (fair)
Presentation and Readability: 4 (good)
Overall evaluation: 3 (borderline)
Reviewer's confidence/expertise: 5 ((expert))
----------- Overall evaluation -----------
This paper propose a new model for addressing the next item recommendation problem by borrowing the idea from knowledge base. The essential idea is that users and items are represented as vectors and then the summation of representation of user and current item is close to the representation of next item. This translation has been extensively studied in knowledge based to learn object representation and relationships between objects. The idea of using this translation for next item recommendation is interesting and makes sense.
(1) The model assumes that items are bought in sequential order that only one item is bought at a time. However, in practice, it is very common for users to buy two or more items together. In this case, how to use the translation model to predict the next item?
(2) The proposed model seem to work only on recommendations in a single domain. For items that are from different domains such one item from electronics and the next item from kitchen, then the model doesn't work. Can the model be improved for recommending items in different domains?
(3) The next item shown in Figure 5 is not convincing. It seems that for such sequence, frequently bought together can give similar result. This brings another question that what's the performance difference between the proposed model and frequently bought items?
(4) Why use max(1, ||\gamma||) as the denominator for normalization instead of ||\gamma||? Using max(1, ||\gamma||) as normalization term may not result in a unit norm vector.
(5) Why do you choose Hit@50? What's the effect of using Hit@K with K being 1, 5, 10?
(6) The translation vector t seems to be unnecessary in Eq.(2) as there are no constraint on t. Consider that without t, we fist learn tu for each user, we can simply extract a common t from all tu, which gives the same effect.
----------------------- REVIEW 3 ---------------------
PAPER: 64
TITLE: Translation-based Recommendation
AUTHORS: Ruining He, Wang-Cheng Kang and Julian Mcauley
Relevance to RecSys: 5 (excellent)
Novelty: 4 (good)
Technical Quality: 4 (good)
Evaluation: 4 (good)
Presentation and Readability: 5 (excellent)
Overall evaluation: 5 (strong accept)
Reviewer's confidence/expertise: 5 ((expert))
----------- Overall evaluation -----------
Very well written paper. Authors have motivated the problem space well. Thorough literature survey. Approach is well described. Excellent data set and methods for comparison. Overall an enjoyable paper to read and highly relevant to the recsys community.
I would have liked a deeper discussion in areas where Transrec performance is not as significant or even poor. Give some intuition to the user where based on data or the nature of the problem this approach may not work.
------------------------- METAREVIEW ------------------------
PAPER: 64
TITLE: Translation-based Recommendation
RECOMMENDATION: accept
The paper represents an interesting, fresh approach to the "next item" problem in recommender systems, by embedding items into a "transition space" and modelling users as "translation vectors" from one item to the next. For an 8-page conference paper, the paper provides a good level of technical, theoretical detail as well as a very comprehensive evaluation. The review team is fairly consistent in its opinion that it is a high-quality paper.