Alibaba search recommendation system has been upgraded again?!

Time:2019-11-10

Alimei’s Guide:As the traffic entrance of search, search shopping guide products carry the important functions of recommendation for user shopping guide and flow diversion of search. Main products include: Home shading, pull-down recommendation, search discovery, navigation, historical search, etc. After several years of exploration and accumulation, each product is more and more mature. Machine learning algorithm is widely used in shopping guide products, and has achieved remarkable results. On the basis of supporting the hand search business, search shopping guide also actively expands the boundary and supports a large number of product lines within the group. Therefore, higher requirements are put forward for the search guide product line: not only the efficiency of its own products needs to be improved to better support the manual search business, but also a flexible framework to support more and wider businesses.

I. system framework

The optimization idea of shopping guide upgrading starts from three directions: 1. Strategy upgrading. By using the idea of deep learning and heterogeneous network, the user personalization is further understood and modeled; at the same time, the number of independent queries caused by Matthew effect is optimized 2. Guided purchase and outward investment. Search, guide and enable shopping in the channels including venue activation page and guess what you like, so as to open up search channels for users 3. Product innovation. On the one hand, innovate and upgrade existing products, such as activation page, pull-down recommendation, etc.; on the other hand, actively try new product forms, such as home page hot words, search dynamic cards, etc.

Search guide core solves the problem of keyword recommendation for consumers, so although there are many products and different forms, there are many commonalities in the underlying architecture, so we designed a general and flexible framework to support.

In the recall stage, we enrich the recall methods, and select different recall strategies according to different channels, scenarios and product forms to get candidate query terms candidates.

In the sorting stage, we not only introduce deep learning into the framework of shopping guide algorithm, but also innovatively join the idea of heterogeneous network, effectively integrate the sequence information of different paths of users with LSTM and other models, and have a deeper understanding of consumers.

In the business strategy stage, we use Jaccard coefficient, edit distance and so on to optimize the problem of semantic repetition. At the same time, we use E & E mechanism to upgrade the scenarios with more serious Matthew effect, and increase the efficiency rotation mechanism to further improve the efficiency.

Next, it introduces several specific products in detail.

II. Detailed scheme

2.1 shading recommendation and optimization

In the algorithm optimization of shading recommendation, we innovatively propose a recommendation method based on heterogeneous information network (HIN). The recommendation framework is shown as follows:

User, item and query are three basic types of nodes in manual search. There are different interaction relationships among these three types of nodes. For example, user clicks item directly, user enters search through query query, and clicks item in search.

However, most of the traditional recommendation methods only focus on Feature Engineering, ignoring the relationship between these different nodes. At the same time, the large-scale data volume (100 million query, billions of users and items) in the field of e-commerce also needs to be considered. So we design and propose a large-scale query recommendation method based on Metapath embedding representation, Metapath guided embedding for large-scale query recommendation (melqr). It uses heterogeneous network to model query recommendation, and uses Metapath to guide user and query representation learning by aggregating local neighborhood information. In addition, we use Metapath to guide user and query representation learning in heterogeneous network All nodes are represented by a fusion method of term embedding, which avoids large-scale parameter problems in network learning.

The model combines the strategies of expanding recall, dynamic display, etc. to increase 10% of the online shading use UV + and lead the transaction amount to increase 10%. It is worth mentioning that the model also uses other products such as search discovery, home page hot words and so on. The improvement of the effect is also very obvious.

2.2 home page hot words optimization

Hot words on the home page is an innovative product of this year’s search in the manual search home page, which can help users find products of interest through keywords and enhance users’ search mind.

The framework and algorithm framework of home page hot words and shading recommendation sharing system

2.3 pull down recommendation optimization

The optimization goal of the previous version of pull-down recommendation is to increase the proportion of pull-down guided PV in search PV, that is, pull-down utilization. The previous version tried to fit the user’s preference for the query shown in the drop-down. However, in the statistical features used, the features used are all pull-down guided data. This brings about a serious problem. In the current product form, only 10 candidate queries can be displayed for each user input. Therefore, a relatively large number of queries will have a relatively high statistical value at the beginning, and a higher statistical value will promote the query to a higher position in the sorting. As a result, there will be a cycle. Over time, under some specific queries, the statistical value characteristics of pull-down recommendation candidates will be very different. Thus, Matthew effect is formed. One of the most serious problems of Matthew effect is that it will cause the query in the dropdown display to over converge to a smaller collection, resulting in a decrease in the number of guided independent queries.

To solve these problems, we systematically reconstruct the pull-down recommendation model, with the framework as follows:

On the one hand, the core idea is to add features and samples of user’s active input, modify Matthew of dropdown itself, on the other hand, strengthen user’s personalized features, add user behavior sequence, etc. After the optimization model goes online, the use of PV for the drop-down itself is increased by 10% +.

2.4 dynamic card optimization

When the user’s search terms are relatively broad, they can’t better represent the user’s search intention. The user’s real-time click behavior on the search results page can reflect the user’s current intention in a more real-time manner. At this time, the user can be recommended search terms that meet their search intention, which can improve the user experience. By recommending relevant search terms to users, we can improve per capita query, then improve per capita PV, and improve the length of stay in search. Product examples are as follows:

Through continuous optimization, the CTR of dynamic card display has been higher than that of goods, content and other cards, indicating that users have strong willingness to click; at the same time, the per capita query of users is increased by 4% +, per capita PV is increased by 1% +, and the user experience is improved.

2.5 other work

In addition to the above work, we also support search discovery, venue search, recommendation vane, search activation page revision and other project optimization.

III. double 11 effect

The shopping guide product line has been optimized and upgraded in many aspects, and has achieved very good results in the double 11: on the one hand, based on the heterogeneous network and in-depth learning, the algorithm strategy has been upgraded, the personalized expression has been enhanced, the Matthew effect has been reduced, and the user utilization rate has been improved, such as the first page shading; on the other hand, the shopping guide ability has been enabled in various channels, and the user search mind has been enhanced , stickiness, improve the efficiency of various channels, such as the hot words on the home page; and innovate the existing products in the form of interaction, enrich the vitality of products, such as pull-down recommendation. On the 11th day of the double year, the re-use of UV for guiding and Purchasing Guide products (shading + hot words + venue search + wind vane) increased by 70% year-on-year, far higher than that of the search market.

IV. future work

  • 1. Mining of graphembedding. We have built the first version of metapath2vec algorithm for heterogeneous networks based on the computing platform graph mind. Compared with deep walk, the effect has been significantly improved. Next, we will continue to try, more flexible and convenient to obtain the information of nodes on different paths, combined with GCN algorithm to optimize the effect.
    1. Item2query mining. I2q data is a basic data of shopping guide, which will greatly affect the shopping guide effect. At present, the logic core is generated by user behavior, so there are problems such as low coverage, no data for cold start products, etc. We hope to generate candidate queries for products by combining the intelligent generation method.
  • 3. Mining from query to graph. Query can have a more intuitive feeling for consumers in the form of graphic display, so we hope that we can combine the algorithm of image processing to fully mine high-quality pictures that can represent query.
  • 4. Activate & pull down new form exploration. In the era of mobile Internet, users’ living habits and usage patterns are changing all the time. We expect to capture users’ needs and increase users’ access time and stay time through innovative product forms.


Author: Sike

Read the original text

This article is from alitech, a partner of yunqi community. If you need to reprint it, please contact the original author.

Recommended Today

The method of obtaining the resolution of display by pyqt5

The code is as follows import sys from PyQt5.QtWidgets import QApplication, QWidget class Example(QWidget): def __init__(self): super().__init__() self.initUI() #Interface drawing to initui method def initUI(self): self.desktop = QApplication.desktop() #Get display resolution size self.screenRect = self.desktop.screenGeometry() self.height = self.screenRect.height() self.width = self.screenRect.width() print(self.height) print(self.width) #Show window self.show() if __name__ == ‘__main__’: #Create applications and objects app […]