Cross-domain Recommendation via Deep Domain Adaptation

03/08/2018
by   Heishiro Kanagawa, et al.
yahoo
The University of Tokyo
0

The behavior of users in certain services could be a clue that can be used to infer their preferences and may be used to make recommendations for other services they have never used. However, the cross-domain relationships between items and user consumption patterns are not simple, especially when there are few or no common users and items across domains. To address this problem, we propose a content-based cross-domain recommendation method for cold-start users that does not require user- and item- overlap. We formulate recommendation as extreme multi-class classification where labels (items) corresponding to the users are predicted. With this formulation, the problem is reduced to a domain adaptation setting, in which a classifier trained in the source domain is adapted to the target domain. For this, we construct a neural network that combines an architecture for domain adaptation, Domain Separation Network, with a denoising autoencoder for item representation. We assess the performance of our approach in experiments on a pair of data sets collected from movie and news services of Yahoo! JAPAN and show that our approach outperforms several baseline methods including a cross-domain collaborative filtering method.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

12/08/2021

A cross-domain recommender system using deep coupled autoencoders

Long-standing data sparsity and cold-start constitute thorny and perplex...
08/28/2018

Superhighway: Bypass Data Sparsity in Cross-Domain CF

Cross-domain collaborative filtering (CF) aims to alleviate data sparsit...
04/26/2021

Recommending Burgers based on Pizza Preferences: Addressing Data Sparsity with a Product of Experts

In this paper we describe a method to tackle data sparsity and create re...
12/10/2020

Using Affective Aware Pseudo Association Method to Connect Disjoint Users and Items for Making Cross-Domain Recommendations

This paper introduces an ingenious text-based affective aware pseudo ass...
05/26/2019

DARec: Deep Domain Adaptation for Cross-Domain Recommendation via Transferring Rating Patterns

Cross-domain recommendation has long been one of the major topics in rec...
06/28/2020

Semi-supervised Collaborative Filtering by Text-enhanced Domain Adaptation

Data sparsity is an inherent challenge in the recommender systems, where...
10/15/2021

Crisis Domain Adaptation Using Sequence-to-sequence Transformers

User-generated content (UGC) on social media can act as a key source of ...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1. Introduction

Traditional recommender systems usually assume user’s past interactions (e.g. ratings or clicks) to make meaningful recommendations. While this seems natural, however, the utility of such system becomes degraded when the assumption does not hold, e.g. when a new user arrives at the service or when we aim to perform cross-selling of products from an unused service. On the other hand, as the variety of web services has increased, information about cold-start users can be obtained from their activities in other services. Therefore, cross-domain recommender systems, which leverage such additional information from other related domains, have gained research attention in recent years as a promising solution to the user cold-start problem (Cantador et al., 2015).

In this paper, we deal with cross-domain recommendations for cold-start users, particularly in the absence of common users and items. Although it is possible to suggest items and obtain feedback from the target users (Rubens et al., 2015), we aim to build a recommender system that can work even when such alternative is not available. In this situation, a major challenge is that traditional methods can not be used for the elicitation of the relations between the two services due to the lack of common users.

A naive approach to overcome this difficulty would be profile-based recommendations in which domain-independent features, such as gender or occupation, are used to obtain user representations shared across domains. While this approach allows the application of single-domain methods, such features are often hard to obtain as users often hesitate to reveal their profiles due to privacy concerns. Therefore, without such features, we need to represent users by their consumption patterns, with which we can utilize the labeled data in a source domain to learn the relation between users and items.

Previous studies dealt with the above challenge by using specific forms of auxiliary information such as search queries (Elkahky et al., 2015), social tags (Abel et al., 2011, 2013), and external knowledge repositories like Wikipedia (Loizou, 2009; Fernández-Tobías et al., 2011; Kaminskas et al., 2013, 2014), which may not be available in some applications. For instance, the information about a TV program that is only available in a particular service may not be available on Wikipedia because few people have written about it. On the other hand, collaborative filtering (CF) approaches were also proposed (Zhang et al., 2010; Iwata and Takeuchi, 2015). While CF methods do not require auxiliary information, they suffer from data sparsity and require a substantial amount of previous user interactions to give a high-quality recommendation.

In contrast, this paper investigates content-based approaches. In particular, we study a deep-learning apparoach to cross-domain recommendation. Deep learning has been successfully applied to recommender systems(Oord et al., 2013; Wang et al., 2015; Elkahky et al., 2015; Covington et al., 2016; Wu et al., 2016; Zheng et al., 2016; Wang et al., 2016; Wu et al., 2017; He et al., 2017)

. In addition to the success in the applications to recommender system, deep learning methods show better performance in the area of transfer learning because of its ability to learn transferable features from the data. In fact, in the area of domain adaptation, deep neural networks show state-of-the-art performance in computer vision and natural language processing tasks

(Ganin et al., 2015; Bousmalis et al., 2016). Based on this observation, we hypothesized that domain adaptation with deep neural networks could also be applied to cross-domain recommendation.

Domain adaptation (Jiang, 2008) is a technique to learn a new domain (the target domain) with a few or no labeled training data using the knowledge acquired from another domain that has labeled data (the source domain). With domain adaptation, a classifier trained in the source domain can be applied to the target domain. Our approach is to treat recommendation as extreme multi-class classification (Partalas et al., 2015; Covington et al., 2016)

where labels (items) corresponding to a user history are predicted. A benefit of this approach is that the problem of cross-domain recommendation can be seen as an instance of domain adaptation, which has been extensively studied in the machine learning literature. To achieve domain adaption, we use a newly proposed neural network architecture for unsupervised domain adaptation, domain separation network (DSN) proposed by

Bousmalis et al. (2016). In addition, to reduce the difficulty of extreme classification and deal with new items, we combine item features via a Stacked Denoising Auto Encoder (SDAE) (Vincent et al., 2010). To validate our approach, we performed experiments on real-world datasets collected from the movie and news services of Yahoo! JAPAN. In summary, our contributions in this paper are as follows:

  1. We proposed a deep domain adaptation approach for cross-domain recommendation.

  2. We examined the capacity of our method in experiments on large-scale real-world datasets and confirm the effectiveness of domain adaptation.

  3. We also showed that our approach has better performance compared to a state-of-the-art cross-domain recommendation method.

The rest of this paper is organized as follows. In Section 2, we give a summary of related works. In Section 3, we first formulate the problem of cross-domain recommendation and introduce the machine learning methods that we use for out model. In Section 4, we elaborate our proposed model. In Section 5, we describe the datasets we used for the study and evaluate our method. Finally, in Section 6, we conclude with a discussion on the future work.

2. Related Work

2.1. Cross-Domain Recommendations

The focus of our work is on addressing the user cold-start problem by leveraging user activities in other domains. One of the major challenges in cross-domain recommendation is the lack of labelled data, which is essential to establish the link between domains. Approaches to this problem can be broadly classified into two types: collaborative filtering and content-based filtering.

Collaborative filtering suggests items based on the rating patterns of users by taking into account the ones of other users with similar tastes. As the only required data is rating patterns, collaborative filtering is a generally applicable approach. By assuming the existence of shared items or users, single-domain methods, such as matrix factorization (Salakhutdinov and Mnih, 2008b, a), can be applied. However, in the absence of user/item overlaps, as no users in the target domain have rated items in the source domain, collaborative filtering does not work (Cremonesi et al., 2011). Two approaches have been proposed to overcome this limitation (Zhang et al., 2010; Iwata and Takeuchi, 2015). In particular, Iwata and Takeuchi (2015) propose a method which is based on Bayesian probabilistic matrix factorization(Salakhutdinov and Mnih, 2008a)

. Their method assumes latent vectors are generated from a common Gaussian distribution with a full covariance matrix, which makes latent factors aligned across domains and enables predictions in the target domain.

In contrast, our approach is a content-based approach. Content-based filtering recommends items by comparing the content of the items with a user profile in accordance with what they consumed before. Unlike collaborative filtering, a domain link can be established without an item/user overlap by using the auxiliary information about users or items. Abel et al. (2011, 2013) presented an approach that constructs common user representations from social tags, such as on Flickr. Fernández-Tobías et al. (2011) and Kaminskas et al. (2013) proposed a method for recommending music to place of interests. Their method is a know-ledge based framework built on the DBpedia ontology, which is used to measure the relatedness between the items in different domains. In their method, items need to be defined on the knowledge repository. Related to these studies, Loizou (2009)

proposed a method that uses Wikipedia as an universal vocabulary to describe items. Their method is to build a graph whose nodes correspond to Wikipedia articles describing items rated by the users, and edges express the relationships between those articles. A Markov chain model is defined using the graph and produces recommendation by assessing the probability of traversing from the nodes expressing a user to particular an item. When articles that directly correspond to items do not exist, they proposed using tags or words associated with items to search for related Wikipedia articles. This requires extra preprocessing of the data, and descriptive articles may not exist. On the otherhand, our method has broader applicability as it only assumes the content information of items.

2.2. Deep Learning and Recommender Systems

Deep learning is a powerful techqnique to learn predictive features from data. The representation of the user or item is essential in recommendation tasks as the performance of recommendation is dependent on the ability to describe user preference and complex user-item relations. For this reason, the application of deep learning to recommender systems has gained research attention. Wang et al. (2015) and Li et al. (2015) combined autoencoder models with collaborative filtering; Wang et al. (2016) used a RNN model for item representation and combined it with CF; Wu et al. (2017) modeled temporal dynamics between users and items wirh RNNs.

Deep neural networks are also used for content-based filtering: music recommendation (Oord et al., 2013), news recommendation (Oh et al., 2014), and movie recommendation (Covington et al., 2016). Among the previous works, the most relevant to our setting is Elkahky et al. (2015), which dealt with user modeling for cross-domain recommendations using deep neural networks. They proposed the multi-view deep structured semantic model (MV-DSSM), which maps users and items into a latent space where the similarity between them is maximized. Their model requires the user’s search queries or user overlaps to obtain user representations shared across domains. Our model, in contrast, does not require those types of data as our model obtain shared user representations via unsupervised domain adaptation. Our model can be seen as an extension of Covington et al. (2016) and Wang et al. (2015). Covington et al. (2016) proposed a collaborative filtering approach via extreme classification in which a classifier is trained to generate a candidate set. Inspired by their work, we approach to cross-domain recommendation with a content-based approach and domain adaptation so that different domains can be linked with semantic information. Also, as in Wang et al. (2015), to perform predictions on items that are not frequently observed, we combine item information with the model using a denoising autoencoder.

3. Preliminaries

3.1. Problem setting

In this paper, we assume that the input data takes the form of implicit feedback such as click logs. We also assume access to the content information of items. Now we formally define our problem setting. Let be a collection of items in a source domain that we wish to recommend. From user logs on these items, we have a labeled data set . Here denotes a label that represents an item consumed by a user, and is raw features representing a list of items used by the user before the item . In a target domain, we only have an unlabeled data set due to the lack of common users between the two domains. As with the source domain, denotes a use history composed of items in the target domain. Our objective is to recommend items in the source domain to users in the target domain. More formally, by exploiting the labeled data and the unlabeled data , we aim to construct a classifier

, which defines a probability distribution over items

given a user history in the target domain. We model this classifier with domain adaptation, which we explain in Section 3.2.

As an illustration, suppose that our task is to recommend videos to users who use a news browsing service. We have user logs of the video service in the source domain and of the news service in the target domain. A video can have text information about itself, such as descriptions of cast or plot. With this content information, we can construct raw user representations from a list of movies using the Bag of Words or the TF-IDF scheme. As with videos, we can construct representations for news users using textual information of news articles. Using content information, we obtain a labeled dataset for the source domain (the video service) and an unlabeled for the target domain (the news service). In this case, the task is to build a classifier that takes as input a news browsing history and assign probability to videos.

3.2. Domain Adaptation

The labled data in the source domain would be useful in mining the relations between items and users; therefore, we want to utilize this data to learn a classifier for the target domain. Domain adaptation serves this purpose by adapting a classifier trained on the labeled dataset in the source domain to the target domain. In the following, we formally define domain adaptation and introduce the architecture for domain adaptation.

3.2.1. Definition of Unsupervised Domain Adaptation

Following (Ganin et al., 2015; Bousmalis et al., 2016), we define the problem of unsupervised domain adaptation. We consider a classification task where is the input space and is the label set. Let and be probability distributions over , which are assumed to be similar but different. We assume that we have an i.i.d sample of labeled data drawn from the source domain. On the other hand, we assume that little or no labeled data is available in the target domain. Instead, we have a large size of unlabeled i.i.d. sample , where is the marginal distribution over in the domain . The goal of unsupervised domain adaptation is to obtain a classifier with a low target risk defined by

using the labeled data and the unlabeled data .

3.2.2. Domain Separation Networks

In this section, we introduce the neural network architecture we use in our method. We use domain separation networks (DSNs) (Bousmalis et al., 2016) to achieve domain adaptation.

Figure 1. Architecture of a domain separation network (Bousmalis et al., 2016)

As illustrated in Figure 1, a DSN model consists of the following components: a shared encoder , private encoders , , a shared decoder , and a classifier . Here , , , , and denote parameters for the shared encoder, the private encoder for the source and the target domain, the shared decoder, and the classifier, respectively.

Given an input vector

, which comes from the source domain or the target domain, a shared encoder function maps it to a hidden representation

, which represents features shared across domains. For the source (or target) domain, a DSN has a private encoder () that maps an input vector to a hidden representation (), which serve as features that are specific to each domain. A decoder reconstructs an input vector of the source (or target) domain from the sum of shared and private hidden representations , (). We denote the reconstructed vector by . The classifier takes a shared hidden representation as input and predicts the corresponding labels.

The training of this architecture is performed by minimizing the following objective function with respect to parameters , , , , :

(1)

where , , and are parameters that control the effect of the loss terms.

Here, is a classification loss by which we train the model to predict the output labels. In the following, we assume that the classification loss is the cross-entropy loss:

where

is the one-hot encoding of the label of the

th example, and is the prediction on the -th input vector.

is the reconstruction error defined as:

encourages the shared and private encoders to extract different type of features from the inputs. As in (Bousmalis et al., 2016), we impose a soft subspace orthogonality condition:

where is the Frobenius norm. Here is a matrix whose -th row is the shared representation of the -th input vector in a source sample , and is similarly defined for private hidden representations. Likewise, and are defined for the target domain.

Finally, the similarity loss encourages the shared encoder to produce representations that are hardly distinguishable. For this, we use the domain adversarial similarity loss (Ganin and Lempitsky, 2015; Ganin et al., 2015). Let be a binary label that represents the domain of the -th sample ( denotes the source domain). To use the adversarial loss, we equip the DSN with a binary classifier parametrized by , which predicts the domain labels of shared representations, and define the similarity loss as:

(2)

where is the predicted domain label of the -th example. During training, the binary classifier and the shared encoder are trained adversarially; the loss is minimized with respect to and maximized with respect to a parameter for the shared encoder . The adversarial training can simply be performed by inserting the Gradient Reversal Layer (GRL) between the shared encoder and the domain label predictor. The GRL is a layer that acts as an identity function during the forward propagation, and during the back propagation, it flips the sign of propagated gradients. With GRL, we only have to minimize Eq. (2).

3.3. Stacked Denoising Autoencoder

We use a stacked denoising autoencoder (SDAE). It is a feed-forward neural network that learns a robust representation of the input data. Let

be a set of input vectors and be a set of vectors each of which is corrupted with noise. The SDAE model takes a corrupted vector as input and then transform it to with its encoder . The original input vector is predicted by reconstructing the hidden representation with its decoder . The network is trained to minimize the reconstruction error with respect to parameters , :

where denotes the prediction for the th corrputed vector.

4. Proposed Method

Figure 2. Proposed architecture

We first belabor our problem setting. To properly suggest items that a user prefer, we learn how likely a user is to choose a certain item from available data. Our objective is to learn a classifier that takes a user in the target domain as input and predict the probability for that the user would select the item .

We model this classifier with a DSN model. The detail of our model is illustrated in Figure 2. The model is a feed-forward neural network that has a DSN to obtain an adapted classifier

. For the output layer, we use the softmax layer. That is, the

-th element of the output layer is

where are softmax weights, and is the activations of the previous layer. Note that the softmax weight can be seen as a variable that represents the item , and the input vector fed into the softmax layer is the representation of a user. Given an input vector , we assign a class by

As the number of items is typically large, predictions only with user features would be a difficult task. In practice, the input data is imbalanced; some labels could be observed just a few times. Also, the types of items preferred by users in the target domain might be different from the ones in the training data. Therefore, the classifier should be able to predict new items not observed in the training data. From these observations, we believe that combining item features would help predictions. In fact, in the video service we used for our experiments, videos that the majority of users watches are short movies of the length of several tens of seconds. Therefore, combining video play times enable the model to identifies which videos are short ones. For this purpose, in addition to using a DSN, we also propose combining a denoising autoencoder to incorporate item features. We impose the mean squared error between the softmax weight and the hidden representation of items calculated by the denoising autoencoder.

In summary, we train the model so that it minimizes the following loss function,

Here, is a loss function defined by Eq. (1). is the squared loss between the softmax weights and item representations:

is the reconstruction loss of the item autoencoder. and are weights that control the interactions of the respective loss terms with the other losses.

In practical applications, the number of labels typically exceeds thousands, which means an exhaustive evaluation of the softmax function is computationally expensive. To make the computation faster, we use candidate sampling (Jean et al., 2015). Also, to generate a list of items to recommend at serving time, as mentioned by Covington et al. (2016), we only have to perform the neighbourhood search in the dot product space of the learned item features and user features (Liu et al., 2004), and thus the computation of the softmax function can be avoided.

5. Experiments

We examine the performance of our proposed method in experiments on a pair of real-world datasets.

5.1. Dataset Description

The two datasets consist of the user logs of the following services: (a) a video on demand service (VIDEO) and (b) a news aggregator (NEWS). The logs used were collected for three weeks in February in 2017. Both datasets contain descriptive attributes of products as follows:

VIDEO DATASET:

The implicit feedback of user watches are used. An access to the video page is regarded as a positive example. The dataset contains 48,152 videos of various categories (e.g. film, music). Each video has its title, category, short description, and cast information. To represent a user and an item, we use these features. In addition, we also use the video play time for item representation.

NEWS DATASET:

The dataset is comprised of news browsing histories. We treat a user who read an article as a positive example. Each news article has its own title and category, and we use them for feature representation. To make domain adaptation easier, we only use articles of entertainment-related categories, such as music, movie, and celebrity news.

5.2. Experimental Settings

We use the first two-week period of the datasets used for training and the rest for evaluation. For the source domain (VIDEO), we create a training set of 11,995,769 labeled examples where a label is a video watched by a particular user, and the input is a list of previously viewed videos. Similarly, for the target domain(NEWS), we build a training set of 10,500,000 unlabeled examples that consists of the view histories of news articles. For evaluation, we form a test data of size using the logs of users who used both services, which is comprised of pairs of a video label and a news history.

We preprocess the text information of user histories with the TF-IDF scheme. We treat a user history as a document and obtain a vocabulary of 50,000 discriminative words, which are chosen according to the tf-idf value. In a similar manner, we convert the item’s text information. An item (video) is treated as a document, and the vocabulary is capped at 20,000 words. A video category is transformed to a one-hot vector. To extract the features of a video playtime, we convert its hour, minute, and second to one-hot vectors. Concatenating these features, we obtain features for items.

We use the 80 percent of the training data for training and the rest as a validation set. Similarly, we sample the 80 percent of the test data. We repeat the experiment 5 times by randomly choosing a training-validation and test pair.

5.2.1. Evaluation Metric for Generated Candidate Lists

Given a user history, we generate a list of items sorted according to the relevance to the user. We aim to evaluate the goodness of the recommendation. To do so, we use the following two metrics:

  • Recall@

    denotes a function which returns if the statement is true and otherwise. The index runs over the test set, and is the size of the test set, and is the label of the -th example.

  • nDCG@

Unlike rating predictions, in implicit feedback, unobserved objects (zero ratings) are uncertain as to whether users have not rated or do not like them. Given that, as we only know that positive ratings are true positive, we use recall and nDCG as the evaluation metrics. With these metrics, we compare our methods with baseline algorithms.

5.2.2. Model Description

First, we describe the setting for our method. We construct a domain separatin network consisted of fully-connected layers with the hidden units:

  • the shared/private encoders: 111The left side is the input layer and the right side is the output ,

  • the shared decoder: ,

  • the classifier: ,

  • the binary classifier: .

In addition, for item representation, we use a stacked denoising autoencoder with the architecture (20000-256-64-256-20000). We denote a DSN with a SDAE by I-DSN.

The exponential linear unit (ELU) (Clevert et al., 2015)

is used as the activation function for all layers. The weight and bias parameters for the fully-connected layers are initialized following

(He et al., 2015). To achieve better generalization performance, we apply dropout (Srivastava et al., 2014) to all fully-connected layers. For DSN, the dropout rate is set to for all the encoders and for the decoder and the classifier. For SDAE, the dropout rate is set to for the input layer and for the rest of layers. We also impose the weight decay penalty on the weight parameters with the regularization parameter chosen from . The parameters , , and in Eq.(1) are set to , , , respectively. For I-DSN, the and are fixed at and . We use the ADAM optimizer for optimization (Kingma and Ba, 2014). The initial learning rate is fixed at . Although we aim to perform unsupervised domain adaptaiton, we use a validation set that consists of logs of common users on the same dates as the training data for model selection to investitigate attainable performance of deep domain adaptation.

5.2.3. Baseline Description

In our experiments, we compare our method with the following baselines.

Most Popular Items (POP):

This suggests most popular items in the training data. The comparison of our method with this shows how well our method achieves personalization.

Cross-domain Matrix Factorization (CdMF):

(Iwata and Takeuchi, 2015)
This is a state-of-the-art collaborative filtering method for cross-domain recommendation. By scraping the same user logs we used for our approach, we obtain a training dataset. We eliminte users with history logs fewer than for the video data and for the news data. We construct a user-item matrix for each domain, with the value of observed entries and of unobserved entries . The unobserved entries are randomly sampled. The statistics of the processed data set is given in Table 1. As with our method, we use the percent of the instances for training and the rest for validation.

We initialize the latent vectors with the solution of maximum a posteriori (MAP) estimation as suggested in the paper

(Iwata and Takeuchi, 2015). The dimensionality of latent vectors is set to . Although is small, as the training involves the computation of the inverse matrices of this dimensionality for MCMC sampling, it takes very long time even for the dimensionality . Also, to make the computation faster, we truncate the iteration at the sample and discard the first samples as burn-in. The hyper parameters , , and are set at the same values as in (Iwata and Takeuchi, 2015). Because of the time computational complexity, it is hard to optimize these parameters. Therefore we make these parameters fixed.

Neural Network (NN):

To investigate the performance of domain adaptation, we also test the neural network without domain adaptation. We use the same architecture as DSN except for the domain adaptation component by setting , in Eq.(1) to . Parameters for weight decay are chosen from the same parameter grid as our method.

News Movie
Instance
User
Item
Table 1. The data statitics for CdMF

5.3. Results of Experiments

Neural network models are selected according to the valud of the softmax cross entropy loss. Method nDCG@ nDCG@ nDCG@ nDCG I-DSN DSN NN CdMF POP

(a) NDCG
(b) Recall
Method Recall@ Recall@ Recall@ Recall@
I-DSN
DSN
NN
CdMF
POP
Table 2. Performance comparison of I-DSN, DSN, NN CdMF, POP

Neural network models are selected according to the value of nDCG@ Method nDCG@ nDCG@ nDCG@ nDCG I-DSN DSN NN POP

Table 3. Performance comparison of I-DSN, DSN, NN, POP.

In Table 2, we show the performance comparison in terms of the two evaluation metrics. For I-DSN, at test time, for unobserved items, we can initialize the softmax weights with the hidden representation of the items returned by the autoencoder. Although we tried this approach but it did not yield better results compared to when the weights were not initialized. Therefore, we report the results of I-DSN with the weights of unobserved items not initialized with the hidden representations.

In both results, CdMF underperforms the other methods. As mentioned in (Hu et al., 2008), it can be considered that CdMF cannot process the implicit feedback and accurately express the popularity structure in the dataset.

We can see that the domain adaptation approaches outperforms NN in terms of both metrics. In terms of recall, I-DSN shows the better average performance compared to POP when the length of suggested lists is and , while DSN cannot beat the results of POP. However, in regards to nDCG, both domain adaptation methods underperform POP. This results suggests that optimizing the model in terms of the softmax cross entropy loss could produce higher recall outputs but not ones of good ranking quality.

Based on the above results, we aim to produce a model that has good ranking performance. To do so, we calculate nDCG@ on the validation set and choose the model by changing the weight decay parameters and the iteration step. The result is shown in Table 2. While all three methods show improvement in their result, DSN shows the best average performance compared to the other methods. Thus, by aligning the objective loss with the evaluation metric, DSN can improve its recommendation performance.

In summary, it has been shown that the performance of a simple neural network was improved by using domain adaptation even in cross-domain recommendation. In addition, our approaches outperformed a state-of-the-art collaborative filtering method. Although the domain adaptation approach overcame other machine learning methods, however, deep domain adaptation sometimes underperformed a simple popularity-based method. Specifically, our method underperformed the popularity-based method particularly when the length of suggested lists is short. This indicates that for services that can suggest a small number of items, only using our approach may not be effective. However, our model can produce a better candidate set with higher recall when the length of suggested items is longer. Also, predicting items from it using other (heuristic) ranking procedures, e.g. combining the result of popularity-based ranking, could improve the result. Furthermore, the application of domain adaptation showed actual improvement when the loss function is aligned with the evaluation metric. Therefore, we conclude that we obtained positive results from these experiments.

6. Conclusions and Discussions

In this paper, we studied a problem of cross-domain recommendation in which we cannot expect common users or items and presented a new deep learning approach using domain adaptation. We investigated the performance of our approach in cross domain recommendations by conducting experiments on a pair of two real-world datasets. We confirmed that our approach actually improved state-of-the-art in the literature. Although the improvement was not overwhelming, we showed that by treating the problem as extreme classification, domain adaptation was applicable and obtained promising results that encourages the further extension of our model. As this is the first attempt in the literature, based on these results, we would like to keep improving in future work.

Specifically, We envision that further improvement can be achieved in the following three ways. First, item feature extraction. Combining item representation with a stacked denoising autoencoder showed the improvement in recall. However, initializing softmax weights with item features did not produce a better result. This implies that the extracted features from autoencoder only contribute to predicting labels observed in the training data. We hypothesize that this could be because only using the text information and the length of videos does not make each video distinct from others. To achieve better generalization performance, we should include other types of information about videos. For example, if a certain program is episodic, then implementing such information may help the prediction because we can deduplicate the same type of videos from the suggested list by treating the videos as one program.

Second, as with item feature extraction, we need improve the training of the classifier so that it can work for classes with few examples. In terms of recall, there was not overwhelming difference between the popularity method and our approach. This indicates that to have better performance, our classifier should be able to better predict rare (not just popular) items. Although dealing with class imbalance is a persisting issue, this could be realized, for example, by combining techniques of few-shot learning (Vinyals et al., 2016; Kaiser et al., 2017; Finn et al., 2017).

Third, as optimizing the model with nDCG saw a large improvement, we also suggest replacing the softmax loss with another ranking loss such as nDCG (Valizadegan et al., 2009).

References