Promoting cold-start items in recommender systems

Promoting cold-start items in recommender systems

Abstract

As one of major challenges, cold-start problem plagues nearly all recommender systems. In particular, new items will be overlooked, impeding the development of new products online. Given limited resources, how to utilize the knowledge of recommender systems and design efficient marketing strategy for new items is extremely important. In this paper, we convert this ticklish issue into a clear mathematical problem based on a bipartite network representation. Under the most widely used algorithm in real e-commerce recommender systems, so-called the item-based collaborative filtering, we show that to simply push new items to active users is not a good strategy. To our surprise, experiments on real recommender systems indicate that to connect new items with some less active users will statistically yield better performance, namely these new items will have more chance to appear in other users’ recommendation lists. Further analysis suggests that the disassortative nature of recommender systems contributes to such observation. In a word, getting in-depth understanding on recommender systems could pave the way for the owners to popularize their cold-start products with low costs.

pacs:
89.75.Hc, 89.20.Hh, 05.70.Ln

I Introduction

Thanks to the blazing development of Internet, e-commerce has flourished over the past decades. With the online buy-and-sell platforms getting increasingly more available products (e.g., more than a billion products in taobao.com), shopping online has become a fashionable style of living and more people choose to purchase on the Internet rather than go to a store. E-commerce makes our life much more convenient, meanwhile, it throws us into a dilemma of information overloads. Facing millions of items online, finding out favourites is rather difficult. As an effective information filtering tool, recommender system is thus of particular significance nowadays (1); (2). In fact, it has already made considerable contributions to the socioeconomic fields in the past decade. For example, 60% of DVDs rented by Netflix are selected based on personalized recommendations, and about a half of sales in Amazon are brought by recommendations (2). Consequently, recommender systems have received huge attentions from both physicists and computer scientists, and many advanced recommendation algorithms are proposed recently, including collaborative filtering (3); (4); (5); (6); (7); (8), content-based analysis (9); (10); (11), dimensionality reduction techniques (12); (13); (14), diffusion-based methods (15); (18); (19); (16); (17); (20); (21); (22), and so on.

One long-standing challenge, called cold-start problem, has plagued almost all recommender systems. Namely, when new users or items enter the system, there is usually insufficient information to produce reasonable recommendation (23). Considering this fact, several potential solutions have been raised. The additional content information (23); (24); (25); (26), tagging information (27); (28); (29) and cross-domain information (30) can be used to marginally relieve this problem, but they don’t work in a purely cold-start setting, where no information is available to form any basis for recommendations. Furthermore, improving diversity and novelty of recommended lists can help new items be pushed out (31); (17); (32).

Practically speaking, as a holder of the recommender system, one can ask for extra information to generate initial profiles on users or items (24), or probe users’ preferences by pushing to them some carefully selected items according to complicated algorithms (33). Both methods are costly and risky. In contrast, an owner would like to popularize her new items. An improper method, called “shilling attacks”, injects a number of mendacious users into the system to raise predicted ratings of new items, and thus enhances the possibility of these new item to appear in the recommendation lists (34); (35). But, it is easily to be detected (36); (37); (38). Furthermore, as a wide-spreading market strategy, advertisements are generally preferred and become more and more prosperous (39). However, to popularize new items costs a lot and imposes an unbearable financial burden for small businesses (40). As mentioned above, how to promote new items under limited marketing resources is a nontrivial challenge and the knowledge of recommendation algorithm may be helpful. Putting aside operational details, if the marketing activities can bring some purchases of certain users, a smart marketing manager will carefully chose the target users so that these purchases can lead to more exposures in the recommendation lists afterwards.

Taking a stand as a marketing manager, in this paper, we focus on how to promote cold-start items by utilizing the knowledge of recommender systems. The main contributions are threefold: (i) We convert this ticklish problem into a clear mathematical model that ignores some insignificant details. (ii) We show that to push new items to active users, a straightforward strategy that will jump into our mind at the first time, is to our surprise a poor-performed strategy. (iii) We propose a degree-based solution that outperforms some baseline methods.

Ii Results

Recommendation can be considered as a variant of link prediction in bipartite networks (41) and thus the better understanding of network structures can in principle improve the quality of recommendations (42); (43); (44); (45). We denote a recommender system by a user-item bipartite network , where and are respectively the sets of users and items, and is the set of links connecting users and items. Consequently, we use the adjacent matrix, , to describe the user-item relations: if user has purchased item , , otherwise (throughout this paper we use Latin and Greek letters, respectively, for user- and item-related indices). Figure 1(a) illustrates a small bipartite network that consists of eight users (gray squares) and eight items (blue circles). , the degree of user , is defined as the number of items linked to . Analogously, the degree of item , denoted by is the number of users connected to . For example, as shown in Figure 1(a), and . The user degree distribution , is the probability that the degree of a randomly selected user, is equal to . The item degree distribution is defined in a similar way. Degree distribution reflects the network heterogeneity (46).

Figure 1: How to add a cold-start item to the user-item bipartite network. Users and items are represented by squares and circles respectively, and solid lines represent the existent links between them. Plot (a) is the original network, and plot (b) is the network after adding the item (the yellow circle). The dotted lines are new links connecting with two existent users.

We consider two real data sets with anonymous users in this paper, including (a) Tmall.com (TM): an open business-to-consumer (B2C) platform where enrolled businessmen can sell legal items to customers; (b) Coo8.com (Coo8): a well established online retailer mainly trading in electrical household appliances and a leading supplier to daily necessities. In order to avoid the isolate nodes in the data sets, each user has bought at least one item, and each item has been purchased at least once. Table 1 shows the basic statistics of the two data sets. Due to the different types of products, these networks have much different average item degrees. As shown in Figure 2, all degree distributions are heavy-tailed and the item degree distributions are generally more heterogenous than the corresponding user degree distributions. These observations complement previous empirical analyses on user-item bipartite networks (47); (48); (49); (50).

Data
TM 103,867 83,342 113,624 1.09 1.36 1.31
Coo8 77,947 18,751 94,457 1.21 5.04 6.46
Table 1: Basic statistical properties of the two data sets. , , and represent the number of users, items and links, and stand for the average degrees of users and items, and denotes the data sparsity.

The nearest neighbors’ degree for user , denoted by , is defined as the average degree over all items connected to (50). For example, in Figure 1(a), . Furthermore, the degree-dependent nearest neighbors’ degree, is the average nearest neighbors’ degree over all users of degree , namely . Corresponding concepts for items, , and , are defined in a similar way and thus omitted here (50). The degree-dependent nearest neighbors’ degree is an appropriate index to characterize the network assortativity (51). As shown in Figure 2, both the two networks are disassortative.

Figure 2: Degree distributions and degree correlations. All degree distributions are power-law-like, with exponents being estimated by the maximum likelihood methods (53); (54). and are respectively showed in the 3rd and 4th rows, where red and black lines representing the results from original and reshuffled networks. Results of reshuffled networks are obtained by averaging over five independent realizations.

Recommender systems typically produce a given-length list of unpurchased items for each user based on his historical purchases. Of nothing comes nothing, that is to say, it is impossible to predict links for an isolate user or item. So only after having been purchased by some users, an item could have the chance to appear in some other users’ recommendation lists. In real e-commerce web sites, to get a new customer is highly costly, and thus under the limited investment, choosing users with considerable coming influence on further recommendations is absolutely critical. Concretely speaking, this problem is described as follow. Given a bipartite network containing users, items and links. A novel item enters this network, and it can at most establish links to users. Given the recommendation algorithm, we need to answer the question that how to choose such users to maximize the frequency that appears in other users’ recommendation lists. For example, in Figure 1(b), item (the yellow circle) comes and needs to link to some existent users. If , then to choose which user, (most active user), (one of the most inactive users) or another one, can make be recommended more times?

We consider four strategies to choose those users: (I) Maximum-degree strategy (MaxD). To rank all users in the descending order of degree, and select the top- users, where users with the same degree are ranked randomly. (II) Minimum-degree strategy (MinD). To rank all users in the ascending order of degree, and select top- users, where users with the same degree are ranked randomly. (III) Preferential attachment strategy (PA). Each user’s probability to be selected is proportional to her degree. (IV) Random strategy (RAN). The users are selected completely randomly. Actually, all strategies above can be unified by a selecting probability of any user , , where is a tunable parameter. More specifically, the strategies MaxD, MinD, PA and RAN correspond to the cases of , , and , respectively.

Among existent recommendation algorithms, item-based collaborative filtering (ICF) has found the widest applications in real e-commerce platforms for its accuracy, stability, scalability and robustness (5); (6); (38). Here, we apply cosine similarity for each pair of items, say

(1)

where and are degrees of items and , respectively. For the target user , we calculate the accumulative score for each item by

(2)

and then rank all the unpurchased items in descending order according to their scores in Eq. (2). The top- items will be recommended to , where is the length of recommendation list.

To compare the degree-dependent strategies, we employ a metric that counts the number of users whose recommendation lists contain the target items, say

(3)

where is the position of the target item among all ’s unpurchased items. Obviously, , since the target item’s degree equals , and the larger value of means better performance. The number of recommended items, , is limited by the user interface, with typical size no larger than 6 (see real recommendation engines of Alibaba Group and Baifendian Inc. as examples).

Since the maximum item degrees for TM and Coo8 are 617 and 933, respectively, in our simulation, we only consider ranging from 1 to 1000. To our surprising, as shown in figure 3, MaxD hardly makes new items recommended while MinD usually shows better performance. Consider the general case where the target item has established a link to user , and and are two of ’s collected items before . For another user who is not connected with . If has collected but not , then both and have the chance to be recommended to . Since in the ICF algorithm, item similarities play the major role, let’s compare the similarities and . Statistically speaking, if is a very active user selected by the MaxD strategy, and are probably less popular as indicated by the disassortative nature of the networks, therefore (i.e., ) may be much larger than and then is probably smaller than , resulting in less probability of to be recommended to . In contrast, if is a very inactive user selected by the MinD strategy, and are probably of larger degrees according to the disassortative nature, resulting in smaller and thus larger probability for to be recommended to . In addition, since is very unpopular, it is also possible that and is only connected with . In such case, for all other users connected with , will be the only recommended item related to .

Figure 3: Performance of the four strategies for original TM and Coo8 bipartite networks. The results of MaxD, MinD, PA and RAN are represented by black squares, red circles, blue triangles and green pentagrams, respectively. Data points are obtained by averaging over 50 independent realizations.
Figure 4: Performance of the four strategies for reshuffled networks. The results of MaxD, MinD, PA and RAN are represented by black squares, red circles, blue triangles and green pentagrams, respectively. Data points are obtained by averaging over 50 independent realizations.

In a word, the disassortativity could contribute to the observations in figure 3. To validate this inference, we reshuffle the original networks by link-crossing method to obtain the null networks (52). Specifically speaking, in each step, two links, say and , are randomly picked out, and if has not collected and has not collected , these two links are rewired as and . In one realization, we repeat such rewiring for times. After that, the reshuffled network has identical degree sequence as the original network but the disassortative nature is vanished as shown in figure 2. Figure 4 reports the performance of the four strategies in the reshuffled networks, from which we can see that the MaxD strategy performs the best. Comparing the results for original and reshuffled networks, we conclude that the advantage of MinD strategy results from the disassortative nature of real e-commerce user-item bipartite networks. In addition, in figure 5 and figure 6, we test the performance of strategies with different . For both TM and Coo8, the negative will lead to better performance while in the null networks, positive is better.

Figure 5: Performance of strategies with different on original and reshuffled TM networks. The black, red and blue lines represent the results for the cases , and , respectively. Data points are obtained by averaging over 50 independent realizations.
Figure 6: Performance of strategies with different on original and reshuffled Coo8 networks. The black, red and blue lines represent the results for the cases , and , respectively. Data points are obtained by averaging over 50 independent realizations.

Iii Discussion

In this paper, we study a practical problem in e-commerce recommender systems: how to promote cold-start items? Under the item-based collaborative filtering systems, we show that the disassortative nature of real user-item networks leads to a non-trivial observation that to link a cold-start item to inactive users will give it more chance to appear in other users’ recommendation lists. This observation is robust for varying recommendation length and linking capacity . It is also applicative to some variants of item-based collaborative filtering, such as the top- nearest neighbors ICF (5).

Notice that, the reported results are affected by both the topological features and underlying recommendation algorithms. We have tested the user-based collaborative filtering (3), under which the MaxD is always better than MinD. In spite of this, this work is still relevant since in most real recommender systems, ICF plays a significant role. In addition, the perspectives and methods reported here are useful for real e-commerce applications, with the core merit is that the in-depth understanding of the structure and algorithms of recommender systems can be transferred into applicable knowledge to better market products.

Iv Acknowledgments

We acknowledge Kuan Fang, Zhidan Zhao and Ying Zhou for useful discussions. We acknowledge Alibaba Group and Baifendian Inc. to share their real business data sets after anonymization. This work was partially supported by the National Natural Science Foundation of China under Grant Nos. 11222543. T.Z. acknowledges the Program for New Century Excellent Talents in University under Grant No. NCET-11-0070, and Special Project of Sichuan Youth Science and Technology Innovation Research Team under Grant No. 2013TD0006.

References

  1. F. Ricci, L. Rokach, B. Shapria, and P. B. Kantor, Recommender Systems Handbook (Springer, 2011).
  2. L. Lü, M. Medo, C. H. Yeung, Y.-C. Zhang, Z.-K. Zhang, and T. Zhou, Physics Reports 519, 1 (2012).
  3. D. Goldberg, D. Nichols, B. M. Oki, and D. Terry, Communications of the ACM 35, 61 (1992).
  4. U. Shardanand, and P. Maes, In Proceedings CHI’95 Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (ACM Press, New York 1995) pp. 210-217.
  5. B. Sarwar, G. Karypis, J. Konstan, In Proceedings WWW’01 Proceedings of the 10th international conference on World Wide Web (ACM Press, New York 2001) pp. 285-295.
  6. M. Deshpande, and G. Karypis, ACM Transactions on Information Systems 22, 143 (2004).
  7. J. B. Schafer, D. Frankowski, J. Herlocker, and S. Shilad, Lecture Notes in Computer Science Springer-Verlag, 2007.
  8. H. N. Kim, A.-T. Ji, I. Ha, and G.-S. Jo, Lecture Notes in Computer Science Springer-Verlag, 2010.
  9. M. Balabanovic, and Y. Shoham, Communications of the ACM 40, 66 (1997).
  10. M. J. Pazzani, and D. Billsus, Lecture Notes in Computer Science Springer-Verlag, 2007.
  11. K. Niemann, and M. Wolpers, In Proceeding KDD’13 Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining (ACM Press, New York 2013) pp. 955-963.
  12. T. Hofmann, ACM Transactions on Information Systems (TOIS) 22, 89 (2004).
  13. G. Takács, I. Pilászy, B. Námeth, and D. Tikk, ACM SIGKDD Explorations Newsletter 9, 80 (2007).
  14. K. Bunte, P. Schneider, B. Hammer, F.-M. Schleif, T. Villmann, and M. Biehl, Neural Networks 26, 159 (2012).
  15. Z. Huang, H. Chen, and D. Zeng, ACM Transactions on Information Systems (TOIS) 22, 116 (2004).
  16. T. Zhou, J. Ren, M. Medo, and Y.-C. Zhang, Physical Review E 76, 046115 (2007).
  17. T. Zhou, Z. Kuscsik, J.-G. Liu, M. Medo, J. R. Wakeling, and Y.-C. Zhang, Proceedings of National Academy of Science of the United Sates of America 107, 4511 (2010).
  18. Y.-C. Zhang, M. Medo, J. Ren, T. Zhou, T. Li, and F. Yang, Europhysics Letters 80, 68003 (2007).
  19. Y.-C. Zhang, M. Blattner, and Y.-K. Yu, Physical Review Letters 99, 154301 (2007).
  20. J.-G. Liu, T. Zhou, and Q. Guo, Physical Review E 84, 037101 (2011).
  21. L. Lü, and W. Liu, Europhysics Letters 83, 066119 (2011).
  22. Y. Zhou, L. Lü, W. Liu, and J. Zhang, PLoS ONE 8, e70094 (2013l).
  23. A. I. Schein, A. Popescul, L. H. Ungar, and D. M. Pennock, In Proceeding SIGIR’02 Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval (ACM Press, New York 2002) pp. 253-260.
  24. S.-T Park, and W. Chu, In Proceeding Recommender Systems’09 Proceedings of the third ACM conference on Recommender systems (ACM Press, New York 2009) pp. 21-28.
  25. I. Cantador, A. Bellogín, and D. Vallet, In Proceeding Recommender Systems’10 Proceedings of the forth ACM conference on Recommender systems (ACM Press, New York 2010) pp. 237-240.
  26. R. Ronen, N. Koenigstein, E. Ziklik, and N. Nice, In Proceeding Recommender Systems’13 Proceedings of the 7th ACM conference on Recommender systems (ACM Press, New York 2013) pp. 407-410.
  27. Z.-K. Zhang, C. Liu, Y.-C. Zhang, and T. Zhou, Europhysics Letters 92, 28001 (2010).
  28. Z.-K. Zhang, T. Zhou, and Y.-C. Zhang, Journal of computer science and technology 26, 767 (2011).
  29. D. Yin, S. Guo, B. Chidlovskii, and B. D. Davision, In Proceeding WSDM’13 Proceedings of the 6th ACM international conference on Web search and data mining (ACM Press, New York 2013) pp. 547-556.
  30. L. Zhang, L.-S. Bai, and T. Zhou, Journal of University of Electronic Science and Technology of China 42, 154 (2013).
  31. C.-N. Ziegler,S. M. McNee, J. A. Konstan, and G. Lausen, In Proceeding WWW’05 Proceedings of the 14th international conference on World Wide Web (ACM Press, New York 2005) pp. 22-32.
  32. T. Zhou, R. Q. Su, R. R. Liu, L. L. Jiang, B. H. Wang, and Y. C Zhang, New Journal of Physics 11, 123008 (2011).
  33. J.-L Zhou, Y. Fu, H. Lu, and C.-J Sun, Journal of Computer Science and Technology 26, 816 (2011).
  34. S. K. Lam, and J. Riedl, In Proceeding WWW’04 Proceedings of the 13th international conference on World Wide Web (ACM Press, New York 2004) pp. 393-402.
  35. B. Mobasher, R. Burke, R. Bhaumik, and C. Williams, In Proceedings of the 2005 WebKDD Workshop (ACM Press, New York 2005).
  36. B. Mobasher, R. Burke, R. Bhaumik, and C. Williams, Lecture Notes in Computer Science Springer-Verlag, 2006.
  37. B. Mehta, T. Hofmann, and P. Fankhauser, In Proceeding IUI’07 Proceedings of the 12th international conference on Intelligent user interfaces (ACM Press, New York 2007) pp. 14-21.
  38. B. Mobasher, R. Burke, R. Bhaumik, and C. Williams, ACM Transactions on Internet Technology 7, 1 (2007).
  39. P. C. Verhoef, and P. S. H. Leeflang, Journal of Marketing 73, 14 (2009).
  40. F. C. Edward, C. J. Hadlock, and J. R. Pierce, The Review of Financial Studies 22, 2361 (2008).
  41. L. Lü, and T. Zhou, Physica A 390, 1150 (2011).
  42. Z. Huang, D. D. Zeng, and H. Chen, Management Science 53, 1146 (2007).
  43. Y. Koren, and J. Sill, In Proceedings of the 5th ACM Conference on Recommender Systems (ACM Press, New York 2011) pp. 117-124.
  44. Q.-M. Zhang, A. Zeng, and M.-S. Shang, PLoS ONE 8, e62624 (2013).
  45. W. Zeng, A. Zeng, H. Liu, M.-S. Shang, and T. Zhou, arXiv 1402.6132 (2013).
  46. A.-L. Barabási, and R. Albert, Science 286, 509 (1999).
  47. R. Lambiotte, and M. Ausloos, Physical Review E 72, 066107 (2005).
  48. B. Goncalves, and J. J. Ramasco, Physical Review E 78, 026123 (2008).
  49. Z. Yang, Z.-K. Zhang, and T. Zhou, Europhysics Letters 100, 68002 (2012).
  50. M.-S. Shang, L. Lü, Y.-C. Zhang, and T. Zhou, Europhysics Letters 90, 48006 (2010).
  51. R. Pastor-Satorras, A. Vázquez, and A. Vespignani, Physical Review Letters 87, 258701 (2001).
  52. C.-J. Zhang, and A. Zeng, Physica A 391, 1822 (2012).
  53. M. L. Goldstein, S. A. Morris, and G. G. Yen, European Physical Journal B 41, 255 (2004).
  54. A. Clauset, C. R. Shalizi, and M. E. J. Newman, SIAM Review 51, 661 (2009).
100203
This is a comment super asjknd jkasnjk adsnkj
""
The feedback cannot be empty
Submit
Cancel
Comments 0
""
The feedback cannot be empty
   
Add comment
Cancel

You’re adding your first comment!
How to quickly get a good reply:
  • Offer a constructive comment on the author work.
  • Add helpful links to code implementation or project page.