Topic Title: Towards a Unified Adaptable Framework of Deep Transfer Learning for Different Commodities Evaluation Network Models of Differentiated Domains

 

Technical Area: Deep Transfer Learning

 

Background

Transfer learning, which transfers the knowledge learned from one scene to another scenario through algorithmic modeling, has always been a hot research field in artificial intelligence. Most machine learning algorithms are based on the same assumption that there is the same feature space and probability distribution between training data and required prediction data. But in practical applications, this assumption is sometimes untenable. The long tail effect is commonly found in our data. Usually, we can use traditional machine learning methods to solve problems of domains in which data can be easily obtained. Most of the time we are interested in some new domains or tasks, but it's hard to get enough data to build models with good generalization capabilities for these tasks. The purpose of transfer learning is to transfer knowledge from existing tasks to current tasks. For the task of introducing Taobao’s commodities to AliExpress scenarios, the Taobao’s commodities are rich in the aspect of content and user behavior data, while the labeled data for these commodities in new scenarios is sparse and not enough to build models from zero. Therefore, in order to quickly fuse the existing models and data in the AliExpress scenario, the transfer learning technology is very suitable for transferring commodity knowledge across different commodity domains.

 

We invite researchers who are either experts or are keenly aware of the challenges and opportunities that their fields bring to transfer learning to work on the new unified adaptable framework of transfer learning that seamlessly integrates different commodity scoring evaluation networks of different countries, allowing us to stay focused on these real-world problems and use cases that transfer learning framework as whole can and should help solve with global optimum settings.

 

Target

However, we notice that most of transfer learning methods focuses on image related tasks, and there is no unified adaptable framework for transfer learning among many similar domains. Thus the whole framework has been far from unified adaptable settings, as many facts should be traded off, such as different number of available features, outputs, and knowledge representations.

 

As a result, we propose to investigate a new unified adaptable framework of transfer learning that can be easily adapted to many similar commodity domain models with differentiated countries.

 

On the technology side, we will take construction and landing of national differentiated traffic guidance algorithm system, including new market cold start, model transfer among different markets, adaptive localization and so on. In business, the system can quickly enter the new market, coordinate operation of long tail market, refine operation key market, and can be self-learning based on feedbacks of user behaviors, and bring more convenient and more intelligent shopping experience for the global consumers.

 

Related Research Topics

There are 4 different methods of transfer learning, which are instance based transfer learning, feature based transfer learning, parameter based transfer learning and relationship based transfer learning. Instance based transfer learning assigns different weights to the samples in the source domain, so as to select the samples that are more useful to the target domain to train the target task. Among the algorithms of instance based transfer learning, the TrAdaBoost algorithm proposed by Wenyuan Dai is more famous, which assumes that the data of the source and target domains have exactly the same characteristics and category labels, but the distribution of the two is different. The parameter based transfer learning method assumes that different models of related tasks share certain parameters, or prior distribution of hyperparamters. Most of the related algorithms use the multi-task learning approach, but transfer learning emphasizes better results on the target task, while multi-task learning emphasizes both the source and target tasks. In multi-task learning, the weights of the source and target domains in the loss function are the same, while most of parameter-based transfer learning algorithms assign a greater weight to the part of the target domain in the loss function so that the target task has a better performance. In addition to the transfer learning of specific models, there are also some transfer learning methods based on ensemble learning, which can use different basic learners. The purpose of feature based transfer learning method is to learn a "good" feature representation for the target task, and the knowledge transferred between the source domain and the target domain is encoded into the learned feature representations. Relationship based transfer learning solves the knowledge transfer between related domains, but requires data to be non-independent and identically distributed.

 

To sum up, the parameter based transfer learning method, which can model the common features of different commodity domains by sharing parameters, can be modeled by utilizing multi-task learning method and deep models widely used by the industry and academia. Therefore, it is very suitable for the applications of AliExpress scenarios of differentiated countries that introduce Taobao’s commodities.

 

In addition, dual learning has been successful in recent years in Neural Machine Translation, image processing and sentiment analysis. The learning framework is modeled by respectively mapping from the feature space of source domain to the feature space of target domain and vice versa, such that the two mapping models are made to approximate the joint probability distribution of source feature space and the target feature space respectively. Cross domain search and cross domain recommendation need to solve the problem of searching and recommending from the user behavior-rich source domain system to the sparse target domain system.

 

Because the user and commodity between the two domains do not overlap, the traditional way of solving the cold start problem by matrix decomposition and matrix filling is not applicable. In order to establish the ranking models from the source domain to the target domain and from the target domain to the source domain, a semi supervised model of cross domain learning can be built by utilizing the dual learning framework, using the rich information of interactions between users and commodities of their respective domains. Secondly, the generative adversarial network (GAN) can be used to model the discriminative model and generative model from the commodity to the label, thus providing more samples for the training of the cross-domain model. In fact, Dual GAN is a generative adversarial network that combines dual learning, and has been successfully applied in the "translation" task between images and tags. Finally, the dual learning framework can be built by separating commonness and difference of the cross-domain model, thus laying the foundation for the large application of the scoring evaluation network in the national differentiation models. In conclusion, dual learning framework can provide a uniform test platform for cross-domain learning model and scoring evaluation network, and facilitate the generalization of national differentiation models.