Investigate why you are looking at Taobao and want to chop your hands.


Editor’s note: This article is from the micro-channel public number “qubits” (ID: QbitAI), Author: dry out, 36 krypton release authorized.After opening Taobao, why can’t you stop trying to chop your hands?Alibaba also wants to find the answer.One of these papers has also appeared in the top AI conference NeurIPS.During NeurIPS 2019 in Vancouver, one of the authors of the paper, Yang Hongxia, a senior algorithm expert at Alibaba Dharma Institute, also thoroughly explained it.Why do you want to chop your hands, can AI understand?This paper from Ali is called Learning Disentangled Representations for Recommendation.His research focuses on the cognitive factors of people in the process of matching people and products.That is: why does a person like a product, what concepts does he care about and plant grass, for what reasons did he click, collect, and place an order, and what cognitive level is his current focus on?Can the recommendation system know, digest and respond accurately?Scientists at Dharma Institute believe that these so-called cognitive factors are not the fine-grained attributes and categories inherent in commodities, but rather a spreadable and interpretable concept for understanding commodities from a human perspective.They are more like memory points that advertisers choose to impress.One difference between a recommendation system and a search scenario is whether it can actively motivate users’ potential interests and help users find and accept unexpected products.Therefore, how to mine potential cognitive concepts and pass the potentially acceptable cognitive concepts to users in a reasonable way may be something that the recommendation system needs to break through.Of course, paying attention to this cognitive process is not for end-to-end prediction of “the next product”, or click-through rate estimation or rating estimation.At least the predecessors’ online experience with large-scale data can show that explainable recommendations with the same product form can’t really improve the final click and conversion effect compared to the black box model.Therefore, compared with interpretable recommendation, cognitive recommendation emphasizes more on the human factor, and its end result must be technology-driven innovation in product form.The new product form can create new needs, user habits and new business scenarios.How to use AI to make you chop your hands. Based on this background, Alibaba decided to focus on two cognitive-related subtasks: First, how are products represented in human cognitive space, and whether such representations are interpretableFor example, if a corresponding dimension can be found, it can represent an independent “semantics”.The interpretability of semantics here is actually a concept related to cognition and communication, that is, it can be understood and transmitted by people.Similarly, does the representation of people in this space also have such semantics?In connection with the development of Disentangled Representation Learning on continuous data, Ali wants to explore whether similar results can be learned from discrete data, especially user behavior data.Secondly, based on such a characterization, can a new type of recommended application be proposed and at least one prototype solution be given.One of the questions explored here is whether, based on the user’s behavior, can we get some cognitive-related decision-making factors and represent the goods and users in a dissociable manner.Their goal is to obtain a vectorized representation of user u. At the same time, the model will also produce a representation of the product for the recommendation system to recall a batch of products based on the user’s representation.According to the characteristics of user behavior on the e-commerce platform, their model adopts a hierarchical design: when inferring a user’s representation, macro disentanglement and micro disentanglement are performed in order.The main consideration of macro dissociation is that user interests are usually very broad, and a user’s click history often involves multiple independent consumption intents (such as clicking on products in different broad categories).And users’ preferences when performing different intentions are often independent. For example, the preference for dark clothes does not mean that the users also like dark appliances.Even price preferences are often non-migratable, such as buying high-end lipsticks and cheap laptops that are not mutually exclusive.In addition, macro-dissociation is also a necessary prerequisite for micro-dissociation.Micro dissociation is to further decompose the user’s preferences when performing an intent to a finer granularity.However, the different types of product attribute sets are very different. If a certain dimension of the user representation vector has been used to characterize the user’s preference for mobile phone power, then this dimension has no meaning for products such as clothing.Therefore, when predicting whether a user will click on a certain clothing, or when learning the representation of a certain clothing through user behavior, these dimensions related to mobile phones should be ignored.Based on this idea, they proposed a model: this is a deep generation model assuming that the user’s representation indicates which macro consumption intentions these products usually correspond to.To optimize this deep probability model, they also adopted the framework of VAE.The following pseudocode can help to better understand the optimization goal: Can this model make you more choppy?Scientists at Ari Dharma Institute said that the dissociation characterization not only brought certain interpretability, but also brought certain controllability.This controllability is expected to introduce a completely new user experience to the recommendation system.For example, since each dimension of the representation is associated with different product attributes, the user’s representation vector can be provided to the user, allowing the user to fix most of the dimensions by himself (such as the style, price, size, etc. of the clothes), And then individually adjust the value of a certain dimension (such as the dimension corresponding to the color), and the system adjusts the recommendation result based on this feedback.This will help users express what they want more accurately and retrieve what they want.They also showed the two batches of products retrieved after adjusting a certain dimension. It can be seen that this adjusted dimension is related to the attribute of the color of the backpack and has a more obvious gradual property: The following is a search in another dimensionIn the two batches of goods obtained, it can be seen that this regulated dimension is relatively related to the attribute of the color of the backpack, and has a more obvious gradual nature: Of course, this study is still insufficient: not all dimensions can be understood by humansSemantics.In an unsupervised situation, it still takes luck to train an interpretable model, and the pitfall of “training multiple models repeatedly and then picking the best model” cannot be avoided.Therefore, it is suggested that future researchers pay more attention to the (weak / semi-) supervision methods and introduce label information.This is the result obtained on the online data, so how about the quantitative experiment on the offline data?They quantitatively measure the degree of ionization (and its relationship to recommended performance) on a small-scale data set.It was initially found that there is a strong correlation between the high degree of dissociation and the good performance of the recommendation. After the introduction of macro dissociation, it did greatly improve the micro dissociation. Their method isThe recommended performance is better than the baseline method.They also measure the top-N recommendation performance of the method on several offline datasets (including one Taobao dataset AliShop-7C).It can be seen that this method is superior to the baseline method, especially on small or sparse data sets.Because Top-N recommendation is not Ali’s original intention to do this, it is satisfactory to be able to make such an effect.What’s the use of doing this?In the thesis, scientists from Dharma Institute also explained that with the development of modern e-commerce recommendation systems, the academic and industrial circles are expected to increase the click rate and predict the next single task of clicking products.It is difficult, and the incremental benefits brought by such an improvement are also difficult to estimate.More user experience issues are put in front of decision makers, such as why they buy and push, why they are clicked products, and how to create real incremental value.Therefore, they now choose to explore the possibility of new recommendation forms around human cognitive behaviors and processes.Finally introduce some authors.One is Ma Jianxin, an intern at Ali Dharma Institute, a 13-year undergraduate student at Tsinghua University, and a professor from Tsinghua University.Another author, Zhou Chang, is also an algorithm expert from Dharma Academy.The paper portal is as follows, if you are interested, you can follow: Learning Disentangled Representations for Recommendation