APP cold start skills and strategies

Cold start is an important beginning in the entire recommendation system. Recommendation systems generally require a large amount of data to make more accurate recommendations. The cold start of an app may directly determine whether a new user will continue to use it. The cold start of a new item also affects the enthusiasm of the producer, so the cold start is very important.

Cold start problems are divided into 3 categories:

1. User cold start: What content should be recommended to new users?
2. Item cold start: To whom should a new item be recommended?
3. System cold start: Based on a "poor and blank" foundation, how to establish the association between new users and items that have never been recommended?

User Cold Start Idea

User cold start, the most common scenario is the cold start of new users. The path for a new user to be converted into an old user is: new user interest acquisition (building an initial portrait of cold start users) -> content consumption and interest convergence -> sedimentation of interest to become an old user. In general, the first step is to "do everything possible" to obtain user portraits or let users actively generate portraits. There are several methods to consider.

Utilize users' social attributes , such as gender, age, region, etc. When a user opens an APP for the first time, many APPs will prompt or leave an entry for the user to fill in relevant information. Even if the user does not actively input, you can try to introduce portrait information from external channels (channel portraits, matrix portraits, applist, etc.) (but you need to pay attention to user overlap and relevance). With this information, coarse-grained personalized recommendations can be made based on social attributes.

By utilizing the user’s relationship chain , we can collect information through operational activities (such as Alipay activities to collect friend relationships and parent-child relationships) or introduce it from the outside (third-party login or open API), and recommend content that friends like to users based on the principle of “birds of a feather flock together”.

By utilizing popular content and knowing “nothing” about the user, based on herd mentality and the 80/20 rule, you can try to recommend popular content to users. This method focuses on the scope and algorithm of popularity, and the effect will be better than random recommendations. The same goes for leveraging high-quality content.

(Left: Weibo, Right: Toutiao)

The indicators of user cold start can focus on the portrait indicators of new users (average number of interests, portrait coverage, portrait accuracy, etc.) and the active performance of new users (click-through rate, retention, etc.).

Suppose an international app can have a better recommendation effect based on nationality and gender at the beginning. How can it obtain this information?

1. Explicit method: As mentioned above, any method including asking users to fill out forms, guiding or motivating them is acceptable.
2. Implicit method: Use some content with particularly large differences in gender/nationality clicks to guess the user's gender and nationality, and then start biased personalized recommendations.

This type of hidden exploration requires skill in selecting items:

1. Popular: There is a certain degree of recognition. The item is too obscure, so most of the results will be "skip".
2. Representativeness and distinctiveness: Items should be distinguishable from each other and represent different interests of users.
3. Diversity: Since people have diverse interests, be sure to cover as many as possible to avoid the embarrassment of having “no options”.

Item Cold Start Idea

Recommended by using item content:

1. Similar item recommendations: Find content that belongs to the same category as the new item, and the new item can take advantage of it and be recommended as a similar item.
2. Item-related recommendations: Build an item information knowledge base based on expert knowledge and establish the relevance between items. For example, in the knowledge graph, another node can be found through a known node and relationship to make expansion recommendations.
3. For example, a user likes "Zhou Dongyu" (node). Zhou Dongyu is the leading actor of the movie "Better Days". Through this relationship, it is associated with the node "Better Days", so "Better Days" is recommended to the user.

(Pictures from the Internet)

Introduction to related algorithms

What are the commonly used algorithms involved during this period? Assume that user A is a new user with only a few portraits.

1. UserCF: User-based collaborative filtering calculates user similarity through user behavior, finds user B who is similar to user A, and recommends items that user B likes to user A.
2. ItemCF: Item-based collaborative filtering calculates item similarity through user behavior, finds item b similar to item a, and recommends item b to user A who likes item a.
4. (Left: UserCF, Right: ItemCF)
5. ContentItemKNN: Content-based filtering calculates the content similarity based on the content features of the items, finds item b similar to item a, and recommends item b to user A who likes item a.

UserCF and ItemCF use the same user behavior data, but with different statistical dimensions. As shown in the simple example below (1 means the user clicked on the item), UserCF calculates the similarity of users horizontally, and ItemCF calculates the similarity of items vertically.

UserCF and ItemCF both have the problem of "first mover" in cold start.

UserCF, new items must appear in the user's display list first, so that more people can give feedback on the item and the item can spread. Therefore, there is a problem of the first driving force, that is, where the first user discovers the new item.

ItemCF calculates user behavior at intervals (the log is huge and time-consuming) to calculate item similarity (if a large number of users have viewed item a and also item b, the two items are considered similar) and outputs an item relevance matrix. When a new item is added, it is not automatically added to the matrix table and a user must first discover the new item.

ContentItemKNN uses the content features of items to calculate item-related tables, and can update the related tables frequently without the problem of the first mover. However, it ignores user behavior and thus ignores the rules contained in user behavior. The results are low in accuracy and high in novelty, and the effect is generally not as good as collaborative filtering. However, if user behavior is strongly influenced by a certain content feature, the content filtering algorithm has its highlights.

The first driving force problem mentioned above, that is, the item cold start problem, also known as "new item trial", is there any way to solve it?

The trial launch of new items is “items looking for users”. If it is "users looking for items", the Matthew effect is likely to occur: popular categories have a lot of exposure and the long tail phenomenon is serious. There are two ways for items to find users:

1. Random test
2. Interest Test

Assume that we define items with less than 500 exposures as new items (information products generally have time limits, such as within 6 hours), represent new items and users as multi-dimensional vectors, calculate the distance between vectors, distribute them to more active users, and weight the sorting and re-ranking restrictions during the cold start phase. During the cold start phase, an item will have a stable click-through rate (or other comprehensive indicators), which will be the basis for its subsequent traffic distribution. Based on the click-through rate performance of small traffic, items with good performance will enter the next larger traffic pool, and items with poor performance will be eliminated or downgraded. The gradient traffic distribution strategy is a relatively common personalized recommendation "horse racing mechanism".

What indicators are used to evaluate the cold start effect of items?

1. Model Accuracy
2. Number of unit trials
3. Effective distribution coverage
4. Item quality rate
5. New item recommendation performance (click rate, duration, etc.)

Another point is that we also need to pay attention to the user's contextual information, including time information and spatial information, and follow some strong rules.

For example, for an e-commerce app, if a new user logs in during the summer, down jackets should not be offered; if a new user logs in during the Mid-Autumn Festival, information about the Dragon Boat Festival should not be offered. But this is not just in the cold start phase, the user's contextual information should be taken into account in the entire recommendation scenario.

Author: Zhang Xiaomiao Miu

Source: Zhang Xiaomiao Miu