============================= Meta Review Comments ======================================
    * Meta Review:
 

This metareview gives a summary of the different reviewers points of views and of the discussion that followed the author rebuttals phase.

Summary of strong points:

-        The ideas are good overall

-        The proposed contribution is assessed though experiments over 4 real-world data sets, showing improvements in the obtained results.

-        Good technical quality. The contribution is technically sound.

-        The paper is easy to follow and the necessary material is provided in the paper to understand the contribution.

-        The claimed contributions are well supported by technical informations and reproducible experiments.  

-        The presented results, conducted with 4 real-world data sets,  show the interest of the contribution.

-        The strengths and weakness are also generally well discussed by the authors in the setting of the considered approaches.

-        The paper is generally well written and well organized. The problem is well stated and the contribution is clearly explained.

 

Summary of weak points:

-        lack of clarity of some terms. Explanations in the rebuttal text are convincing. Authors should include these explanations in the final version if the paper is accepted.

-        Limited novelty: the paper proposes a new recommendation algorithm that combines sequential and social information, while dealing with the cold start problem. There is some novelty in that the proposed combination improve the results of the combined algorithms considered separately. However, the paper does not sufficiently discuss the impact of the proposed contribution on the general area of recommender systems.

-        Authors need to include more comments on the limitation of their approach.

-        Authors need to discuss the results at a broader scope, considering other kind of existing technics for recommender systems, to show the impact of the contribution at a broader scope.

-        Include more explanations. In the rebuttal authors convince reviewers about certain concerns. Authors will add more explanations in the  final version of the paper.

 

 
 

   
   ============================= Reviewer Comments  #22059 =========================
 * Originality: 6
 * Technical Quality: 5
 * Significance: 5
 * Relevance: 8
 * Quality of writing: 5
 * Overall Score: 5
 * Comments:
This paper proposes a novel model, called SPMC, tackling the cold-start problems for recommender systems. The model utilizes the feedback from sequences and social interactions, and predicts the next action of a user based on three ingredients, i.e., the user’s tastes, recent actions and the actions of the neighbourhood. The ideas are good overall, while the paper can be improved from the following aspects:

 

In the modelling section, the authors simply model the predictor as sum of user preferences, sequential dynamics, social-temporal dynamics and item bias. However, the weights of the three factors are ignored. Specifically, some users’ ratings to items are mainly dependent on their intrinsic features, such as preferences, while the others may be softheaded, so that their ratings can be affected by the social context to a large extent. In the proposed model, it is not clear whether the item bias parameter and the model training is capable of resolving this, please give clarifications.

 

Some terms are not clearly defined. For example, (1) most “recent” actions. How do you define the fuzzy word “recent”? Strictly based on a pre-defined threshold or a time decay? (2) “positive” feedback. Since most users’ ratings are presented as figures, how do you define positive? Then, do you consider negative feedback? Moreover, since different users have different standard, given a rating range: [0,10], the score 5 could be positive for user u_i, but can be negative for user u_j. How do you handle this?

 

In the learning model, only the positive item set as considered. How about if this positive feedback set is empty? Does the training fail? How about if a user only gives negative feedback? 

 
   
   ============================= Reviewer Comments  #22818 =========================
 * Originality: 6
 * Technical Quality: 8
 * Significance: 6
 * Relevance: 8
 * Quality of writing: 8
 * Overall Score: 6
 * Comments:
The paper addresses the topic of recommender system, while dealing with sparse information and cold start problems. The contribution is related to recommender systems, based on collaborative filtering, which combines two state-of-the art technics to leverage the recommendation by both sequential (recent activity sequence) and social (friends influence) information. The proposed contribution is assessed though experiments over 4 real-world data sets, showing improvements in the obtained results when compared to the isolated instances of the combined algorithms.



Originality : The paper proposes a new recommendation algorithm that combines sequential and social information, while dealing with the cold start problem. Although the proposed algorithm is a combination of existing state of the art matrix-factorization technics, there is some novelty in that the proposed combination improve the results of the combined algorithms considered separately. However, the paper does not sufficiently discuss the impact of the proposed contribution on the general area of recommender systems (are the considered algorithms the best known algorithms in the research area? How does these algorithms compete with other type of technics (not based on matrix-factorization), content based approaches, or hybrid approaches? what are their limitations? etc.

Technical Quality: The contribution is technically sound. The necessary material is provided in the paper to understand the contribution. The claimed contributions are well supported by technical informations and reproducible experiments.  The presented results, conducted with 4 real-world data sets,  show the interest of the contribution. The strengths and weakness are also generally well discussed by the authors in the setting of the considered approaches. It would have been interesting to discuss the results at a broader scope, considering other kind of existing technics for recommender systems, to show the impact of the contribution at a broader scope.

Significance : The proposed contribution is significant in that it concretely shows how combining social information and information on the user’s sequence of recent activities (sequential information) improve the performance of recommender systems, while dealing with sparse information and cold start problems.

Relevance: The paper is fully in the scope of the conference and of interest for both learning and recommender systems communities.

Quality of writing: The paper is generally well written and well organized. The problem is well stated and the contribution is clearly explained.

 
   
   ============================= Reviewer Comments  #28055 =========================
 * Originality: 7
 * Technical Quality: 5
 * Significance: 7
 * Relevance: 7
 * Quality of writing: 6
 * Overall Score: 6
 * Comments:
 

The main goal of this paper is resolve the cold-start problem in the Factorized Personalized Markov Chains (FPMC). FPMC uses the user basket’s history. The authors try to improve the recommender system when this history hasn’t got enough information. To do it, they added social factors to calculate the recommendation. The novel model is called SPMC.

The novelty of this paper is the inclusion of social factors in the method FPMC. On the other hand, the use of social factors to recommend products is an active area in this moment and there are many papers published about this.

The equations are difficult to understand and the paper lacks from explanations. Weights are not specified. The general meaning of the equations is easy to understand but additional explanations are required for latent factors.

Experiments show good results within 4 different datasets. They improve between 1% and 20% to FPMC accuracy rate. And in the cold-start problem they have the best results. 

In summary, the method is original and it works very well, but the paper would benefit from a review of equations and their explanations.

 
   
   ============================= Reviewer Comments  #32237 =========================
 * Originality: 4
 * Technical Quality: 6
 * Significance: 6
 * Relevance: 8
 * Quality of writing: 8
 * Overall Score: 6
 * Comments:
In this paper the authors consider the problem of item recommendation in settings where both sequential and social information exists. Previous work has focused either on the sequential aspect of recommendation, or on leveraging social context. The authors use both simultaneously in their model, and show their approach is particularly beneficial for dealing with the cold start problem in four large real world datasets.

The problem is quite interesting - undoubtedly both sequential context and social cues play a significant role in recommendation, and therefore leveraging both simultaneously seems like a promising direction. The paper is well written and motivated, and the approach seems solid, albeit somewhat straightforward. However, the paper is not without issues. My comments/questions are as follows.

Core technical questions:

* Unless I am mistaken, essentially the model only captures one step transition probabilities - this seems potentially rather limiting. In general it seems that temporal context is more sensitive to time than to number of transitions, or at least the two are codependent.

* The choice of a linear combination of socio-temporal context, sequential dynamics, and inherent user preferences, seems somewhat odd given that they all seem strongly codependent.

* Are all friends weighted the same? It would make a lot more sense to weight them relative to their mutual affinity. Also, it seems that different friends influence differently in different contexts. How is the friend set F_u determined?

* The derivation is contingent on the assumption that "all users are independent and for each user all adjacent item pairs are independent", but this assumption seems clearly wrong...

Questions regarding the experiments:

* In both the Ciao and the Epinions datasets, it seems that increasing the threshold does not improve the performance of SPMC - this seems rather counter-intuitive to me, what is the reason for this?

* How statistically significant are these results? In some cases the margins seems rather small.

* What happens when the cold start is even colder? Say N=1? Does SPMC still outperform the rest?

* Have you looked at performance using other prediction metrics (say mean average precision)?

Lastly, regarding the related work section, at least with respect to sequential dynamics, there seems to be quite a bit of related literature that's overlooked, including but not limited to the following:

- "Playlist Prediction Via Metric Embedding", Chen, Moore, Turnbull and Joachims, KDD 2012

- "Which app will you use next?: collaborative filtering with interactional context", Natarajan, Shin and Dillon, 2013

- "DJ-MC: A Reinforcement-Learning Agent for Music Playlist Recommendation", Liebman, Saar-Tsechansky and Stone, AAMAS 2015

- "RLCF: A Collaborative Filtering Approach Based on Reinforcement Learning With Sequential Ratings", Lee, Oh, Yang & Park, Soft Computing 2016