Cold start recommendation method, device and storage device based on user interest migration

Abstract

The technical problem to be solved by the present invention is to use the historical behavior information of new e-commerce users on the content platform to extract user interest preferences and complete personalized user cold start recommendations in response to the defects of the existing technology. To achieve the above purpose, the present invention is based on a cold start recommendation method, device and storage device for user interest migration, and uses the behavioral footprints of new users on other platforms to construct user group attributes. Then, based on the theory that people are divided into groups, clustering is performed according to user interest characteristics, and users are divided into multiple subsets. Then, based on the historical behavior of users in each subset, product preferences are calculated as candidate sets for this type of users. By adopting the technical solution, it is possible to more accurately predict products that new users may like, so that new users can find products that may be of interest to them more quickly, and find suitable social circles for themselves more quickly, so that new users can get a more friendly user experience on the new platform, thereby improving the retention rate of new users.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a user cold start recommendation method based on user interest migration, and an apparatus and storage device involved in the cold start recommendation method. Background Art

[0002] As e-commerce continues to expand, the number of product categories and products on various e-commerce platforms has rapidly increased. Customers need to spend a considerable amount of time to find the products they want. This behavior of browsing through a large number of irrelevant products not only significantly increases the user's shopping time cost, but also causes consumers to lose their patience after exhausting their shopping experience, thereby damaging the profits of merchants. Therefore, how to enable users to quickly find the products they need in the product library and how to maximize the exposure benefits of their products has emerged as a result of recommendation systems.

[0003] As an information filtering system, a recommendation system can learn users' interests and preferences based on their historical behavior and then recommend items they'll enjoy. However, for newly registered users, there's no historical record to rely on, and without historical behavior to predict their interests, personalized recommendations fail, leading to the user cold start problem. However, the initial user experience almost always determines whether a new user will stay, making the user cold start a crucial component of a recommendation system.

[0004] Conventional solutions in the industry fall into the following categories: 1. Recommendations based on best-selling lists; 2. Coarse-grained personalized recommendations based on registration information; 3. Recommendations based on collected auxiliary information. However, these approaches utilize relatively limited user information, making it difficult to provide comprehensive and relatively accurate personalized recommendations for cold-start users.

[0005] In the prior art, the invention patent "Cold start recommendation method, device and electronic device" with patent publication number CN111814032A discloses a cold start recommendation method, device and electronic device, wherein the cold start recommendation method includes: obtaining the user feature value of the new user for the set user feature based on the user data generated by the new user through the use of a third-party application; obtaining the classification feature value of each product in the product set according to the user feature classification; obtaining the matching degree between the new user and each product based on the user feature value and the classification feature value of each product; and obtaining a product recommendation list for the new user based on at least the matching degree between the new user and each product.

[0006] This patent effectively utilizes third-party user data, but the user data involved is primarily demographic, which in practice cannot accurately predict user interests. Furthermore, the patent uses a single factor to sort the candidate set, leaving room for improvement in the accuracy of push notifications.

[0007] Therefore, we need to find a cold-start recommendation technology based on user interest migration, which can leverage the behavioral traces of new users on other platforms, explore their interest preferences, and complete personalized cold-start recommendations for users on new platforms. This can help new users find the items they need faster and more accurately, thereby improving their initial perception of the platform and enhancing the recommendation effect of the entire cold-start algorithm. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the above-mentioned defects of the existing technology and provide a method, device and storage device for solving cold start based on user interest migration technology, which utilizes the historical behavior information of new e-commerce users on the content platform to extract user interest preferences and complete personalized user cold start recommendations.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: a cold start recommendation method based on user interest migration, comprising:

[0010] Build a user set with both e-commerce and content attributes, and obtain the historical behavior logs and content tag data of these users on the content platform;

[0011] Construct user features that represent the users in this group, cluster the users using the k-means++ model, calculate the product preferences of users in each group, and use them as the product recommendation list for the group.

[0012] For new e-commerce users, we construct a user feature vector based on their recent access logs on the content platform. We then use a distance formula to calculate the distance between the user and the centroid of each group. The group with the minimum distance is then selected as the group suitable for the user, and a product recommendation list for that group is recommended to the user.

[0013] The content platforms mentioned here can be existing platforms from other developers, such as social media platforms like TikTok and WeChat, or they can be content platforms from the same developer that are interconnected with the new platform. In either case, users must be informed and authorized to access content information from these content platforms.

[0014] Preferably, the step of clustering users using the k-means++ model specifically includes:

[0015] Based on the user's access logs on the content platform, LDA is used to build a user topic model and obtain the probability distribution of each user on the topic as the user feature vector;

[0016] Use the user feature vector as the input of the k-means++ model to cluster the users and obtain the centroid representing the final group and the user classification label;

[0017] Count the most popular product collections in each group as the recommendation list for users in that group.

[0018] LDA (Latent Dirichlet Allocation) is a document topic generation model.

[0019] Preferably, the specific steps of using LDA to construct a user topic model and obtain the probability distribution of each user on the topic as the user feature vector include:

[0020] The user-word matrix is ​​constructed using the user's access log on the content platform. Based on the user-word matrix, the user topic model is constructed using LDA to obtain the user-topic vector as the user feature vector.

[0021] Preferably, the user-word matrix specifically includes:

[0022] Eliminate noise data from log data, including content without label information and user data with too few user behaviors;

[0023] According to the content-tag dictionary, the content in the user's historical browsing log is mapped into tags to obtain the user content tag document;

[0024] The user content tag document is segmented and stop words are removed to obtain the vectorized representation of the document; the number of occurrences of each word in each document is calculated to obtain the user-word vector composed of the number of occurrences of feature words, and the user-word matrix composed of all user-word vectors.

[0025] Preferably, based on the user-word matrix, LDA is used to construct a user topic model to obtain a user-topic vector as a user feature vector, including:

[0026] Determine the range of topic number k based on experience, traverse k to train the LDA user topic model, observe the topic situation under the training model with different k values, and determine the appropriate topic parameter k;

[0027] Based on the optimal number of topics k, the distribution of users on each topic is obtained through training as the user feature vector.

[0028] Preferably, the user feature vector is used as the input of the k-means++ model to cluster the users, and the centroid representing the final group and the user classification label are obtained. Specifically, the following are performed:

[0029] Normalize the user feature vector.

[0030] Select the appropriate number of clusters K based on experience, and determine the K initial cluster centers based on the principle of maximizing the initial cluster centers;

[0031] Standard k-means clustering is performed to determine the final group centroid and user classification labels.

[0032] Preferably, calculating a set of the most popular products in each group as a recommendation list for users in the group specifically includes:

[0033] For each group of users:

[0034] Obtain the behavior logs of users in this group over a period of time, including key behaviors such as browsing, searching, adding to cart, and placing orders;

[0035] Execute for each product in turn:

[0036] (1) Count the number of behaviors of each product in weekly dimension, and perform maximum and minimum normalization processing on each indicator, which is recorded as

[0037] (2) Weight each behavior based on the combined weighting method i , where ∑w i =1

[0038] (3) The weighted sum is used to calculate the comprehensive score of the product in the statistical period t (t)

[0039]

[0040] in: w is the normalized number of occurrences of behavior i in the tth week of product; i is the weight coefficient of behavior i;

[0041] (4) Considering the timeliness of historical data, the historical scores are downgraded by the time decay coefficient to obtain the final score of the product:

[0042]

[0043] Where f(t-t0) is the time decay function, which is as follows:

[0044]

[0045] The products are sorted by their final scores and the top-scoring products are selected as candidate recommendations for each user group. The top-scoring products are randomly distributed on the page rather than sequentially.

[0046] Preferably, for a new e-commerce user, a user feature vector is constructed based on the user's recent access log on the content platform, and a distance formula is selected to calculate the distance between the user and the centroid of each group. The group with the minimum distance value is used as the group suitable for the user, and a product recommendation list of the group is recommended to the user, including:

[0047] Construct a feature vector for a new user based on the user's recent access logs on the content platform;

[0048] Calculate the distance between the new user and each centroid using the distance formula to determine the group to which the user belongs;

[0049] Output a list of recommendations.

[0050] This approach breaks away from the traditional labeling model and instead determines the attributes of user groups through comprehensive interest characteristics, making positioning more accurate.

[0051] The present invention also includes an electronic device, which includes: a memory and a processor, wherein the memory is used to store computer instructions, and the processor is used to call the computer instructions from the memory and execute a cold start recommendation method based on user interest migration under the control of the computer instructions.

[0052] The present invention also includes a computer-readable storage medium, which stores computer instructions. When the computer instructions in the storage medium are executed by a processor, a cold start recommendation method based on user interest migration is implemented.

[0053] In the prior art, the industry's commonly used cold-start recommendation algorithm constructs product features in the same format as user features, then selects a set of products that match the user's demographic characteristics as a candidate set for the user. However, the present invention utilizes the behavioral footprint of new users on other platforms to construct user group attributes. Then, based on the theory that people are divided into groups, users are clustered according to their interests, dividing them into multiple subsets. Product preferences are then calculated based on the historical behavior of users in each subset, forming a candidate set for each user. Using multiple behavioral factors and determining appropriate weights based on a combined weighting method, a comprehensive index is constructed for ranking, recommending items to new users that share their preferences. This cold-start strategy more accurately predicts products that new users are likely to like, allowing them to more quickly find products they may be interested in and quickly establish a suitable social network. This provides new users with a more user-friendly experience on the new platform, thereby improving new user retention. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0055] Figure 1 is a flowchart of the cold start recommendation method based on user interest migration in the embodiment of the present invention.

[0056] Figure 2 is a flowchart of the method for obtaining user characteristics in the embodiment of the present invention.

[0057] Figure 3 is the process of calculating the popular product list for group users in the embodiment of the present invention.

[0058] Figure 4 is a flowchart of the method for calculating the recommended list for new e-commerce users in the embodiment of the present invention. Specific embodiments

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0060] The following will clearly and detailedly describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments are only some embodiments of the present invention. <{

[0061] The present invention relates to the K-means++ algorithm. In the original K-means algorithm, K points in the dataset are randomly selected as the clustering centers at the beginning, while K-means++ selects K clustering centers according to the following idea: assuming that n initial clustering centers (0 < n < K) have been selected, then when selecting the (n + 1)-th clustering center: the points farther away from the current n clustering centers have a higher probability of being selected as the (n + 1)-th clustering center. When selecting the first clustering center (n = 1), it is also done by a random method. This algorithm verifies the improved idea of K-means: that is, the points farther away from the current existing clustering centers have a greater probability of being selected as the next clustering center.

[0062] The present invention discloses a cold start recommendation method based on user interest migration, as Figure 1 shown, including:

[0063] S1. Build a user set with both e-commerce and content attributes, and obtain the historical behavior logs and content tag data of these users on the content platform;

[0064] S2. Construct user features that represent the users in this group and cluster them using the k-means++ model. Calculate the product preferences of users in each group and use them as the product recommendation list for that group, including:

[0065] S21. Based on the user's access log on the content platform, use LDA to build a user topic model and obtain the probability distribution of each user on the topic as the user feature vector;

[0066] S22. Use the user feature vector as the input of the k-means++ model to cluster the users and obtain the centroid representing the final group and the user classification label;

[0067] S23. Count the top N popular products in each group, that is, the collection of the most popular products, as the recommendation list for users in that group;

[0068] S3. For new e-commerce users, construct a user feature vector similar to step S2 based on their recent access logs on the content platform. Use a distance formula to calculate the distance between the user and the centroid of each group. The group with the minimum distance is considered the most suitable group for the user, and a product recommendation list for that group is recommended to the user.

[0069] Step S1 is to construct a user set that has both e-commerce and content attributes, and obtain the historical behavior logs and content label data of these users on the content platform. Among them, having both e-commerce and content attributes means that the user is both a content user and an e-commerce user. The historical behavior log data of users on the content platform can come from various platforms, such as APP, WeChat applets, Douyin, etc. For self-developed platforms, user log data can be read directly from the database; for third-party platforms, user behavior data on third-party platforms can be obtained through API interfaces. Under the condition that the content has the same standard label, the content is not limited to articles, videos, etc. Content tags are generally established by manual tagging during content creation and automatic tagging by models, and can be directly obtained from the database when used. However, no matter which method is used to obtain it, it is necessary to actively inform the user and obtain the user's authorization permission.

[0070] like Figure 2 As shown, this is Example 1:

[0071] Step S21 is based on the user-topic probability distribution as the user feature, and includes the following steps:

[0072] Step S211: Utilize user access logs to construct a user-term matrix, including:

[0073] Step S2111. Eliminate noise data in the log data, including content without tag information and log data with too few user behaviors. Among them, content without tag information means that the corresponding tag cannot be matched in the content tag library. In order to reduce the computational cost, it is directly filtered; in addition, for logs with too few behaviors, it is easy to cause certain interference to the LDA user topic model. Generally, a threshold (such as 10, 20, etc.) can be determined according to the distribution of user access times, and user logs that do not meet the threshold can be directly filtered. Thus, for each user u, the user behavior vector [a1, a2…a s ], where a i Representative content.

[0074] Step S2112: Based on the content-tag dictionary, the content in the user's browsing history is mapped to tags to obtain a user content tag document. Because different content has different formats, it is necessary to map content to a tag dimension based on model requirements. Generally, each piece of content has multiple tags, including categories, keywords, and so on. After obtaining the content tag dictionary from the database, all content in the user behavior vector is mapped to content tags. After concatenation, the user content tag document is obtained.

[0075] Step S2113. Segment the user content label document and remove stop words to obtain a vectorized representation of the document; calculate the number of occurrences of each word in each document to obtain a user-word vector composed of the number of occurrences of feature words, and a user-word matrix composed of all user-word vectors. For user content label documents, most of them are composed of phrases or short sentences, so word segmentation is still required. It is worth noting that the dictionary that comes with the existing word segmentation tool is mainly trained based on resources such as the People's Daily corpus, and the recognition of related nouns in professional fields is not high. A custom dictionary is needed to improve the accuracy of identifying proper nouns. In addition, the words after word segmentation are processed to remove stop words to filter out some meaningless words and symbols. For example, for the content label vector shown in the following table:

[0076]

[0077] After processing, the input user / document-word matrix is ​​obtained as follows:

[0078]

[0079] Step S212: Based on the user-word matrix, LDA is used to construct a user topic model to obtain a user-topic vector as a user feature vector, including the following steps:

[0080] Step S2121. Determine the range of the number of topics k based on experience, traverse k to train the LDA user topic model, observe the obtained topic situation, and determine the appropriate topic parameter k. As an unsupervised model, LDA needs to be given the number of topics k during training. Based on the understanding of the basic framework of the platform content, a range of k can be determined by experience. Traverse k and train multiple LDA user topic models, and determine the appropriate k value by observing the distribution of topic-words. Alternatively, you can also construct a perplexity index, draw a curve of the change in the number of topics and perplexity, and consider the inflection point of the curve as the balance point where the uncertainty of the user belonging to the topic and the maximum benefit of the number of topics are maximized, as the k value of the model.

[0081] Step S2122. Based on the optimal number of topics k, the distribution of users on each topic is obtained through training as the user's feature vector. LDA is used to iteratively sample the user's tag text to obtain the user's probability vector on the topic, i.e., U i ={P(z i |u1),P(z i |u2)…P(z i |u j )}. Among them, P(z i |u j ) represents user u j In topic z j The probability value under , is used as the user's feature vector.

[0082] Example 2: Similar to Example 1, the embedding method is used as the user's feature vector. Unlike the LDA model, this method only needs to use the user's access footprint. The ID feature of the content is mapped to another low-dimensional space through the item2vec method to obtain the vectorized expression of the content. Then, based on the user's recent content access records, the user's vectorized expression is obtained by weighted average. The main steps are as follows: the user's access content list is divided into multiple session segments. Specifically, the user behavior list is sorted by time, and the time interval between the two previous and subsequent behaviors is calculated. If the interval is greater than a certain threshold (for example, 30s, 60s, etc.), it is segmented. At this time, the user's behavior sequence is converted into a session sequence. Based on the session sequence, the word2vec training method skip-gram is used to obtain the vectorized representation of the content. Based on the user's most recent N content access logs, the user's vectorized expression is obtained by weighted average as the user's feature vector. Among them, the weighting coefficient is the time decay factor, which is used to characterize the change of user interest decaying over time. The exemplary time decay formula is as follows:

[0083]

[0084] Where λ is the cooling coefficient, t-t0 is the time interval from the behavior time to now

[0085] In step S22, the user feature vector is used as the input of the k-means++ model to cluster the users and obtain the centroid and user classification of the final characterization group, including:

[0086] S221. Obtain the user's feature vector and perform normalization. To prevent the problem of sample distance measurement failure caused by dimensional differences, perform Z-score normalization on the feature. The formula is as follows:

[0087]

[0088] S222. Based on experience, select an appropriate number of clusters, K, and determine K initial cluster centers based on the principle of maximizing the initial cluster centers. To determine the initial cluster centers, first randomly select a cluster center, m1. Calculate the distance between the remaining samples and cluster center m1. Select the sample with the largest distance as the next cluster center, m2. Repeat this process K-1 times until all initial cluster centers are selected.

[0089] S223. Calculate the final cluster center and sample classification through standard k-means, and record the centroid vector and user classification label of each group.

[0090] In step S23, the top N popular products in each group are counted as the recommendation list for users in this group. When analyzing the popularity of the first-order products of new users, we found that most new users chose products with relatively high popularity. As for the definition of popularity, from the indicator level, there can be sales volume, click-through rate, conversion rate, etc.; from the data statistical period, there are offline and real-time. When calculating the popular product list of each group this time, a comprehensive indicator is constructed using the key behavior data of offline users as an evaluation indicator of the group users' preference for the product. Figure 3 As shown, the main steps are as follows:

[0091] For each group of users:

[0092] Step S231: Obtain the behavior logs of the group of users over a period of time, including key behaviors such as browsing, searching, adding to shopping cart, and purchasing.

[0093] Step S232. For each product:

[0094] Step S2321. Count the number of actions for each product on a weekly basis, and perform maximum and minimum normalization on each indicator. The format is as follows:

[0095]

[0096] Step S2322. Weight each behavior based on the combined weighting method i , where ∑w i =1 Among them, this paper adopts the combined weighting method combined with the hierarchical analysis method and the entropy weight method for comprehensive evaluation. Its main steps are as follows:

[0097] (1) Analyze user behavior to establish a behavioral hierarchy, use the hierarchical analysis method to establish a judgment matrix, and assign values ​​to the judgment matrix. For the extracted key behaviors of browsing, searching, adding to shopping cart, and purchasing, a 4-dimensional judgment matrix can be defined and assigned values ​​based on experience:

[0098]

[0099] Among them, d ij Represents the degree of importance of behavior i over behavior j; when i = j, d ij =1; According to the matrix definition, there is d ji =1 / d ij

[0100] (2) Perform maximum and minimum normalization processing on the judgment matrix, the formula is as follows:

[0101]

[0102] (3) Calculate the information entropy of the behavior importance index. The information entropy formula is as follows:

[0103]

[0104] in, n is the order of the judgment matrix

[0105] (4) Quantify the weight of each user's behavior. The formula is as follows:

[0106]

[0107] (5) Matrix consistency test: Check whether the hierarchical division is reasonable, including:

[0108] 1) Calculate the maximum eigenvalue λ of the judgment matrix max

[0109] 2) Calculate the consistency test index CR, the formula is as follows:

[0110]

[0111] If the calculated CR consistency index is less than 0.1, the manually defined judgment matrix is ​​considered reasonable, and the behavior weights calculated based on this judgment matrix can measure the differences between different behaviors. Therefore, after passing the consistency test, the behavior weights can be recorded as [w1, w2, w3, w4].

[0112] Step S2323. Use weighted summation to calculate the comprehensive score of the product in statistical period t (t)

[0113]

[0114] in: w is the normalized number of occurrences of behavior i in the tth week of product; i is the weight coefficient of behavior i.

[0115] Step S2324. Considering the validity of historical data, the historical scores are downgraded by the time decay coefficient to obtain the final score of the product:

[0116]

[0117] Step S233: Sort the products by their final scores and select the top N products as the recommended list products for each group of users. The specific number of top N products needs to be configured according to the component situation.

[0118] S3. For new e-commerce users, construct a user feature vector based on the user's recent access logs on the content platform, similar to step S2. Use a distance formula to calculate the distance between the user and the centroid of each group. The group with the minimum distance value is regarded as the group suitable for the user, and recommend a product recommendation list of this group to the user. Figure 4 , as shown below:

[0119] Step S31. For new e-commerce users, construct the user's feature vector based on the user's recent access log on the content platform and in the same manner as step S2.

[0120] Step S32. Use the distance formula to calculate the distance between the user and the centroid of each group to determine the group to which the user belongs. Use the same distance metric formula used in the k-means++ model to calculate the distance between the new user and the centroid of each group, and select the group with the minimum distance as the user's group. A commonly used distance metric formula is as follows:

[0121]

[0122] Among them, x i is the i-th element of vector X, y i is the i-th element of vector Y

[0123] Step S33. Output the recommendation list. Obtain the popularity list of the group to which the user belongs, and filter it again based on the user's other information to serve as the recommendation list for the new e-commerce user. Among them, the user's other information refers to some user information collected during registration, including the user's baby's age, baby's gender, user's location address, etc. Based on each feature dimension and the characteristic performance of the user's purchase, its popularity list can be filtered again. For example, for a mother of a 6-month-old baby, the recommendation list can filter out products that are not suitable for babies under 5 months old.

[0124] It should be noted here that the information used in the re-filtering based on other information of the user is different from the interest preference information in the subject matter of the present invention. The interest preference in the present invention is a product set with elements. For example, a mother has seen a lot of content about making fruit puree and diapers in the past week. In the prior art, it is often believed that complementary foods and diapers are the user's recent interest tags. In the technical solution of the present invention, it is further discovered that for users, complementary foods with apple and strawberry flavors, and diapers with cute cartoon graphics are the users' real interests. If complementary foods of other flavors are recommended, it will not attract the attention of this user, which will waste a lot of information resources.

[0125] Based on this information, the present invention performs commodity-based vector representation, which can convert user tags into vector features to accurately represent user interest preferences.

[0126] In addition, in order to ensure the operation of the present invention, it also includes:

[0127] An electronic device comprises: a memory and a processor, wherein the memory is used to store computer instructions, and the processor is used to call the computer instructions from the memory and execute a cold start recommendation method based on user interest migration under the control of the computer instructions.

[0128] And a computer-readable storage medium storing computer instructions, which, when executed by a processor, implements a cold start recommendation method based on user interest migration.

[0129] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A cold-start recommendation method based on user interest migration, characterized by: include: Build a user set with both e-commerce and content attributes, and obtain the historical behavior logs and content tag data of these users on the content platform; Construct user features that represent the users in this set, cluster the users using the k-means++ model, calculate the product preferences of users in each group, and use them as the product recommendation list for the group. The steps of clustering users using the k-means++ model include: Based on the user's access logs on the content platform, LDA is used to build a user topic model and obtain the probability distribution of each user on the topic as the user feature vector; The user feature vector is used as the input of the k-means++ model to cluster the users and obtain the centroid representing the final group and the user classification label. The specific steps of using LDA to build a user topic model and obtain the probability distribution of each user on the topic as the user feature vector include: Using the user's access log on the content platform, we build a user-word matrix. Based on the user-word matrix, we use LDA to build a user topic model and get the user-topic vector as the user feature vector. Counting the most popular product sets in each group as a recommendation list for users in that group; the user-term matrix specifically includes: Eliminate noise data from log data, including content without label information and user data with too few user behaviors; According to the content-tag dictionary, the content in the user's historical browsing log is mapped into tags to obtain the user content tag document; Perform word segmentation and stop word removal on user content tag documents to obtain a vectorized representation of the document. Calculate the number of occurrences of each word in each document to obtain a user-word vector consisting of the number of occurrences of feature words, and a user-word matrix consisting of all user-word vectors. Based on the user-word matrix, use LDA to build a user topic model and obtain the user-topic vector as the user feature vector, including: Determine the range of topic number k based on experience, traverse k to train the LDA user topic model, observe the topic situation under the training model with different k values, and determine the appropriate topic parameter k; Based on the optimal number of topics k, the distribution of users on each topic is obtained through training as the user feature vector; For new e-commerce users, we construct a user feature vector based on their recent access logs on the content platform. We then use a distance formula to calculate the distance between the user and the centroid of each group. The group with the minimum distance is considered the group that best matches the user, and we recommend a list of products for that group to the user. The following steps are used to calculate the most popular product sets within each group as the recommended list for users in that group: For each group of users: Obtain the behavior logs of users in this group over a period of time, including key behaviors such as browsing, searching, adding to cart, and placing orders; Execute for each product in turn: (1) Count the number of behaviors of each product in weekly dimension, and perform maximum and minimum normalization processing on each indicator, which is recorded as (2) Empower each behavior based on the combined weighting method ,in (3) The statistical period is obtained by weighted summation. The comprehensive score of this product in: For product Zhou Xing Normalized number of occurrences; For behavior The weight coefficient of (4) Considering the timeliness of historical data, the historical scores are downgraded by the time decay coefficient to obtain the final score of the product: in, is the time decay function, which is as follows: According to the final ranking of the products, the items with the highest scores are selected as the candidate recommendation list for each group of users.

2. The cold-start recommendation method based on user interest migration according to claim 1, characterized in that: The user feature vector is used as the input of the k-means++ model to cluster the users. The final centroid representing the group and the user classification labels are obtained. Specifically, Normalize the user feature vector. Select the appropriate number of clusters K based on experience, and determine the K initial cluster centers based on the principle of maximizing the initial cluster centers; Standard k-means clustering is performed to determine the final group centroid and user classification labels.

3. The cold start recommendation method based on user interest migration according to claim 1, characterized in that: For new e-commerce users, we construct a user feature vector based on their recent access logs on the content platform. We use a distance formula to calculate the distance between the user and the centroid of each group. The group with the minimum distance is selected as the group that matches the user. We then recommend a list of products for this group to the user, including: Construct a feature vector for a new user based on the user's recent access logs on the content platform; Calculate the distance between the new user and each centroid using the distance formula to determine the group to which the user belongs; Output a list of recommendations.

4. An electronic device, characterized in that: The electronic device includes: a memory and a processor, the memory is used to store computer instructions, the processor is used to call the computer instructions from the memory, and execute the cold start recommendation method based on user interest migration as described in any one of claims 1-3 under the control of the computer instructions.

5. A computer-readable storage medium storing computer instructions. When the computer instructions in the storage medium are executed by a processor, the cold start recommendation method based on user interest migration as described in any one of claims 1 to 3 is implemented.

Images