AIRec personalized recommendation recall model tuning and participation

Introduction to This article is a summary of video sharing of "AIRec personalized recommendation recall model tuning and participation (e-commerce, content community as examples)", mainly by Alibaba product expert Zhilu to share with you the AIREC personalized recommendation recall model And the actual tuning demonstrations for these recall models in the e-commerce and content industries.

1. Typical recommended scenario

(1) Introduction of the scene concept

Scene, it is a concept set in smart recommendation. Whether the scene is on the upload data table or on the console, there will be related concepts. The scene can be understood as an entrance to user traffic, and a traffic entrance represents a certain access mind. For example, if a user enters a big promotion page, his mind is to see if there are currently interested products or products that have a desire to buy that are participating in the big promotion activities. During the big promotion period, you normally open the APP and enter a page to browse whether there are new and interesting products recently. At this time, you will have a different mind. Here is the distinction between user traffic entrances.

Once the user traffic entry is determined, the user access logic to be created for a page is basically determined to achieve the goal, which means that the logic from the user’s traffic entry to the user’s access logic determines our selection logic on this page. The selection logic is something that we can intuitively experience, including some strategies on the placement. In the end, we will bind it to a series of algorithm logic, which means that a scene represents the only selection logic and The only combination of algorithm logic.

If there are multiple recommended scenarios, but many of the recommended scenarios are essentially a set of algorithm strategies, we can merge them into one scenario at this time; but if there are differences, we can split it into different scenarios.

(2) The process of

To create a new scene without using cloud services, the general approach is to first circle the items, circle the database from the database which items are needed for the current scene to be used as recommendations, and then make the link open, such as all the items in the platform User behavior data, analyze how we can make such a new scene page to avoid training and prediction from scratch when we make recommendations to users, but to reuse some data.

In the overall process, first is data docking, and then data verification. If it is based on self-built mode, it is also necessary to prepare feature engineering, including to assemble the links of recall ordering, and the strategies involved in business orchestration, so relatively speaking Its timeline will be very long, including if you want to set up a recommendation system, you must also consider offline near-line and online link designs.

(3) Quickly build a personalized recommendation page

We can quickly build personalized recommendation pages by using smart recommendation products. Starting from the docking service, completing the data docking and interface debugging, it will automatically pull up the industry-customized algorithm template and start various calculation logics, including table reflow logic tasks. After the service is started, you can quickly customize and publish the scene.

After landing and going online, it is necessary to customize the scene and optimize the business. There are two main methods. The first is to solve the problem of business adaptation through the dimension of the algorithm. The second is to solve this problem through the dimension of operational strategies, such as customizing some product selection rules and delivery rules, such as tilting the support strategy, and adjusting the overall strategy according to the timeliness of the recommended items.

The path of scene tuning is to customize and optimize the algorithm and operation strategy of the scene separately after we complete the service first.

2. Introduction to the classic algorithm model

(1) Collaborative filtering

Algorithm logic

The collaborative filtering of the intelligent recommendation application is itemCF, and the algorithm logic is that according to the input platform behavior data, combined with the judgment of whether to click or not, a table similar to the PPT is generated, the scoring table between each item is calculated, and the ID is clicked. After the item is equal to 1, the probability value of the point ID is 2. This probability value represents the similarity between the two items.

Mode of operation

First, find the item on the left. When looking for the item on the left, you need to combine real-time user behavior. For example, the user clicks on the lipstick product and hits the unique lipstick product ID. In the next brush, knowing that the user is interested in the current lipstick, recommend some more similar lipsticks, which are easier to get more clicks, and then further converted into order data, according to the scoring situation in the figure, the item with ID 2 Recommend to this user, this is the entire collaborative filtering, from the user's behavior to the retrieval of the form we calculated, and finally to the user to add to the recall link such a process.

Optimization operator

Convergence optimization of parent category and subcategory: Compared with the item association of beer diapers, it can help learn items that are actually more similar in nature, because they belong to the same parent category or subcategory. The logic is to look at first. If these two, for example, 1 and 2 in it, belong to the same parent category or the same subcategory, it can be considered that their similarity is relatively high. , You can make it more likely to appear in the link.

Swing: The system will consider the pairs of some users. The user behavior has increased the judgment logic for the importance of the overall item similarity calculation process. For example, the behavior of two users and the sequence of clicks they clicked are not too big. Similarity, but on the contrary, if two users are not similar, but they do have two items that have been clicked together, they will find that these two items may themselves have a relatively large degree of similarity. If a user’s behavior to him is relatively similar, and two users’ behaviors are more similar. In this process, the two of them click on the items that they hit together, and I might lower the right to it during the calculation. This method actually makes use of the ability of user collaboration to a large extent to discover the correlation between items.

(2) User history preference recall

Algorithm logic

We characterize a user’s preferences, mainly through the user’s behavior, such as the behavior of the past 30 days, and real-time behavior calculation and analysis, we will find some characteristics in the e-commerce industry that affect our consumption decision, such as the brand of the product , Shop labels and product categories. These may be the more important features that affect consumer behavior, so look at these important features and what preferences are mapped to users. Then analyze the current user based on the user’s historical behavior. He may have a preference for which categories and which brands he has a preference for now or in the past. Based on his real-time interest, we can also predict which ones it may be interested in in the future. What content of the brand has preferences, and the fusion of the two constitutes the user portrait that we can often say.

After constructing the user portrait, we then combine the portrait to map these features on the item table to make a combination and display. In the e-commerce industry, you can also see that what we can customize in this link is its closed and enabled status, as well as our maximum number of recalls, including a priority within us, which will involve Many features, such as category, brand, store, and label.

So when choosing these characteristics, one aspect of our industry, under our business model, are these characteristics the main consumer decision-making characteristics of users.

But another aspect is the maintenance of these features. For example, we may label very well, and I can label the style of the dress well. For me, I can use its advantages to a large extent. I can increase the priority of the label. This is a strategy and way that we can combine such recall links to optimize.

(three) vector recall

Algorithm logic

Vector recall embedding is a commonly used algorithm when the feature dimension is high. By mapping the multi-dimensional features into a vector expression, the vector distance is calculated, and the similarity score is produced. For example, based on the vector recall of the title, we will first analyze the word segmentation of the item title through NLP, and after getting a word vector, we use the word2vector method to calculate the similarity correlation between the word vectors, so when there are 2 When we want to analyze whether the items are similar, we can first find out the vector expression of the item and compare the distance between the vectors. The shorter the distance, the higher the similarity. Tag-based vector recall is also a similar idea. Based on the user behavior sequence, it can be understood that in a conversation, the user will produce a series of click sequences. These click sequences are like a sectence of interest circulation, using the same idea, like A coherent title expressed by a product, we can also calculate the similarity between items in a similar way.

If you don’t know how much vector recall can bring to our effect, we can also carry out some platform-based experiments, and then check the experimental report to see if the recall is this way. Played a more important role.

(4) New product algorithm strategy

The new product algorithm strategy is easier to understand, which is the promotion strategy of some of our newly released products or content. In the promotion process, we must first let the entire system know which products and which content is new. This requires real-time updates of our fields, such as pub\_time and fields, or more accurate updates.

In this process, we all know that the new product is a cold-start problem. If we do not have any behavioral conditions, and do not know the new quality conditions to do a distribution, it may hit our performance data, because there are The quality of some new products is not very good. If we recommend them, we will lose some clicks and some purchases.

In this process, we can provide a strategy-based solution. As mentioned above, we have calculated some interest tags of the user based on the user’s historical preferences. At this time, we can also use the algorithm strategy of the new product. For example, we can support based on the user's preference category, prefer brands to support, and we can support based on tags. Except for some industries, our requirements for new products will be a little higher. I don’t necessarily require him to be personalized, but may require him to rank based on the overall popularity scores of our new products after it’s released. I just want to find one The most potential new product, and then it is also possible that I will give priority to the newly released content first, we will adjust the strategy, such as the priority of the sixth point to make an adjustment, in fact, how do we choose a suitable one? The algorithm model is also based on our essential business requirements. For example, the content industry, especially the point where we need to stimulate creation, we have to open up new products, regardless of its traffic, and its caliber. A bigger hole allows new products to be exposed more effectively.

(5) Other typical recall algorithm models

In addition to the algorithms mentioned above, we also have some other typical recall algorithms and sorting algorithms. The intelligent recommendation standard version has archived some for everyone, such as collaborative filtering, user preference recall, new product recall, and vector recall. In addition, these recall links allow everyone to optimize some parameters. If you have higher-level requirements, such as want to use high-level algorithm models, we may need to process the data ourselves, perform some feature engineering processing, produce a score sheet, and then register it with our online model. Used in combination, this is a function that the premium version of the product will provide.

III. E-commerce industry optimization best practices

The optimization best practices of the e-commerce industry mainly revolve around three dimensions

The first is how to provide real-time feedback experience in combination with the refresh process of c-end users.

The second one is how to do a good job of feature tilt and effect improvement when we make this template, from standardized products to personalized recommendations embedded in our business system.

The third is personalized recommendation. It actually belongs to a traffic portal. How do we let users maximize the value of our platform marketing reach from the traffic portal? We may have some exposure filtering and click filtering strategies. To combine to configure and use.

(1) Real-time feedback experience is improved

First, how to understand real-time feedback? That is, when our user has some behavior in real time, we will further follow up and feedback in the next recommendation result to him. For example, I gave you a screenshot of a Taobao page. First of all, we saw that it exposed a lot of products. Of course, the latter two products were not fully exposed. We considered it to be an invalid exposure. After 4 products were exposed, the user may He is more interested in the makeup gift box endorsed by Zhang Yunlei. He will click on the product to view the details. This is a click behavior. If he is more interested, he may also make an additional purchase.

In the process, we will find that the user is more interested in cosmetics and skin care related content, and his recent purchase behavior indicates that he has a purchase intention. We can first collect this click behavior and send it back to the recommendation system in real time. In addition, we hope that when the second swipe, three swipes, n swipes, we will also give feedback based on its interests. At this time, we can use the system that was just in the system As I mentioned in the filtering algorithm, we combine a convergence optimization with category. For example, we are based on the category of the product. The category of the current product is the category of make-up or make-up suits belonging to domestic brands. Under these categories, we can give priority to recall, and recommend them to users that are more similar to the current products. This means that when we configure the algorithm priority, we can increase the priority of category convergence.

(two) feature tilt and effect enhancement

For example, the characteristics of some products we gave to the right are firstly beauty sets, then the brand is Zhiyouquan, the store is Zhiyouquan Tmall flagship store, and the label is the celebrity co-branded gift box, cosmetics. There may be some malls that emphasize the concept of stores, and some may sell many brands in one store. We need to consider which store and brand are more important to our industry. We will add some stores and brands. One feature, in addition, we believe that our label is very important in consumer decision-making, and we can adjust the priority of the label.

The logic of judgment is what we think of consumer decision-making or interest-based decision-making. What are its primary characteristics? What are the secondary characteristics? Whether our maintenance is relatively high-quality, to adjust a priority of this recall to improve our recommendation effect.

(3) Marketing Reach Strategy Application

When buying a product, especially when a girl is making a purchase decision, she may buy a lot more. After a period of time, she has not placed an order, but if she repeatedly recommends it to her, she will think that she still wants to buy it. It placed an order. This is actually the psychological logic of consumers, that is, after an exposure, click, collection and purchase behavior occurs, the purchase may not be made quickly, and the order may be placed after the intention is touched in a certain period of time. In the process We can consider trying to use some marketing strategies in the platform to help users make consumer decisions.

For example, here we configured the exposure filter time to be 3 days, and the click filter time to be 1 day, which means that we saw these products for the first time, for example, the sun hats and clothing were exposed and clicked in these 4 products. The other is only exposure. After one day, it may not be in this order. On our page, it may have another product that I clicked on before. The system gave him a chance to repeat the exposure. If the user clicks again this time, the system You can also give him another chance to repeat the exposure. If this re-exposure opportunity triggers a user’s purchase, the system will no longer recommend it at the time of such an exposure filtering.

If the system gives a chance to re-recommend, but the user does not click on it, it means that the user does not have it for the time being, and the system will not recommend it again. The ideal state is to go in the process of repeating the recommendation. Promoting consumer decision-making and placing orders is also a common strategy in our e-commerce industry.

Like some long videos and long content, we may also use similar strategies when it involves the optimization of our stay time.

IV. Best Practices in the Content Industry

(1) Multi-region / more than Feature divided page construction

If the content community involves multiple regions or a division of multiple features, it may also involve some user logic. For example, for some users, we need to do some special filtering for it. Youth mode, such as special VIP users. , Do not show him a certain type of label. In this process, if we choose the previous form of scene construction, many scenes may emerge. Hundreds of scenes are very difficult to maintain for operation and maintenance. At this time, we have an internal recommendation function called online attribute filtering. You can use the characteristics of geographic location, the characteristics of sub-categories, and the characteristics of special tags on the video, to perform some intersection and union on it, and finally get it. The result of a recommendation filtering is then bound to a scene. We can perform multiple logical assembly of such attribute filtering on this scene to produce multiple recommended landing pages, so as to improve our operation and maintenance efficiency. Facilitate us to tune faster.

(2) Timeliness adjustment of recommended content

Timeliness refers to the approximate dimensions of the time distribution of the content we hope in the results of our recommendation? For example, in some of our industries that have strong timeliness requirements, especially the news industry, we would like to recommend the content if it is more than 5 days old, and the publication time exceeds 5 days, then no recommendation will be made, and after it is published to it In the process of failure, we may also have some operations such as loading and unloading. In this process, we can combine our timeliness to set the item filtering rules, and in the content we urgently need to publish, we can give it a weighted operation, and then I can also carry out the control of the shelves to ensure that it expires. An effective distribution can be obtained within.

(3) High-quality author motivation

The last point is the incentives of quality authors, especially in the support of new content. For new products, for example, I will request that the releases within the last 20 days and the releases within the last 7 days can be effectively promoted, because this is also the content community’s One of the vitality, we hope that the newly released content can be effectively exposed. First of all, we can set the caliber of a new product. We believe that what is released within a few days is a new product, and then it is the traffic to it. The traffic is from a statistical dimension. Yes, such as how many products and content are distributed on our entire platform. How much of this content belongs to the overall traffic distribution within the caliber, rather than the traffic distribution specific to each user, and the new product distribution strategy just mentioned Whether to distribute according to interest, or according to the release time, or according to the popularity, this can also be adjusted according to our business needs.

V. Conclusion

The above is the content shared with you this time. If you are more interested in this product, you can try our first purchase of 100 yuan in the first month. After the trial is completed, you can also upgrade the supporting standard version to unlock our higher levels. The intervention optimization of the recall model and the function of the experimental platform.

Thank you for reading.

Copyright Statement: content of this article is contributed spontaneously by Alibaba Cloud real-name registered users. The copyright belongs to the original author. The Alibaba Cloud Developer Community does not own its copyright and does not assume corresponding legal responsibilities. For specific rules, please refer to the "Alibaba Cloud Developer Community User Service Agreement" and the "Alibaba Cloud Developer Community Intellectual Property Protection Guidelines". If you find suspected plagiarism in this community, fill in the infringement complaint form to report it. Once verified, the community will immediately delete the suspected infringing content.