The cold start problem, frequent with recommender systems, addresses the issue in cases where we don't know enough about our users (e.g., the user hasn't rated anything yet, or there are no user activities) in that specific domain. In our paper we present a simple and robust transfer learning approach where we model users' behavior in a source domain, transferring that knowledge to a new, target domain. First, we vectorize the items by using word2vec for each dataset independently. Second, we calculate the transformation matrix that connects the source dataset to the target dataset by using their common users.
The economic advantages of cross selling products between different domains (e.g., movies or books) are appealing and cross-domain recommendation has become a hot topic in recent years. However, while it is not always trivial how to identify patterns that successfully model user behavior, doing so across domains is even more difficult. If one tends to like David Lynch movies, we may be able to recommend “The Elephant Man”, but what does this knowledge translates to in another similar yet different domain, such as the books domain? What about in an apparently unrelated domain, such as the restaurants domain?
In this paper we propose a very simple and robust transferlearning based approach that can successfully model the behavior patterns of users across different domains, showing promising results even with only a small number of common users.
In recent years, numerous transfer learning approaches
have been proposed. Sahebi et.al. [
Consider the case when we have a source domain and a new, target domain with no common items, but common users. Such cases are very frequent in real-life situations. We propose a cross-domain recommender methodology that can accurately recommend items from the entire inventory of the target domain, leveraging on the common users. We illustrate the scope of our paper in figure 1. Our method captures analogies in user behavior across domains by modeling user preference in the source model, transferring this knowledge to a new, target domain. The prerequisite for our proposal is only the existence of a relatively small group of common users.
The proposed method consists of two steps. First, we generate the vector representation for each item, separately for both the source and target domains. We use word2vec for this generation process. Word2vec is usually employed for text analysis, but in our process we use it for user rate logs, treating items as words, with each users’ positive items organized into paragraphs. Secondly, using the common users, we learn a transformation matrix M between the source and target domains.
N M = arg min XkM uis − uik
t M i=1 (1) where uis(∈ RK ) is the vector of user i in the source vector t space, ui in the target vector space. User vectors are calculated as the mean of the item vectors rated by the user as positive. K is the dimension of each vector. N is the number of common users. M (∈ RK×K ) is the transformation matrix between source and target spaces. Source dataset leveraged recommendation is performed by calculating the user vector and converting it to the target vector space by using the transformation matrix M . Once this conversion is achieved, the converted user vector can be compared to the item vectors in the new domain and items with most similar vectors can be recommended to this user. In this paper, the similarity measure used is the cosine similarity.
In our experiments we used the ratings from the yelp dataset1. User ratings range from 1 to 5, but we only consider positive ratings (4 and 5), ignoring the rest. Items are tagged by various labels (e.g. Restaurants, Shopping).
We validated our proposal on the 3 most frequent categories in terms of item count: Restaurants, Shopping and Food. For each source-target domain, we removed the items which were labeled in both categories. For example, in the Restaurants-Shopping pairing, after removing the common items, we retained 131,825 and 10,899 users respectively, among them, 8,642 were common users. Table 1 summarizes the statistics of the above categories. As test data, we used 20% of the common users in all source-target pairing configurations.(e.g., 1728 users in the Restaurant-Shopping settings). The rest of data is used for training. In order to observe the effect of the number of common users, we incrementally increased the number of common users in our training data, starting from 10% of the training data and increasing by 10% at each step. For the evaluation metric we used recall@100, as we are interested in correctly identifying the items the users actually rate as positive in target domain, using the data only from the source domain. Besides recall@100, we experimented with other 1https://www.yelp.com/dataset challenge metrics such as recall@10 and @50, precision@10, @50 and @100. However, the same tendencies were observed in all experiments. As baseline, we treat the source and target domains as a single dataset, generating one common vector space model using word2vec in a similar way with our proposal, with the recommended items being the ones whose vector is closest to the user’s vector. 5.
We proposed a simple, but robust transfer learning approach that can bridge the gap across differing domains. We found that our method is most efficient (1) when the number of users in the source domain is smaller than the number of the users in the target domain and (2) when the number of common users is small across the two domains.
We are investigating the efficiency of other item vector representation methods besides word2vec, such as matrix factorization within our transfer learning framework.