News recommendation has been a must-have service for most mobile device users to know what has happened in the world. In this paper, we focus on recommending latest news articles to new users, which consists of the new user coldstart challenge and the new item (i.e., news article) coldstart challenge, and is thus termed as dual cold-start recommendation (DCSR). As a response, we propose a solution called neighborhood-based transfer learning (NTL) for this new problem. Specifically, in order to address the new user cold-start challenge, we propose a cross-domain preference assumption, i.e., users with similar app-installation behaviors are likely to have similar tastes in news articles, and then transfer the knowledge of neighborhood of the coldstart users from an APP domain to a news domain. For the new item cold-start challenge, we design a category-level preference to replace the traditional item-level preference because the latter is not applicable for the new items in our problem. We then conduct empirical studies on a real industry data with both users' app-installation behaviors and news-reading behaviors, and find that our NTL is able to deliver the news articles more accurately than other methods on different ranking-oriented evaluation metrics.
Intelligent recommendation systems [5] have been a ubiquitous service in our daily life, which has saved us a lot of time in finding proper information such as music, goods and news articles. For instance, personalized news recommendation [1, 2] has been one of the must-have services for most mobile device users, which plays an important role in helping users keep up with the current affairs in the world. In this paper, we focus on recommending latest news articles to new users, i.e., the users are newly registered in a certain news recommendation service and have not read any news articles before, and the news articles have not been read by any users before. We term it as dual cold-start recommendation (DCSR), denoting both cold-start users and cold-start items.
For the dual cold-start problem, previous news recommendation methods [1, 2] are not applicable, because they rely on users’ historical reading behaviors and news articles’ content information that are not available in our case.
We turn to address the cold-start recommendation problem from a transfer learning [3, 4] view. Although there are no users’ behaviors about the cold-start users and cold-start items in the news domain, there may be some other related domains with users’ behaviors. Specifically, we leverage some knowledge from a related domain, i.e., APP domain, where the users’ app-installation behaviors are available. We find that most cold-start users in the news domain have already installed some apps, and that may be helpful in determining his/her preferences in news articles. In particular, we assume that users with similar app-installation behaviors are likely to have similar interests in some news topics. In other words, close neighbors in the APP domain are likely to be close neighbors in the news domain.
With the above cross-domain preference assumption, we propose to take the neighborhood in the APP domain as the knowledge and try to transfer it to the target domain of news articles. Specifically, we design a neighborhood-based transfer learning (NTL) solution that transfers knowledge of neighborhood from the APP domain to the news domain, which addresses the new user cold-start challenge. With the neighborhood, some well-studied neighborhood-based recommendation methods are applicable for news recommendation.
We conduct empirical studies on a real industry data in order to verify our cross-domain preference assumption and the effectiveness of our transfer learning solution. Experimental results show that the two domains of apps and news articles are indeed related and can share some knowledge for preference learning.
tion. Mathematically, the preference prediction rule for user u to item i is as follows,
In our studied news recommendation problem, we have two domains, including an APP domain and a news domain.
Firstly, in the APP domain, we have a set of triples, i.e., (u, g, Gug), denoting that user u has installed Gug times of mobile apps belonging to the genre g. The data of the APP domain can then be represented as a user-genre matrix G as shown in Figure 1.
Secondly, in the news domain, we have a user-item matrix R denoting whether a user has read an item. Each item i is associated with a level-1 category c1(i) and a level2 category c2(i). We thus have a set of quadruples, i.e., (u, i, c1(i), c2(i)), denoting that user u has read an item i belonging to c1(i) and c2(i). Finally, we have a user-category matrix C after pre-processing, where each entry denotes the number of items belonging to a certain category that a user has read.
Our goal is to recommend a ranked list of new items (i.e., latest news articles) to each new user, who has not read any items before. We can see that it is a new user cold-start and new item cold-start problem, which is thus termed as dual cold-start recommendation (DCSR). Note that we only make use of items’ category information, but not content information.
We put some notations in Table 1. 2.2
The main difficulty of the DCSR problem is the lack of preference data for new users and new items. Specifically, there are two challenges, including (i) the new user cold-start challenge, i.e., the target users (to whom we will provide recommendations) have not read any items before; and (ii) the new item cold-start challenge, i.e., the target items (that we will recommend to the target users) are totally new for all users. Under such a situation, most existing recommendation algorithms are not applicable. 2.3
In most recommendation methods [5], the user-user (or
item-item) similarity is a central concept, because the
neighborhood can be constructed for like-minded users’ preference
aggregation and then for the target user’s preference
predic(
rˆu,i = 1
X rˆu′,i, |Nu| u′∈Nu where Nu is a set of nearest neighbors of user u in terms of a certain similarity measurement such as cosine similarity, and rˆu′,i is the estimated preference of user u′ (a close neighbor of user u) to item i. The aggregated and normalized score rˆu,i is taken as the preference of user u to item i, which is further used for item ranking and top-K recommendation.
For our studied dual cold-start recommendation problem, we can not build correlations between a cold-start user in the test data and a warm-start user in the training data using the data from the news domain only. The main idea of our transfer learning [3] solution is to leverage the correlations among the users in the APP domain with the assumption that users with similar app-installation behaviors are likely to be similar in news taste. For instance, two users with the installed apps of the same genre business may both prefer news articles on topics like finance.
With the cross-domain preference assumption, we first calculate the cosine similarity between a cold-start user u and a warm-start user u′ in the APP domain as follows, su,u′ =
Gu·GTu′· pGu·GTu·qGu′·GTu′· where Gu· is a row vector w.r.t. user u from the user-genre matrix G. Once we have calculated the cosine similarity, for each cold-start user u, we first remove users with a small similarity value (e.g., su,u′ < 0.1), and then take some (e.g., 100) most similar users to construct a neighborhood Nu.
For the item-level preference rˆu′,i in Eq.(
Finally, with the Eqs.(
X |Nu| u′∈Nu
X |Nu| u′∈Nu
Nu′,c1(i), Nu′,c2(i),
In our empirical studies, we use a real industry data, which consists of an APP domain and a news domain. APP Domain In the auxiliary domain, i.e., APP domain, we have 827,949 users and 53 description terms (i.e., genres) of the users’ installed mobile apps, where the genres are from Google Play. Considering our target task of news recommendation, we removed 14 undiscriminating or irrelevant genres such as tools, communication, social, entertainment, productivity, weather, dating, etc. Finally, we have a matrix G with 827,949 users (or rows) and 39 genres (or columns), where each entry represents the number of times that a user has installed apps belonging to a genre.
News Domain In the target domain, i.e., news domain, we have two sets of data, including a training data and a test data. The training data spans from 10 January 2017 to 30 January 2017, and contains 806,167 users, 747,643 items (i.e., news articles), and 16,199,385 unique (user, item) pairs. We can see that a user has read about 16199385/806167 = 20.09 articles on average from 10 January 2017 to 30 January 2017. The test data are from the data on 31 January 2017, which contains 3,597 new users, 28,504 new items (i.e., news articles), and 4,813 unique (user, item) pairs. We can see that a cold-start user read about 4813/3597 = 1.34 articles on 31 January 2017. Note that we have |C1| = 26 level-1 categories and |C2| = 222 level-2 categories about the items in the news domain.
For performance evaluation, we adopt some commonly used evaluation metrics in ranking-oriented recommendation such as precision, recall, F1, NDCG and 1-call. Specifically, we study the average performance of the top-15 recommended list generated for each cold-start user in the test data. 3.2
We compare our proposed transfer learning solution with a random method and two popularity-based methods using category information.
• Random recommendation (Random). In Random, we randomly select K = 15 items in the test data for each cold-start user. • Popularity-based ranking via level-1 category (PopRankC1). In PopRank-C1, we first calculate the popularity pc1 of each level-1 category c1 ∈ C1 in the training data, and then use rˆi = pc1(i) in Table 1 as the score to rank each item (i.e., article) i in the test data. For the most popular level-1 category, there may be more than K = 15 items (i.e., articles) in the test data, we then randomly take K items (i.e., articles) from that level-1 category for recommendation. • Popularity-based ranking via level-2 category (PopRankC2). In PopRank-C2, we use rˆi = pc2(i) in Table 1 as the prediction rule similar to that of PopRank-C1.
For the number of neighbors in our neighborhood-based transfer learning method, we first fix it as 100, and then change it to 50 and 150 in order to study its impact. We denote our transfer learning solution with level-1 category as NTL-C1 and that with level-2 category as NTL-C2, where their prediction rules are shown in Eq.(6) and Eq.(7), respectively. Note that for Random, PopRank-C1, PopRank-C2, and NTL with randomly selected neighbors, we repeat the experiments for 10 times, and report the average results. 3.3
We report the main results in Table 2. From Table 2, we can have the following observations: • The overall performance ordering is PopRank-C1, Random, PopRank-C2 ≪ NTL-C2 < NTL-C1, which clearly shows the effectiveness of our proposed transfer learning solution to the challenging dual cold-start recommendation problem. • The performance of PopRank-C2 and PopRank-C1 are rather poor in comparison with our proposed solution. The reason is that popularity-based methods are nonpersonalized methods and will simply select one most popular level-1 category or level-2 category for all users, which ignores the difference in users’ news reading preferences. • For the relative performance of NTL-C1 and NTL-C2, we can see that NTL-C1 performs better as expected because the level-1 category may introduce more smoothing effect for the cold-start problem.
We further study the impact of the neighborhood size. The results of our NTL-C1 using 50, 100 and 150 neighbors are shown in Figure 2. We can see that the results are relatively stable with different numbers of neighbors, and configuring it as 100 usually produces the best performance. 0.08 0 50
100 Neighborhood size 150
In order to gain some deep understanding about the transferred neighborhood, we study the performance of randomly choosing the same number (i.e., 100) of neighbors in our NTL-C1. We report the results in Figure 3, from which we can have the observations: • The neighborhood constructed using the app-installation behaviors is better than that of the random counterpart, which shows that the two domains are related and can indeed transfer knowledge from one domain to the other. • The difference between the two types of neighborhood is not as large as that between popularity-based methods and our NTL in Table 2, which can be explained by the fact that a large portion of users’ preferences or tastes in news articles are similar.
In this paper, we study an important and challenging news recommendation problem called dual cold-start recommendation (DCSR), which aims to recommend latest news articles (cold-start items) to newly registered users (cold-start users). Specifically, we propose a neighborhood-based transfer learning (NTL) solution, which is able to address the new user cold-start challenge and the new item cold-start challenge by the transferred neighborhood from the APP domain and the category-level preferences in the news domain, respectively. Empirical results on a real industry data show that our NTL performs significantly more accurate than the few applicable methods, i.e., popularity-based ranking using category information.
For future works, we are interested in selecting some representative genres and categories in two domains and building a mapping between them, which will be further used to study the neighborhood of the items.