ExcUseMe: Asking Users to Help in Item Cold-Start Recommendations

Summary

This paper is mainly about the solution of item cold start problem in the real world recommender system. In an on-line recommender system, the cold start problem is caused by lacking of the historical data to generate meaningful suggestions. In other words, the cold start problem can be treated as the user or item that newly enter into the system. In general, the cold start problem was solved by include new context or content features. However, in this paper, the author focus on the specific solution for collaborative filtering (CF). Their idea is conducting a small scale user experiment to those users who might interested in the new item. The user interaction within the experiment can be used as a reference for the new item recommendation.

The major challenge behind this idea is: Who is the target user? In this paper, they proposed a novel algorithm "ExcUseMe". The authors assumed the user is randomly visited the online recommender system. In this continuing user stream, the system required to determine who should be included in the new item experiment. The approach can be divided into two stages: 1) the learning phase; 2) the selection phase. In 1), the algorithm is computing the likelihood of the user provide feedback of the item. In stage 2), the user is selected by their likelihood ranking. The vector similarity between the users are also considered in a semi-greedy manner.

The experiment simulated the real world on-line setting, they sampled the n% users as selection pool. The performance was evaluated by RMSE (Root-mean-square deviation) metric for three large scale datasets. The baseline was compared with Random, Frequent Users, Distance and Anava approaches. The experiment result indicated the proposed ExeUseMe algorithm outperform all the baseline models, moreover, with a lower computational cost. The experiment result also pointed out the importance and contribution of positive feedback from users. In other words, the key to drive this method as a meaningful output is the user participants.

This paper highlight a new approach for CF to solve the cold start problem. Besides, this approach is also practical that can be adopted into the real world on-line system. The main contributions of this paper are 1) the real-world user selection method; 2) efficiency algorithm and 3) the high model performance in most of the scenarios. There are some comments in this paper: 1) this algorithm is valuable in an on-line experiment setting, but the experiment is simulated with off-line setting and dataset. A further simulation experiment can better demonstrate the advantage of this approach; 2) the data filtering section limited the user with between 20 - 300 ratings. This setting reduced the probability to provide useful suggestions for long-tail items; 3) In real-world recommender system, the small amount of user feedback lead to the result of a recommendation. There might be some incentive to inverse control the experiment result in some ways.

Reference

Michal Aharon, Oren Anava, Noa Avigdor-Elgrabli, Dana Drachsler-Cohen, Shahar Golan, and Oren Somekh. 2015. ExcUseMe: Asking Users to Help in Item Cold-Start Recommendations. InProceedings of the 9th ACM Conference on Recommender Systems (RecSys '15). ACM, New York, NY, USA, 83-90. DOI=http://dx.doi.org/10.1145/2792838.2800183