E-commerce recommendation system cold start method based on user behavior sequence
By collecting and processing user behavior data, and combining federated learning and domain adaptation techniques, an interest profile of new e-commerce users is constructed, which solves the data fusion problem in new user recommendation and improves recommendation performance and user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BANGLIDE TECHNOLOGY (SHANGHAI) CO LTD
- Filing Date
- 2026-02-27
- Publication Date
- 2026-05-19
AI Technical Summary
Existing e-commerce recommendation systems lack effective cross-platform, multi-source behavioral data fusion during the cold start of new users, resulting in the inability to build initial interest profiles, low recommendation click-through rates, and a long user-system integration period.
By collecting and preprocessing user behavior feature data within e-commerce and across platforms, calculating behavior authenticity values, and using federated learning methods to fuse multi-source features, perform domain-adaptive feature transfer, construct structured interest profiles, and generate personalized recommendation lists.
While protecting user privacy, the system increased the click-through rate of new user recommendations by 45%, significantly shortening the user's adjustment period with the system.
Smart Images
Figure CN122066498A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of e-commerce technology, specifically to a cold start method for an e-commerce recommendation system based on user behavior sequences. Background Technology
[0002] In e-commerce recommendation system applications, cold start specifically refers to newly registered users entering an e-commerce platform. Because they have not generated effective behavioral data such as consumption, in-depth browsing, and adding to cart, the recommendation system lacks basic data that can be used to analyze user interests and preferences, making it impossible to provide personalized recommendation application scenarios. This is also the core challenge faced by e-commerce recommendation systems when recommending to new users.
[0003] Existing e-commerce recommendation systems, when performing cold start recommendations for new users, can only rely on a small amount of basic behavioral data collected by the e-commerce platform itself, and cannot effectively integrate multi-source behavioral data across platforms. Furthermore, traditional multi-source data integration methods are prone to leakage of user privacy information, making it difficult to build an effective initial interest profile for new users. Ultimately, this results in low click-through rates for new user recommendations and a long interest integration period between users and the system. Therefore, this paper proposes a cold start method for e-commerce recommendation systems based on user behavior sequences to solve these problems. Summary of the Invention
[0004] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a cold start method for e-commerce recommendation systems based on user behavior sequences, thus solving the problems mentioned in the background section.
[0005] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a cold start method for an e-commerce recommendation system based on user behavior sequences, comprising the following steps: Step 1: Collect behavioral feature datasets of users who are cold-starting users within the e-commerce platform and users associated across platforms, and perform preprocessing. Step 2: Based on the preprocessed behavioral feature dataset, perform objective and comprehensive calculations to obtain the user behavior authenticity value; Step 3: Integrate the behavior authenticity value into the behavior feature extraction process, and use the federated learning method to fuse multi-source cross-platform behavior features to obtain a unified cross-platform behavior feature vector; Step 4: Perform domain-adaptive feature transfer processing on the cross-platform unified behavioral feature vector to obtain the e-commerce behavioral feature vector; Step 5: Based on e-commerce behavior feature vectors and behavior authenticity values, construct a structured initial interest profile for cold-start users; Step 6: Based on the matching relationship between the initial interest profile and the characteristics of e-commerce products, generate a personalized recommendation list for users in the cold start phase.
[0006] Preferably, the behavioral feature dataset includes effective dwell time, behavioral operation depth, behavioral reverse operation rate, cross-platform category collaborative matching degree, cross-platform behavioral effectiveness convergence value, behavioral timestamp, and behavioral object identifier.
[0007] Preferably, the specific steps for obtaining the authenticity value of user behavior are as follows: Perform dimensionless standardization on all data in the behavioral feature dataset; Perform comprehensive calculations on the standardized data; The user behavior authenticity value is obtained with a range of 0-1.
[0008] Preferably, the specific steps of the standardization preprocessing are as follows: Filter out abnormal behaviors and remove invalid data from the behavioral feature dataset; Unify behavior type coding and timestamp format to achieve data format standardization; Anonymous IDs are generated by performing an irreversible hash operation on the user identifier, resulting in a standardized behavioral feature dataset.
[0009] Preferably, the specific steps for fusing multi-source cross-platform behavioral features using the federated learning method are as follows: Each platform extracts behavioral feature vectors containing behavioral realism values locally and only uploads the encrypted model gradients; The master node aggregates the gradients uploaded from each platform using a federated averaging algorithm and updates the global model parameters; it iterates the training until the global model converges, and obtains a unified behavioral feature vector across platforms.
[0010] Preferably, the specific steps of the domain-adaptive feature transfer processing are as follows: Set the cross-platform unified behavioral feature vector as the source domain feature, and extract the shared features through a domain adaptive neural network; Similarity verification of shared features is performed by combining e-commerce product category feature database; Shared features that meet the similarity threshold are normalized to obtain e-commerce behavior feature vectors.
[0011] Preferably, the specific steps for constructing the structured initial interest profile of cold-start users are as follows: Calculate the category matching degree of each e-commerce category based on e-commerce behavior feature vectors; The interest level of each behavioral category is determined by combining the behavioral authenticity value; By integrating interest levels, category matching degree, and behavioral temporal features, a structured initial interest profile is obtained.
[0012] Preferably, the specific steps for matching the initial interest profile with e-commerce product features are as follows: Perform precise matching and association matching between the interest categories in the initial interest profile and the product categories and product attributes; Products that meet the preset matching threshold are selected to form a personalized candidate product set.
[0013] Preferably, the specific steps for generating the personalized recommendation list are as follows: The candidate product set is objectively ranked based on the matching degree between the product and the profile, and the freshness of the product. There is no subjective weight setting in the ranking process. Select a preset number of products based on the sorting results; The selected products are integrated to generate a personalized recommendation list for users starting from scratch.
[0014] Preferably, the specific steps for incorporating the behavioral authenticity value into the behavioral feature extraction process are as follows: The behavioral authenticity score is used as the criterion for determining the validity of behavioral features; Invalid behavioral features with a behavior authenticity value lower than a preset threshold are removed, and only valid features are retained for subsequent feature extraction.
[0015] Beneficial effects The present invention has the following beneficial effects: This method for cold-starting an e-commerce recommendation system based on user behavior sequences calculates the authenticity value of user behavior and combines federated learning to achieve cross-platform behavior transfer and multi-source data fusion. Under the premise of protecting user privacy, it constructs an initial interest profile for new users, solving the problems of low recommendation click-through rate and long user-system integration period caused by the lack of an effective initial interest profile in the cold start of e-commerce new user recommendations. It achieves a 45% increase in cold-start user click-through rate and significantly shortens the user-system integration period.
[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0017] Figure 1 This is a flowchart of a cold start method for an e-commerce recommendation system based on user behavior sequences, according to the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] This invention provides a technical solution: a cold start method for an e-commerce recommendation system based on user behavior sequences, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect behavioral feature datasets of users who are cold-starting users within the e-commerce platform and users associated across platforms, and perform preprocessing. The behavioral feature dataset includes effective dwell time, behavioral operation depth, behavioral reverse operation rate, cross-platform category collaborative matching degree, cross-platform behavioral effectiveness convergence value, behavioral timestamp, and behavioral object identifier. All of the above data are obtained through non-intrusive collection methods. Data within the e-commerce platform is collected in real time from the e-commerce platform's front-end behavioral logs, while cross-platform data is collected encrypted by each cross-platform local node through a dedicated data interface. The collection targets are newly registered / cold start users within the e-commerce platform with no effective consumption behavior and cross-platform users associated with the user's device. The collection process only acquires behavioral data and does not store the user's original identity information.
[0020] The collected behavioral feature dataset is subjected to abnormal behavior screening and invalid data is removed. Behaviors with an effective dwell time of less than 1 second, a frequency of the same type of behavior of more than 50 times within 1 hour, and no active device interaction are judged as abnormal behaviors and are directly removed. Unify behavior type encoding and timestamp format, perform unified digital encoding on behavior types of different platforms, convert all behavior timestamps to UTC time format, and complete data format standardization; Anonymous IDs are generated by performing irreversible hash operations on user identifiers. The SHA256 hash algorithm is used to calculate information such as user device feature codes and platform temporary identifiers to generate cross-platform unique and irreversible anonymous user IDs. These IDs replace the original user identifiers in the dataset, resulting in a standardized behavioral feature dataset.
[0021] Step 2: Based on the preprocessed behavioral feature dataset, perform objective and comprehensive calculations to obtain the user behavior authenticity value; All data in the behavioral feature dataset are subjected to dimensionless standardization. The min-max standardization algorithm is used to map the values of all feature data to the 0-1 range, eliminating the dimensional differences between different feature data. Then, by using basic mathematical operations such as summation, averaging, and multiplication, the standardized data is comprehensively calculated, taking into account core factors such as the effectiveness of the behavior and cross-platform collaboration, to obtain a user behavior authenticity value with a range of 0-1. The specific methods for obtaining the authenticity value of user behavior are as follows: In the formula: This represents the authenticity value of user behavior, ranging from 0 to 1. It is used to characterize the authenticity and validity of user behavior in the corresponding behavior category. The closer the value is to 1, the more accurately the behavior reflects the user's interests and preferences. The closer the value is to 0, the more invalid or random the behavior is. This represents the standardized effective dwell time of the behavior, with a value range of 0-1. It is the result of min-max standardization of the actual dwell time of the user on the behavior page under the corresponding category after excluding the loading time, and is used to characterize the actual attention time of the user to the category content. This represents the standardized depth of behavioral operations, ranging from 0 to 1. It is the result of the number of user behavioral operation steps in the corresponding category after min-max standardization, and is used to characterize the degree of user's active operation on category content. This represents the standardized cross-platform category collaboration matching degree, with a value range of 0-1. It is the standardized result of the cross-platform category attention and the corresponding product category of the e-commerce platform after coding matching, and is used to characterize the consistency of interest orientation between cross-platform behavior and e-commerce behavior. This represents the standardized reverse operation rate, ranging from 0 to 1. It is the proportion of "click-close" non-stop operations by users in the corresponding category out of the total number of behaviors, used to characterize the proportion of invalid behaviors in the category.
[0022] Step 3: Integrate the behavior authenticity value into the behavior feature extraction process, and use the federated learning method to fuse multi-source cross-platform behavior features to obtain a unified cross-platform behavior feature vector; The validity of behavioral features is determined by the behavioral authenticity value. Invalid behavioral features with a behavioral authenticity value lower than a preset threshold are removed, and only valid features are retained for subsequent feature extraction. The preset effective threshold for behavioral authenticity value is 0.3. Behavioral features with a behavioral authenticity value < 0.3 are removed from the dataset, and only valid behavioral features with a behavioral authenticity value ≥ 0.3 are retained. The valid features are then associated and bound with their corresponding behavioral authenticity values.
[0023] Each platform extracts behavioral feature vectors containing behavioral authenticity values locally and only uploads the encrypted model gradient. Each cross-platform and e-commerce platform acts as a local node of federated learning. Based on the effective behavioral features bound to behavioral authenticity values, it uses the Word2Vec algorithm to extract 64-dimensional local behavioral feature vectors. At the same time, it calculates the model gradient of the local feature vectors based on the logistic regression model. After encrypting the model gradient using a homomorphic encryption algorithm, it is only uploaded to the e-commerce master node. The master node aggregates the gradients uploaded from each platform using the federated averaging algorithm and updates the global model parameters. The e-commerce master node performs a weighted arithmetic average calculation on the encrypted gradients to obtain the global model gradient and synchronously updates the global model parameters of the federated learning. Iterative training continues until the global model converges, resulting in a unified cross-platform behavioral feature vector. The process of local feature extraction, encrypted gradient uploading, and global aggregation and update is repeated. When the global model loss value fluctuates by less than 0.001, the model is considered to have converged. The e-commerce master node then generates a 64-dimensional unified cross-platform behavioral feature vector for each anonymous user.
[0024] Step 4: Perform domain-adaptive feature transfer processing on the cross-platform unified behavioral feature vector to obtain the e-commerce behavioral feature vector; The cross-platform unified behavior feature vector is set as the source domain feature. Shared features are extracted by the domain adaptive neural network. The 64-dimensional cross-platform unified behavior feature vector is input into the domain adaptive neural network (DANN). The network feature extractor extracts the shared feature vector. The boundary between the source domain and the target domain is blurred by the adversarial training of the domain discriminator. Similarity verification of shared features is performed by combining the e-commerce product category feature library. The 64-dimensional product category feature library of the e-commerce platform is set as the target domain feature. The cosine similarity algorithm is used to calculate the similarity between the shared feature vector and the corresponding category feature vector. The preset effective similarity threshold is 0.6. Shared features that meet the similarity threshold are normalized to obtain e-commerce behavior feature vectors. Invalid shared feature vectors with similarity <0.6 are removed. The valid shared feature vectors are normalized using the min-max normalization algorithm to map the vector values to the 0-1 range, thus obtaining e-commerce behavior feature vectors adapted to e-commerce recommendation scenarios.
[0025] Step 5: Based on e-commerce behavior feature vectors and behavior authenticity values, construct a structured initial interest profile for cold-start users; Based on the e-commerce behavior feature vector, the category matching degree of each e-commerce category is calculated. The cosine similarity algorithm is used to perform one-to-one matching calculation between the e-commerce behavior feature vector and the feature vector of each category in the e-commerce product category feature library to obtain the user category matching degree value of each e-commerce category. The specific method for obtaining the user category matching score is as follows: In the formula: This represents the user category matching score, ranging from 0 to 1. It is used to characterize the degree of matching between the user's e-commerce behavior characteristics and the corresponding e-commerce product category. The closer the value is to 1, the higher the user's interest and preference for that category. This represents the user's e-commerce behavior feature vector, a 64-dimensional feature vector with values ranging from 0 to 1 for each dimension. It is the e-commerce adapted feature after cross-platform behavioral features have undergone domain adaptive transfer. This represents the feature vector of e-commerce product categories. It is a 64-dimensional feature vector with each dimension taking values from 0 to 1. It represents the basic features of e-commerce product categories after standardization. Representing vectors with vector The dot product is used to represent the basic value of similarity between two vectors; Representing vectors The length of the mold, Representing vectors The modulus is used to normalize the dot product result.
[0026] The interest level of each behavioral category is determined by combining the behavioral authenticity value. A dual determination is made based on the user's behavioral authenticity value (Q) and user category matching value (Y). Core interests are defined as Q≥0.7 and Y≥0.7, potential interests are defined as 0.4≤Q<0.7 or 0.4≤Y<0.7, and weak interests are defined as Q<0.4 and Y<0.4. The interest level, category matching degree, and behavioral temporal features are integrated to obtain a structured initial interest profile. The temporal compactness of the behavioral sequence of each behavioral category is extracted. Categories with a temporal compactness ≥0.5 are marked as high-concentration interest categories. Anonymous user ID is used as the unique identifier. The interest level, behavioral authenticity value, category matching degree, and temporal concentration label of each e-commerce category are integrated and stored in a structured manner according to the core interest, potential interest, and weak interest levels to form a structured initial interest profile of cold start users.
[0027] Step 6: Based on the matching relationship between the initial interest profile and the characteristics of e-commerce products, generate a personalized recommendation list for cold start users; The interest categories in the initial interest profile are precisely matched and associated with product categories and product attributes. Precise matching means that the core interest categories and the core attributes of the product are completely matched, while associated matching means that the potential interest categories and the relevant attributes of the product are matched. Products that meet the preset matching threshold are selected to form a personalized candidate product set. The preset matching threshold is 0.5. Only products with a matching degree ≥ 0.5 are included in the candidate product set. The number of products in the candidate product set is controlled in the range of 200-500.
[0028] The candidate product set is objectively ranked based on the matching degree between the product and the profile and the freshness of the product. There is no subjective weight setting in the ranking process. First, the freshness value of the products in the candidate product set is calculated, and then the product and the category matching degree value between the product and the profile are multiplied to obtain the comprehensive matching value. The products are ranked from high to low according to the comprehensive matching value. The specific method for obtaining the product freshness value is as follows: In the formula: This represents the freshness value of the product, ranging from 0 to 1. It is used to characterize the freshness of the product since it was put on the shelves. The closer the value is to 1, the newer the product is since it was put on the shelves. This represents the current UTC timestamp used for calculating product freshness, in days, which is the number of days converted from the UTC time of the calculation date. The timestamp is the UTC date and time of the product listing, expressed in days. It represents the number of days since the product was listed (UTC time). 30 indicates a fixed time calculation period, expressed in days, meaning that only the freshness of the product is considered within 30 days of its listing.
[0029] The specific method for obtaining the overall product matching value is as follows: In the formula: This represents the overall matching value of the product, ranging from 0 to 1. It is used to characterize the overall matching and freshness of the product with the user's interests. The closer the value is to 1, the higher the recommendation priority. This represents the category matching score between the product and the user profile, ranging from 0 to 1. It is the user category matching score calculated in step five, representing the degree of matching between the product and the user's interests. This represents the freshness value of the product, ranging from 0 to 1. It is the freshness value of the product calculated above, which indicates the freshness of the product at the time it was put on the shelf.
[0030] Select a preset number of products based on the sorting results. Then, select the top N products with the highest comprehensive matching values according to the front-end recommendation display requirements of the e-commerce platform. N is set to 20 recommendation slots on the APP homepage and N is set to 10 recommendation slots on the product details page. At the same time, control the proportion of interest levels: 60% for core interest products, 30% for potential interest products, and 10% for weak interest expansion products. Integrate the selected products to generate a personalized recommendation list for cold start users. Organize the selected products according to the comprehensive matching value sorting results to generate a personalized recommendation list containing product ID, product category, comprehensive matching value, and matching interest level. Directly connect to the e-commerce platform's front-end recommendation interface to realize personalized recommendation display during the cold start phase.
[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0032] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A cold start method for an e-commerce recommendation system based on user behavior sequences, characterized in that, Includes the following steps: Step 1: Collect behavioral feature datasets of users who are cold-starting users within the e-commerce platform and users associated across platforms, and perform preprocessing. Step 2: Based on the preprocessed behavioral feature dataset, perform objective and comprehensive calculations to obtain the user behavior authenticity value; Step 3: Integrate the behavior authenticity value into the behavior feature extraction process, and use the federated learning method to fuse multi-source cross-platform behavior features to obtain a unified cross-platform behavior feature vector; Step 4: Perform domain-adaptive feature transfer processing on the cross-platform unified behavioral feature vector to obtain the e-commerce behavioral feature vector; Step 5: Based on e-commerce behavior feature vectors and behavior authenticity values, construct a structured initial interest profile for cold-start users; Step 6: Based on the matching relationship between the initial interest profile and the characteristics of e-commerce products, generate a personalized recommendation list for users in the cold start phase.
2. The method according to claim 1, characterized in that, The behavioral feature dataset includes effective dwell time, behavioral operation depth, behavioral reverse operation rate, cross-platform category collaborative matching degree, cross-platform behavioral effectiveness convergence value, behavioral timestamp, and behavioral object identifier.
3. The method according to claim 1, characterized in that, The specific steps for obtaining the authenticity value of user behavior are as follows: Perform dimensionless standardization on all data in the behavioral feature dataset; Perform comprehensive calculations on the standardized data; The user behavior authenticity value is obtained with a range of 0-1.
4. The method according to claim 1, characterized in that, The specific steps of the standardized preprocessing are as follows: Filter out abnormal behaviors and remove invalid data from the behavioral feature dataset; Unify behavior type coding and timestamp format to achieve data format standardization; Anonymous IDs are generated by performing an irreversible hash operation on the user identifier, resulting in a standardized behavioral feature dataset.
5. The method according to claim 1, characterized in that, The specific steps for fusing multi-source cross-platform behavioral features using the federated learning method are as follows: Each platform extracts behavioral feature vectors containing behavioral realism values locally and only uploads the encrypted model gradients; The master node aggregates the gradients uploaded from each platform using a federated averaging algorithm and updates the global model parameters; it iterates the training until the global model converges, and obtains a unified behavioral feature vector across platforms.
6. The method according to claim 1, characterized in that, The specific steps of the domain-adaptive feature transfer processing are as follows: Set the cross-platform unified behavioral feature vector as the source domain feature, and extract the shared features through a domain adaptive neural network; Similarity verification of shared features is performed by combining e-commerce product category feature database; Shared features that meet the similarity threshold are normalized to obtain e-commerce behavior feature vectors.
7. The method according to claim 1, characterized in that, The specific steps for constructing a structured initial interest profile for cold-start users are as follows: Calculate the category matching degree of each e-commerce category based on e-commerce behavior feature vectors; The interest level of each behavioral category is determined by combining the behavioral authenticity value; By integrating interest levels, category matching degree, and behavioral temporal features, a structured initial interest profile is obtained.
8. The method according to claim 1, characterized in that, The specific steps for matching the initial interest profile with e-commerce product features are as follows: Perform precise matching and association matching between the interest categories in the initial interest profile and the product categories and product attributes; Products that meet the preset matching threshold are selected to form a personalized candidate product set.
9. The method according to claim 8, characterized in that, The specific steps for generating a personalized recommendation list are as follows: The candidate product set is objectively ranked based on the matching degree between the product and the profile, and the freshness of the product. There is no subjective weight setting in the ranking process. Select a preset number of products based on the sorting results; The selected products are integrated to generate a personalized recommendation list for users starting from scratch.
10. The method according to claim 1, characterized in that, The specific steps for incorporating behavioral authenticity values into the behavioral feature extraction process are as follows: The behavioral authenticity score is used as the criterion for determining the validity of behavioral features; Invalid behavioral features with a behavior authenticity value lower than a preset threshold are removed, and only valid features are retained for subsequent feature extraction.