Author: Origin, Zhengchao

I. Introduction to the cold start problem

1 What is cold start

In the recommendation system, there are thousands of users and thousands of items. The essential task of the recommendation system is to recommend the items that users are interested in. In this system, both users and items are constantly updated. How to recommend interesting items to new users and how to recommend new items to interested users is the cold start problem of the recommendation system. Therefore, the cold start problem of recommendation system mainly includes two categories: user cold start and item cold start.

2 the importance of cold start

The flow and uncertainty of users are objective facts, as are the listing, updating and removal of items. In the era of information overload, the uncertainty of users is more obvious. How to recommend good items to these uncertain users is one of the main functions of the recommendation system. Since users and items are constantly being created and are the norm on the Internet, cold start problems can be with a product throughout its life cycle.

Every product on the Internet is looking at MAU and DAU, and in an era where traffic is king, users play a crucial role in whether a product can survive and survive well. Whether new users are satisfied with the product or not is directly related to the user growth and revenue growth of a product. In business, the customer is king, and on the Internet, the saying still holds true, and it’s played out to perfection.

In addition, whether a product can bring forth new products is the key to attract users. In a sense, the quality of goods directly determines the quality of a product.

Therefore, how to solve the problem of new users and new items, namely the cold start problem, is very important for the recommendation system.

3 cold start method

According to the different characteristics of users and items, different cold start methods will be adopted, which will be explained separately next.

3.1 Cold Startup

Non-personalized recommendation

Popular recommendation is a good method, although there is no personalized, but many people have herd mentality, according to the 80-20 principle, recommend popular items to new users, can meet the needs of 80% of users. Such as popular movies, popular songs, and popular short videos.

Use registration information to recommend

At present, many apps require users to register before they can use them, so personalized recommendations can be made according to the registration information. For example, dating websites can recommend beautiful women to men and handsome men to beautiful women.

In addition, according to the registered age, region, occupation, education, income and other information to form a portrait, and then according to these portrait personalized recommendation.

Make recommendations based on points of interest

At present, some apps require users to select their interest points before using them, so that the recommendation system can make good recommendations. For example, news apps require users to select the tags they are interested in, game apps require users to select the types of games they are interested in, and music apps require users to select the music styles they are interested in.

Make recommendations based on a small number of behaviors

Some users are less active and have few behaviors, but personalized recommendations can be made based on these rare user behaviors. For example, users can make recommendations based on a short video they have seen.

Use a tentative approach to make recommendations

The method of detection and utilization is one of the common methods of recommendation system. Firstly, recommend several items randomly to the user, and then obtain the user’s interest according to the user’s feedback. This method is mainly applicable to apps that consume less time and can quickly locate users’ interests, such as news and short video apps.

Make recommendations based on the interest migration strategy

Some companies have mature apps, or there are multiple types of recommendations on one app, so they can transfer recommendations based on other users’ interests. For example, some apps recommend both music and short videos, so they can recommend related short videos according to users’ interests in music.

3.2 Cold start of items

Make recommendations based on Side information

Goods naturally have some attribute information, such as the merchant, classification and price of goods, as well as the language, style, music style and musical instrument of music, etc. The recommendation system can recommend these basic information to the corresponding interested users.

Make recommendations based on a small number of behaviors

Some items have a small amount of behavioral information, so personalized recommendation can be made based on this small amount of behavioral information. For example, if a user plays a short video in its entirety, the short video can be recommended to similar users.

Use a tentative approach to make recommendations

The method of detection and utilization is also applicable to the cold start of items. Firstly, a cold start item is randomly distributed to a group of users and recommended to the corresponding interested users according to the feedback of users.

4 evaluation index of cold start method

To evaluate the quality of a cold start method, consider the following three points:

coverage

The first evaluation index that needs to be considered is coverage, which directly determines the quality of the online effect. If the coverage is too low, no matter how good the effect is within the online coverage range, the overall effect will be greatly reduced. As for the methods described above, both the cold-start method of goods based on Side Information and the non-personalized cold-start method of users have high coverage, which can cover almost 100%. However, the user cold start method based on a small number of behaviors and the user cold start method based on interest migration are relatively strict, and their coverage is not so high.

The accuracy of

The second evaluation index that needs to be considered is accuracy. For example, the user cold start method based on interest migration has relatively high recommendation accuracy due to the large amount of user information, while the method based on Side Information has high coverage, but its accuracy is not so high.

Can explain

The interpretability of recommendation in the recommendation system is very important for both users and the recommendation system. Now many recommendation systems pay more and more attention to the interpretability of recommendation. Similarly, for cold start problems, good interpretability also helps improve the accuracy of recommendations. For example, the user cold start method based on the point of interest can well explain the recommended items for users.

From the above evaluation indicators, none of the methods has all the advantages, so in the practice of recommendation system, a variety of cold start methods coexist, so as to achieve the effect of complementary advantages of various methods.

Above, we briefly introduced the definition of cold start problem, the general methods to solve the two types of cold start problem, and the evaluation criteria of cold start method. Next, we will introduce the practical scheme to solve the cold start problem of songs in the music recommendation system.

Second, cloud music song cold start practice

1 Service Background

At present in netease cloud music independent musicians exceeds 400000, independent musicians release of a large number of outstanding new work every day, how to quickly accurate will the outstanding works of new distributed to the target audience in the playlist, cold start and then enter the complete song song growth system is netease cloud music recommendation system to solve an important problem.

Due to the particularity of cold start songs, it is difficult to directly apply the recommendation model established for non-cold start songs. Therefore, an effective song recommendation model should be established for cold start songs.

2 song cold start facing problems

2.1 Absence of song features

The fundamental problem faced by cold-launch songs is the lack of historical user interaction data for songs, resulting in the absence of features and samples.

Lack of song statistics

It includes the frequency and conversion characteristics of various behaviors of songs, such as playing, downloading, collecting and sharing, etc., which are usually an important part of song recall and ordering model. Cold start songs cannot be directly used in existing models because they do not contain these characteristics.

Missing samples to train the embedding vector for cold start songs

Embedding in recall and sorting models of recommendation systems is usually trained end-to-end, while cold start songs do not exist in the word list, so the corresponding song representation cannot be directly obtained.

2.2 Interpretability of services

The ultimate goal of song cold start system is to serve the business. In addition to successfully distributing cold start songs, it is hoped that the process of cold start can be explained as much as possible. An interpretable cold start system will help businesses better answer questions such as what songs are more likely to succeed in cold start, thus providing successful experience for subsequent cold start songs.

3 Solution

The core idea to solve the cold start of songs is to increase available data as much as possible. The most widely used method is the cold start method of Side Information, which is usually simple to implement and has low requirements on data characteristics, and has good business interpretability.

The cold-start scheme based on content label will be introduced from two perspectives of cold-start song recall and cold-start song sorting.

3.1 Cold start recall

As user interaction behavior records of cold-start songs cannot be collected, cold-start songs usually cannot be recalled using conventional I2I or vector. However, cold-start songs can be recalled using content labels of cold-start songs. The process of recall is shown in the following figure

The first step is to pre-normalize the content labels corresponding to cold-start songs in the recall part, including only retaining the main music style, unifying minor languages, unifying album artists and singing artists, etc.

In the second step, cold-started songs are classified according to the content label, and each category is given candidate points for recall according to the conversion rate with time decay. The calculation method of candidate points is as follows:

It consists of three parts

Part 1: Time decay factor, days for cold start days, T for half-life, this part of the whole means that the more recent the new song score is higher.

Part 2: Smooth tag conversion, calculated as

Like7d and play7d respectively represent the number of people collecting and playing songs on the platform in the last 7 days, and K is used to adjust the initial conversion rate. This section represents the conversion rate of cold-launched songs using the tag dimension.

Part 3: Smooth conversion rate of cold start song itself. The results of the whole calculation process are as follows: 1) Smooth conversion rate is used as a candidate score for recall; 2) If there is no cold start song conversion data in the early stage of cold start, the content label conversion rate corresponding to the cold start song is used instead; 3) With the advance of cold start distribution time, the conversion effect of song content tags decreases and is gradually replaced by the conversion effect of actual songs.

3.2 Cold startup Sort

Similar to the recall part, the features of cold-start songs cannot be directly used in sorting, but can be degraded to the tag features of cold-start songs. The missing features of cold-start songs are represented by the music style trained by the non-cold-start song recommendation sorting model, the statistical features of languages and artists as well as the embedding vector.

The structure of the sorting model is shown below

Includes song-side and user-side feature builds, as well as sequencing models for user and cold-launched songs.

Song side

In this paper, the sorting model for non-cold start song training is presented to deduce the music style, the corresponding languages and artists’ embedding vector as well as the corresponding statistical features of song tags. Then query the features of embedding and statistics respectively according to the music style, language and artist label of cold start songs.

The user side

The preference sequence of non-cold start songs for wind, language and artist is calculated. After querying embedding and pooling, the label preference of users is expressed as the corresponding preference vector.

Ranking model: This part calculates the matching score by inner product of the vectors of the three dimensions of user’s music style, language and artist, and then adds the statistical features of song label dimension and the attributes of user dimension into the logistic regression model to predict whether the user. The training of the model is carried out on the user’s interaction samples of non-cold start songs, and the weight output of each is obtained after the training is completed.

Online prediction

As the user side preference vector, user attribute characteristics, cold start song side tag conversion rate and song label vector can be completed offline, online prediction only needs to query the song label, and then match with the user preference label respectively. Finally, by multiplying the weight of each dimension in the logistic regression model and inputting sigmoid, the ranking score of cold start songs can be obtained. Online prediction has less computation and faster prediction speed.

Finally, after scoring the candidate cold start songs by users, top songs are inserted into the distribution traffic of the existing song recommendation system to complete the distribution of cold start songs.

4 Service Effect

The most direct indicator to measure whether cold start songs are successfully distributed is the exposure coverage of cold start songs, that is, the number of users who have been exposed to cold start songs. However, cold-start songs are not always of high quality, and recommending low-quality songs to users may affect users’ playing experience. Therefore, cold-start songs cannot be distributed without limit. A good song cold start system should be on the basis of no or as little impact on the existing song recommendation system, so that more cold start songs are distributed.

The figure below compares the online experiment, with the recall and ordering of cold-start recommendation as the experimental group and the non-cold-start recommendation as the benchmark group. During the experiment, under the premise of not affecting the user experience index of the existing song recommendation system, the exposure coverage of cold start songs in the experimental group with cold start song system increased by more than 40% compared with the benchmark group.

5 subtotal

This section introduces the tag-based song cold start recommendation method. The method is simple and effective, and is an important part of NetEase cloud music song recommendation system at present. It is an important cornerstone for building a healthy music distribution ecosystem, and provides the first step to help the growth of many new songs with cold start.

In future installments of this series, we’ll continue to share more solutions and practical experiences with cloud music on the classic problem of cold startup.

This article is published by NetEase Cloud Music Technology team. Any unauthorized reprinting of this article is prohibited. We recruit all kinds of technical positions all year round, if you are ready to change jobs, and you like cloud music, then join us at [email protected]