Baidu Feed recommendation is comprehensive information with hundreds of millions of users a month Streaming recommendation platform. The platform covers a variety of content types such as graphics, videos, updates, mini programs, Q&A, etc. It not only provides click-and-click recommendations similar to single or double columns, but also includes various recommendation forms such as video immersion. At the same time, the recommendation system is a multi-stakeholder system that not only includes C-side user experience. Content producers play an important role in the recommendation system. Baidu Feed has a large number of active practitioners, producing massive amounts of content every day.
The essence of the content platform recommendation system is to achieve a win-win situation for all parties. For the user side: the platform needs to continue to recommend high-quality, fresh, and diverse content to users to attract more users. Users, contribute more time; for the author side: users’ positive incentives promote authors to produce more high-quality content. On the contrary, if the high-quality and fresh content published by the author does not receive quick and sufficient exposure, Authors will choose to exit the platform, which is not conducive to the sustainable development of the platform. Based on the above discussion, several keywords can be extracted: freshness, high quality, variety, author publishing, and retention. This is closely related to what this article will discuss, cold start. First of all, more resources should be able to obtain sufficient display, and by collecting more content feedback, the amount of content that can be recommended by the system can be increased, thereby increasing the diversity of user consumption resources; secondly, new resources can be quickly produced to increase the number of users. The freshness of the content will in turn drive the market’s duration, DAU, and CTR; on the author’s side, it will increase the number of active authors and the amount of content published by stimulating the author’s enthusiasm.
#There are some differences between cold start of new resources and regular recommendation algorithms. The challenges faced by cold start can be summarized into three main aspects: 1. Data scarcity: New resources often do not have enough user behavior data to support personalized recommendations in the initial stage. This results in the recommendation algorithm being inaccurate.
The first is the challenge of accurate recommendation. With the development of recommendation algorithms over the past decade or so, from the initial matrix decomposition to the later widespread application of deep learning, the role of ID-type features in the model has gradually become more prominent. However, because the number of cold-start samples of new resources is rare or non-existent, ID-type features are insufficiently trained on cold-start samples, thus affecting recommendation accuracy.
Second, the Matthew effect is prevalent in recommendation systems, that is, resources that have been recognized by users are more likely to be recommended, thereby getting more exposure and clicks, further consolidating their status. Conversely, new resources struggle to get recommendations and may even be ignored entirely. Therefore, recommendation systems need to be continuously optimized to make them more fair and objective.
Finally, We need to provide certain cold start support to new resources, so how can we deal with them more efficiently and fairly? New resources for support? This introduces the two concepts of fairness and impartiality. Fairness refers to: each content product can obtain certain exposure opportunities in the early stage of cold start and have a chance to compete fairly. Fairness means: we must reflect the value of high-quality content, and the quality of the content must be able to affect the weight of Lengqi's support. Therefore, when it comes to new resources, it is also a big challenge to find the right balance between fairness and justice so that high-quality resources can stand out and maximize overall benefits.
The following are commonly used recall methods for new resources. Because the number of interactions between new resources and users is very small, the traditional i-to- The i (item-to-item) and u-to-i (user-to-item) recall methods do not apply. Therefore, cold start mainly relies on content recommendation methods. For example, direct recall methods based on the most basic user portraits, content tags and classifications have a low degree of personalization and relatively poor recall accuracy.
SecondlyAs more and more authors have personalized attributes on major content platforms, cold start based on attention relationships has become an effective method. However, the attention is relatively sparse and cannot satisfy the posts of many authors with low fan followings; so we go a step further and use algorithms to mine potential fans of authors to expand the influence of cold start based on attention. For example, users who often consume the author but do not follow it, and calculate the potential attention relationship based on the user-author attention relationship composition.
In addition, multi-modal recall is also an effective method. With the development of cross-modal, multi-model, and large-model technology, the integration of various modal information of content in recommendation systems has significant effects, especially in cold-start recommendation systems. CLIP is a pre-training method based on comparing text and images. It mainly includes two modules: a text encoder and an image encoder. It maps text and image information to the same space, providing better help for downstream tasks. There will be certain problems in directly using this vector for recall. This vector represents the prior information of the content. Prior similarity does not necessarily mean that users will like it. We need to combine the prior representation with the behavior-based representation in the recommendation system. The learned posterior representations of the data are associated.
The specific mapping method is based on the distribution of sufficient embedding and sufficient learning resources. Some samples can be collected and used as labels for training the projection network. This projection network maps the cross-modal prior representation to the posterior behavioral representation of the recommender system. One advantage of this approach is that the existing recall and ranking models in the recommendation system can be seamlessly used without adding any models. For example, for the Twin Towers model, we only need to utilize the existing user-side vectors without making any changes, and then use the projection network to project the new resources into the posterior representation space of the Twin Towers model, so that we can simply and quickly A twin tower recall is online. Similarly, the existing graph recall and tree-based recall can also be implemented at a low cost.
Of course, this mapping method has a small disadvantage, that is, regression is more difficult. In CB2CF, this is a regression problem, and regression is generally difficult to learn. Therefore, we can also use a pairwise approach to learn mapping relationships. Specifically, the positive samples can be set to similar item pairs learned by item CF. The negative samples can be obtained through global negative sampling, etc. The input also includes some prior and dynamic information of the item, and then such a mapping.
By utilizing the prior information of the content, it is basically possible to effectively implement the commonly used recall methods on cold start.
Since the early cold start mainly targets some zero Design with clickable resources, and once you have some early cold starts, the content will collect a certain amount of positive feedback seeds. At this point, we can recall through the lookalike method.
An important advantage of Lookalike is that it is extremely real-time. This method mainly comes from the field of Internet advertising. In the past, advertisers would select some potentially interested users as seed users, and then the system would look for similar users of these seed users to spread. In the recommendation system, we can subscribe to online real-time streaming logs to obtain positive feedback on resources collected during previous cold starts, such as clicks, plays, interactions, attention, etc., and even negative feedback, such as users who swipe quickly. Then based on these seed users, the system can obtain the representation of the item through the user's embedding, various aggregation methods or adding some self-attention mechanisms. This representation can be updated very quickly, and then spread outward based on this representation, which has a very high timeliness.
In terms of models, the optimization points of cold start can be summarized into three paradigms: ID discarding, ID generation and dynamic parameters of the model. These three paradigms can actually be used in combination with each other.
For ID discarding optimization, due to the small number of overall resource samples, the model easily caters to the head resources. Therefore, the ID learning of the head resources is very sufficient, and the feature importance in the model is also Extremely high. However, the occurrence of cold-start resources is less and ID learning is insufficient. There are two ways of thinking about this problem: one is to avoid using IDs as much as possible, and the other is how to make better use of IDs.
The first paradigm is dropout optimization, and one of the classic methods is DropoutNet. During the training process, DropoutNet will randomly discard item ID and user ID features to maximize the model's emphasis on non-ID features and enhance the model's generalization ability. Doing so can actually improve the cold start effect of new users or new resources.
In addition, some comparative learning methods have also emerged in recent years. Contrastive learning is a self-supervised learning method that does not rely on manual annotation and can construct a large number of samples, which helps optimize the multi-band cold start problem because we can construct additional samples to strengthen the status of the cold start data. For example, in the two-tower model, an auxiliary contrast loss can be added on the item side. The parameters of the two towers are shared. Using contrastive learning loss can affect the network parameters and embedding features of the resource tower. Through the masking method, samples with ID features and other cold start features are masked in different proportions, thus taking into account The generalization ability of the model and the specificity of cold start resources.
#The next step is generative optimization. As mentioned earlier, unreliable ID features should be used as little as possible, but currently the better approach is How to make it more reliable. The conventional idea is to initialize the embedding of ID based on the prior characteristics of ID. Through reasonable initialization, the prediction of new resources can be more accurate and converge faster. Taking the twin tower model as an example, usually, new features will be randomly initialized or initialized with all zeros, which will lead to the prediction of new resources. Inaccurate and slow to converge. Therefore, you can use some a priori features of the content, such as tags, content tags, author tags, etc., as well as some similar IDs (such as popular IDs), to select some ID embeddings with sufficiently high posteriori and high distribution resources as tags, Then train a generator to generate an embedding of the ID to replace the initial value. Of course, you can also directly average the ID embeddings of the new resource and the most similar top K popular resources as the embedding initialization of the new resource. This method has a relatively stable effect and very low cost, and is often used in the industry.
For the problem that popular IDs dominate the model and the model relies more on ID features, we can use multi-task and multi-scenario ideas for optimization. Still taking the twin-tower model as an example, the prediction of cold start and non-cold start resources can be split into two independent goals. Through common multi-objective models, the model pays more attention to new content. A classic approach is the CGC network, shown on the left in the figure above. In this kind of network, all tasks share the embedding layer, and then independent expert networks are learned through cold start tasks and non-cold start tasks respectively to improve the ability of cold start prediction. Another method is to adjust the parameter weights of different resource types in the network through dynamic weights, as shown in the right part of the above figure. In this network, the rightmost network is a cold start indicator, which receives information about cold start resources (such as the current number of click impressions and resource types), and then outputs the weights of each layer of the network to control information under different resource types. The transmission channel in the network allows the model to predict more accurately under cold start conditions.
New resources need to be produced as soon as possible to improve the author’s publishing The realization of experience and recommendation, but due to the Matthew effect, we need to give a certain tilt to new resources. General cold start tilt can be divided into two flows: base flow and boost flow. Basic traffic means fairness, and we need to give all resources some inclusive traffic for testing. Boosting traffic provides differentiated support based on the estimated potential of the author's quality resources and the performance of primary traffic.
The cold start support mechanism has two parameters at the abstract level: time and distribution volume, that is, through forced insertion, power adjustment, etc., resources are allocated within a given time Reach the given distribution volume target. For different businesses, we will set different distribution volumes and required times. For example, for ordinary resources, 100 impressions may be enough within 24 hours; for new and hot resources, it may be faster, such as 3,000 impressions within half an hour. At the same time, a larger cold start quota may be set for new authors.
Specific to the formula, t in the formula represents the normalization of the current release time divided by the time required for the target, that is, the current time progress, and x represents the current distribution progress. We want t and x to be equal, which means distribution at normal progress. If x is less than t, it means that the current cold start speed is slow and it is necessary to increase the weight or force the coefficient. θ in the formula can control the degree of inclination of resource allocation in the early stage.
However, the premise of this formula is that the traffic of the product in different periods is uniform, but the actual situation does not meet this assumption. The traffic distribution of general Internet products has peak and trough differences, so it needs to be adjusted according to the actual situation. For example, if a piece of content is published at 2 a.m., it may only require 25 deliveries by 8 a.m. because there is less traffic during the early morning hours. Therefore, t in the formula needs to be integrated based on the actual flow distribution.
Another key issue is that in the early stages of resource distribution, the To which users are resources allocated? The most common approach is to try to recommend new resources to old users rather than new users, because old users are usually more tolerant and can avoid inaccurate new resource recommendations that would harm new users. In addition, if the improvement of cold start resources is regarded as an intervention, based on the Uplift idea, the impact of intervention on user duration and retention can be learned, and users who will not have a negative impact on the intervention can be selected for cold start.
The above two points are based on the influence of C-side users. However, the selection of cold-start audiences will also affect the subsequent communication development of resources. From the perspective of information dissemination, the two-level communication theory divides information dissemination into two steps. First of all, among the large amount of information generated every day, some groups of people have the ability to filter and promote information, which we call opinion leaders. Then, the resources amplified and promoted by these opinion leaders will spread on a large scale.
In the current era, the role of opinion leaders also exists on social platforms, well-known media, TV stations, etc. For recommendation systems, there is also the concept of key node user resources. They filter high-quality resources and make recommendations, thereby affecting the consumption behavior of other users.
So, how to tap these key users? Through the above discussion, key users have two characteristics: first, they have a high ability to identify resource quality, and second, their recommended content has a high probability of being accepted by other users. Therefore, there are two mining methods:
First, divide the resources into high-quality and low-quality resources according to their posterior conditions and use them as labels. Then the user IDs who initially clicked on these resources are used as features to predict the posterior situation of the resources. The weight of each user ID learned by the model can be regarded as the key index of the user.
Secondly, through the online user collaborative filtering recommendation system, the recommendation success rate between users is mined. Users with a higher recommendation success rate can be regarded as key users in the recommendation system. Through these two methods, key users in the graph are mined and recommended to them first when resources are cold started.
There are some special things that need to be paid attention to when designing the experimental system for cold start content Because the samples of the recommendation system are shared, the feedback collected by the experimental group will also be learned by the control group, making it difficult to accurately measure the effect of the cold start strategy. Therefore, we need to conduct content isolation experiments to evaluate the impact of the cold start strategy on the entire system.
A common experimental design is to completely isolate users and resources, such as shown in the lower left part of the above figure. Among them, 50% of users can only see 50% of the content, and different resource groups use different cold start strategies. This allows you to evaluate the impact of the cold start strategy on the entire system. However, this method may have a greater impact on the experience of C-side users, because they can only see part of the content.
Another soft approach is to completely isolate users and resources during the cold start phase, such as the first 3000 times, and then implement different cold start strategies for different groups. After a cold start, resources can be distributed to all users. Such a design can reduce the impact on the C-side user experience.
Through experiments, we can analyze the following indicators:
A1: The judgment of hot and cold towers is usually based on the amount of resource distribution. Generally speaking, resources with lower distribution volumes are considered cold towers, while resources with higher distribution volumes are considered hot towers. For example, a resource that has been distributed less than 100 times can be considered a cold start resource. Of course, it is necessary to analyze based on the prediction accuracy of the online model, and determine specific judgment standards based on the actual situation.
A2: Quality boosting of cold start traffic usually involves an assessment of resource potential. Judging resource potential can combine multiple signal sources. For example, to determine whether it is a new hot topic in the field, you can comprehensively consider the information of the entire network, including the hot list information of each product, as well as topic discussions and attention in related fields, etc. For the evaluation of the value of resources, the quality of the author may be considered, including factors such as his performance in the early stages and interactions. By comprehensively utilizing this information, a more comprehensive estimate of resource potential can be made.
A3: When solving for ideal t and actual t, it can be reflected by observing the exposure curve. The exposure curve shows the exposure of resources in different time periods. The ideal t refers to the theoretical exposure progress calculated based on the set target time, while the actual t is determined based on the current actual exposure progress. of. In order to ensure that the actual exposure is consistent with the overall market trend, the proportion of the overall traffic needs to be stably monitored to ensure that the progress of the cold start is consistent with the overall traffic trend. If the progress of cold start is slow, you may need to increase the exposure or adjust other recommended strategies to speed up the progress. If the progress is too fast, you may need to slow down the exposure speed to avoid overexposure of resources.
A4: Regarding the cold start problem, it is actually very difficult to accurately measure the exact value of the effect. Nowadays, we usually compare the experimental group and the control group to see which one is better.
The above is the detailed content of Baidu recommended resource cold start practice. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
