Alibaba proposed a dynamic ranking model based on heterogeneous content streams to improve users’ click and purchase rates

An End-to-end Model of Dynamic Ranking On Heterogeneous Content Streams

Gao Zi-zhe, Gao Zheng, Huang Heng, Jiang Zhuo-ren, Yan Yu-liang/Zizhe Gao,Zheng Gao,Heng Huang,Zhuoren Jiang,Yuliang Yan

Research Background:

Search engine plays an extremely important role in the field of e-commerce, which can well guide users’ potential purchase behavior. Traditional e-commerce search engines usually refer to commodity search engines. Users type a Query and return a list of commodities. However, with the development of “we media”, more and more users are more willing to share their shopping experience and show their views in the form of articles, comments and videos. In this article, these are collectively referred to as content streams. In order to provide users with more shopping help, content search engine came into being. Recommend high-quality content streams to users when they search for products, and help users choose products they like and which they might like.

At present, there are many challenges in sorting heterogeneous data. First of all, the cross-domain knowledge provided by commodity search engines and content search engines should be fully utilized so that users’ behavioral preferences in commodity search engines can be applied to content search engines. Second, existing algorithms need to support sorting of multimedia content.

In this paper, our goal is to solve the problem of heterogeneous data sorting in product search engines and content search engines, and to recommend rich personalized content streams to users. We divide the algorithm into two parts: 1) Sorting of heterogeneous content stream types, that is, determining what type of content stream is displayed in each pit, such as articles, videos or commodity lists; 2) Isomorphic content stream content ordering. The second step uses the well-known DSSM model. Under this content stream type, the content of the content stream is sorted and the content with the highest similarity is selected for insertion. This article focuses on the first step.

The proposed algorithm:

This paper proposes two algorithms for sorting content stream types, independent multi-arm slot machine algorithm and personalized Markov deep neural network algorithm.

In the independent multi-arm slot machine algorithm, we need to calculate a ratio \(\theta \), calculated by IPV and PV. If \(\theta \) is higher, users are more likely to click when they see the content stream in the search list. For each search pit, we first calculate a prior distribution of \(\theta \), Here we use Beta distribution \(\theta _i\sim B(\alpha _i,\ Beta _i)\, where I stands for post, list, video. \(\alpha _i\) stands for historical IPV click data of type I, \(\beta _i\) indicates historical browsing data. The expectation of \(B(\alpha _i,\beta _i)\ is \(\frac{\alpha _i}{\alpha _i+\beta _i}\), and the posteriempirical probability distribution is updated by a real-time streaming data task. Is expressed as the following probability formula:

In this way, the content flow type for all pits is independent, with the pseudo-code as follows:

The choice of dependent heterogeneous data flow type is determined by three factors: the user, the Query, and the previous pit type. First, users can express different preferences within the same Query. For example, if a user searches for “dress,” one user may prefer an introductory article, while another may prefer a video introduction. Moreover, no user likes a single type of presentation, and more or less all like a diverse array of content stream types. Different users should be shown different sorting results for the same Query. The proposed personalized Markov deep neural network algorithm consists of two steps, including the representation task learning of user and query and the prediction learning of pit type.

Low-dimensional representations of users and Queries We build a graph that contains users, Queries, and content. Node2vec is used to learn the embedding of user and query.

The middle part of the figure is the embedding representation of the training node. The input layer is the unique thermal coding of nodes. The weight matrix W is the embedding of all nodes, which maps the nodes into a D-dimensional space.

Pit type prediction Our objective function is defined as

Where X represents the input feature of the ith pit position. In order to simplify our pMDNN model and speed up the operation, we only use the information of the pit position before the current predicted pit position. However, this raises the question of how to predict the type of the first pit. Here, we use cross-domain knowledge. We extract the information from the user’s most recent browsing in the product search engine and map it to the content search pit feature to meet the input requirements of the current model. Our model input layer is the user’s embedding, Query embedding and the previous pit embedding. Can be expressed as

Three full connection layers access the input layer. Each layer uses a linear classifier and cross entropy as the Loss function. Relu was selected as the activation function, and Softmax was applied as the activation function in the output layer.

Experimental results:

We deployed the proposed model in A/B test bucket environment and selected 5 main indicators to compare the two models iMAB and PMDnn. pv represented the number of displayed contents; Pv click indicates how much of the content displayed is clicked; Uv is how many users use the content search engine, uv click is how many users click on the content stream; In the case of UV CTR, the ratio of whether the user clicks on the content stream.

The following table shows the experimental results, among which the experimental results of pMDNN are better than that of iMAB. Uv Click and UV CTR, in particular, are very important for our scenario, because the growth of UV Click means that more users are turning to content search engines because it helps them shop better. At the same time, the growth of UV CTR shows that users using content search engines are receptive to our recommended content streams. As for the improvement of PV Click, it also represents that the model we proposed is more in line with the personalized needs of users.

Based on PV Click and UV CTR, we can conclude that pMDNN applies cross-domain knowledge and that globally optimized multi-pit type is indeed superior to pit-independent iMAB.

References:

[1] Robert Busa-Fekete and Eyke Hullermeier. A Survey of Preference -based Online learning with Bandit Algorithms. [2] Ali Mamdouh Elkahky, Yang Song, and Xiaodong He. 2015. A multi-view deep learning approach for cross domain user modeling in recommendation systems. In Proceedings of the 24th International Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 278-288.