A commodity intelligent recommendation method and system based on user behavior analysis

Abstract

The application discloses a kind of based on user behavior analysis's commodity intelligent recommendation method and system, it is related to commodity recommendation technical field.A kind of based on user behavior analysis's commodity intelligent recommendation system, including have: behavior analysis module and commodity recommendation module.The application can accurately identify the behavior type of user by obtaining the historical commodity browsing data of user and carrying out classification screening, to better understand the interest and purchase intention of user, the classification and analysis of this data can effectively support personalized recommendation, ensure that recommended content and user demand are more matched;Based on the comparative analysis of user purchase intention feature and browsing intention feature, the potential purchase intention and browsing preference of user can be accurately identified, to optimize commodity recommendation content.

Description

Technical Field

[0001] The present invention relates to the technical field of product recommendation, and in particular to a method and system for intelligent product recommendation based on user behavior analysis. Background Art

[0002] Among existing product recommendation technologies, e-commerce platforms typically use collaborative filtering or content-based methods to predict users' interests and preferences by analyzing their clicks, add-to-cart behavior, and other behavioral sequences, and recommend relevant products. However, these methods have obvious limitations: First, they are inadequate in distinguishing between users' deep browsing intentions and accidental behaviors such as invalid jumps, resulting in biased judgments about users' true interests. Second, existing technologies typically treat product recommendations and discount strategies as two independent systems, failing to incorporate induced discount strategies into the recommendation calculations of user interests, and lacking effective means to facilitate final transactions in the recommendation algorithm.

[0003] To solve the above technical problems, the present invention conducts parallel modeling and joint learning of the evolution of users' content interests and discount sensitivity. While accurately predicting user interests, it actively learns and executes the optimal conversion induction plan, significantly improving the conversion efficiency from intention to transaction. Summary of the Invention

[0004] The present invention aims to provide a method and system for intelligent product recommendation based on user behavior analysis, which can accurately predict user interests while actively learning and executing the optimal conversion induction plan, significantly improving the conversion efficiency from intention to transaction.

[0005] A method for intelligent product recommendation based on user behavior analysis includes the following steps:

[0006] Obtain several items of historical product browsing data from the user in the previous cycle; classify and filter the product browsing data to obtain transaction conversion data, purchase intention data, browsing loss data, and invalid jump data;

[0007] Analyze transaction conversion data to extract the corresponding preferential treatment dataset for each transaction conversion data item; analyze user behavior based on the preferential treatment dataset to obtain user purchase intention characteristics; and conduct comparative analysis based on the intended add-to-cart data and browsing loss data to obtain user browsing intention characteristics.

[0008] A user purchase interest evolution model is established based on user purchase intention characteristics and user browsing intention characteristics; products are recommended to users within the prediction period based on the user purchase interest evolution model.

[0009] As a preferred technical solution of the present invention, the specific steps of classifying and screening the product browsing data include:

[0010] For each item browsing data, obtain the corresponding raw behavior data and user interaction characteristics; raw behavior data includes but is not limited to user-item identification, browsing timestamp, and subsequent behavior type; user interaction characteristics include but are not limited to product dwell time and page scroll depth;

[0011] Based on the subsequent behavior type, the corresponding product browsing data is directly divided into transaction conversion data, purchase intention data, and remaining product browsing data;

[0012] For the remaining product browsing data, the semantic relevance scores between the user's recently browsed products are obtained. At the same time, a quantitative score is performed based on the product dwell time and page scroll depth to obtain the browsing effectiveness score. The semantic relevance score and the browsing effectiveness score are weighted to obtain the interaction depth score.

[0013] If the interaction depth score is lower than the set browsing threshold, the remaining product browsing data will be classified as invalid jump data; otherwise, the remaining product browsing data will be classified as browsing loss data.

[0014] As a preferred technical solution of the present invention, the specific steps of performing user behavior analysis based on the induced discount dataset include:

[0015] From the transaction conversion data, associate each transaction conversion data item with one or more inducement discount data items received before the transaction is completed, thereby constructing an inducement discount data set; the inducement discount data set includes at least one offer acceptance data item and zero or more offer rejection data items; calculate the purchase inflection point feature vector in the inducement discount data set; the purchase inflection point feature vector specifically includes: the maximum discount amount difference between the offer acceptance data item and all the offer rejection data items; the time elapsed from the first receipt of the inducement discount data item to the final acceptance of the offer data item; and the total number of offer rejections before the offer data item is accepted;

[0016] The purchase inflection point feature vectors corresponding to all transaction conversion data are combined to obtain the purchase inflection point feature space; the purchase inflection point feature space is clustered using an unsupervised clustering algorithm to obtain several user intention type clusters; each user intention type cluster is semantically annotated to obtain the user purchase intention feature.

[0017] As a preferred technical solution of the present invention, the specific steps of performing comparative analysis based on the purchase intention data and the browsing loss data include:

[0018] Analyze the intended add-to-cart data to obtain the corresponding feature data of the added-to-cart items; analyze the browse loss data to obtain the corresponding feature data of the lost items; calculate the similarity between the feature data of the added-to-cart items and the feature data of the lost items, and combine the intended add-to-cart data and browse loss data with similarity greater than the item similarity threshold to obtain the dual behavior-item dataset;

[0019] Perform feature extraction on the dual behavior-product dataset to obtain a composite interaction feature vector;

[0020] A product importance analysis model is constructed using the composite interaction feature vector as input and the corresponding contribution data type as the target label;

[0021] All composite interaction feature vectors are input into the product importance analysis model for analysis, and composite interaction feature vectors with higher contribution are extracted; intention rule analysis is performed on all composite interaction feature vectors with higher contribution to obtain user browsing intention features.

[0022] As a preferred technical solution of the present invention, the specific steps of establishing a user purchase interest evolution model based on user purchase intention characteristics and user browsing intention characteristics include:

[0023] Perform differential calculation on user browsing intention features and user purchase intention features to obtain the purchase conversion gain sequence;

[0024] Sort each user's browsing intention features and user's purchase intention features in the order of browsing timestamps to obtain the user's purchase interest evolution sequence;

[0025] The user purchase interest evolution model includes a content interest evolution channel and a discount sensitivity evolution channel; the content interest evolution channel is trained based on the user purchase interest evolution sequence, where the output is the predicted target product set in the next cycle; the discount sensitivity evolution channel is trained based on the purchase conversion gain sequence, where the output is the predicted discount sensitivity state in the next cycle.

[0026] As a preferred technical solution of the present invention, the specific steps of recommending products to users within a prediction period based on the user's purchase interest evolution sequence include:

[0027] Obtain the user's behavior sequence within the prediction period; input the behavior sequence into the user's purchase interest evolution model for dual-channel analysis to obtain a list of recommended products and a matching personalized inducement discount strategy;

[0028] Recommend products to users using a list of recommended products and a matching personalized discount strategy.

[0029] An intelligent product recommendation system based on user behavior analysis, comprising:

[0030] The behavior analysis module includes a purchase behavior analysis unit and a browsing behavior analysis unit. The purchase behavior analysis unit is used to obtain a number of historical product browsing data of users in the previous cycle; the product browsing data is classified and screened to obtain transaction conversion data, intention to purchase data, browsing loss data, and invalid jump data; the browsing behavior analysis unit is used to analyze based on the transaction conversion data and extract the inducement discount data set corresponding to each transaction conversion data; user behavior analysis is performed based on the inducement discount data set to obtain user purchase intention characteristics; based on the intention to purchase data and browsing loss data, comparative analysis is performed to obtain user browsing intention characteristics;

[0031] The product recommendation module includes an intelligent recommendation unit, which is used to establish a user purchase interest evolution model based on user purchase intention characteristics and user browsing intention characteristics; and recommend products to users within the prediction period based on the user purchase interest evolution model.

[0032] The present invention has the following advantages:

[0033] 1. The present invention can accurately identify the user's behavior type by obtaining the user's historical product browsing data and classifying and screening it, thereby better understanding the user's interests and purchasing intentions. The classification and analysis of this data can effectively support personalized recommendations, ensuring that the recommended content is more closely aligned with user needs. Based on the comparative analysis of the user's purchase intention characteristics and browsing intention characteristics, the user's potential purchase intentions and browsing preferences can be accurately identified, thereby optimizing the product recommendation content. By establishing a user purchase interest evolution model and based on the predicted product set output by the model, the user can be provided with product recommendations that meet his or her interests, thereby improving the user's purchase conversion rate.

[0034] 2. The present invention can accurately capture the changing trends of user interests by differentially calculating the purchase conversion gain sequence and combining it with timestamp sorting. This dynamic modeling method enables the recommendation system to flexibly adapt to changes in user interests in different time periods, better predict the products that users may be interested in in future cycles, and improve the timeliness and accuracy of recommendations; by using an unsupervised clustering algorithm to cluster the purchase inflection point feature vectors, it can better identify the purchasing behavior characteristics of different user groups and generate a personalized recommendation list for each user group; through cluster analysis, it can effectively discover potential user groups and their corresponding purchase intentions, and further optimize the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic diagram of the structure of an intelligent product recommendation system based on user behavior analysis adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order for those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0037] Embodiment 1, a commodity intelligent recommendation method based on user behavior analysis, comprising the following steps:

[0038] Obtaining a plurality of commodity browsing data of a user in a previous period; classifying and screening the commodity browsing data to obtain transaction conversion data, intention purchase data, browsing loss data and invalid jump data;

[0039] The transaction conversion data refers to the commodity browsing behavior data of the user finally completing the purchase. This kind of data contains the record that the user finally achieves the transaction through purchase or other conversion behaviors (such as placing an order, payment, etc.) after browsing a certain commodity. The feature is usually directly related to the subsequent behavior type (such as purchase), which means that the user makes a purchase decision after browsing the commodity. The intention purchase data refers to the explicit purchase intention of the user when browsing the commodity, but the transaction has not been completed. It is usually manifested as adding the commodity to the shopping cart or the collection, etc. This indicates that the user has a certain interest in the commodity and may complete the purchase at some future time. The feature is that the user may add the commodity to the shopping cart or mark the commodity as like, attention, etc. after browsing the commodity, but does not immediately complete the purchase.

[0040] The browsing loss data refers to the user's behavior data of not continuing other interactions after browsing a certain commodity, such as not adding the commodity to the shopping cart, not performing other in-depth browsing, etc., and finally failing to complete the purchase. This kind of data reflects that the user's interest in the commodity cannot be maintained or converted into a purchase behavior. The user may have lost interest during the browsing process, resulting in loss. The feature is that the user's browsing behavior is not further deepened (such as short commodity stay time, low page scrolling depth, etc.), which finally leads to no conversion. The invalid jump data refers to the behavior data of the user quickly leaving or switching to other pages after browsing the commodity. These users hardly have any interaction or in-depth browsing when browsing the commodity, which shows that they have no substantive interest in the commodity. This kind of data usually contains very short stay time and very low page scrolling depth. The user hardly interacts with the page or stays on the commodity page, which belongs to invalid commodity browsing data.

[0041] The specific steps of classifying and screening the commodity browsing data include:

[0042] For each piece of commodity browsing data, the corresponding original behavior data and user interaction features are obtained. The original behavior data includes but is not limited to user-commodity identification, browsing timestamp and subsequent behavior type. The user interaction features include but are not limited to commodity stay time and page scrolling depth.

[0043] In the process of commodity browsing data classification screening, the original behavior data and user interaction features are the comprehensive records of the user's interaction process with the commodity on the platform, which help to analyze the user's interest in the commodity, behavior pattern, and whether to generate purchase or other conversion behavior.

[0044] User-commodity identification is used to uniquely identify the interaction relationship between a user and a commodity. Each piece of commodity browsing data is bound to a specific user and a specific commodity. Usually, a combination of user ID and commodity ID is used to ensure that the behavior of a specific user when browsing a specific commodity can be uniquely identified.

[0045] The browsing timestamp records the specific time point when the user starts browsing the commodity. The timestamp provides a time reference for subsequent analysis, helping to analyze whether the user's behavior is concentrated in certain time periods. The timestamp is usually in milliseconds or seconds, representing the number of seconds since a certain starting time.

[0046] The subsequent behavior type records the next step behavior of the user after browsing the commodity. These behaviors can include browsing ending (i.e., the user leaves directly after browsing a certain commodity), adding to cart (i.e., the user adds the commodity to the shopping cart), purchasing (i.e., the user completes the purchase), etc.

[0047] User interaction features reflect the depth of user interaction when browsing the commodity, helping to evaluate the intensity of user interest in the commodity. These features help to understand the potential motivation of user behavior. Commodity stay duration indicates the actual time the user stays on a certain commodity page when browsing. This feature can reveal the user's interest in the commodity. Longer stay time usually indicates that the user has a higher interest in the commodity. Stay duration is usually recorded in seconds. For example, 120 indicates that the user stayed on the commodity page for 120 seconds. Page scroll depth records the user's scrolling on the commodity page, reflecting whether the user has thoroughly browsed other parts of the commodity page. Greater scroll depth may indicate that the user has a more comprehensive understanding of the commodity. Page scroll depth is usually a percentage, indicating the progress of the user's page scrolling. For example, 70% indicates that the user has browsed 70% of the page depth. This value can be calculated by tracking the action of the page scrollbar or derived by analyzing the user's page interaction data.

[0048] According to the subsequent behavior type, the corresponding commodity browsing data is directly divided into transaction conversion data, intended add-to-cart data, and remaining commodity browsing data.

[0049] Check the subsequent behavior type of each product browsing data; if the subsequent behavior type is a purchase or similar transaction behavior, the data is directly classified as transaction conversion data; if the subsequent behavior type is an intentional behavior such as adding to a shopping cart or adding to favorites, the data is classified as intentional add-to-cart data; for other browsing data that does not immediately generate a purchase or add-to-cart behavior, that is, the subsequent behavior type is browsing or jumping without further interaction, it is classified as remaining product browsing data. These remaining data will be further analyzed as invalid jump data or browsing loss data based on the interaction depth score.

[0050] For the remaining product browsing data, the semantic relevance scores between the user's recently browsed products are obtained. At the same time, a quantitative score is performed based on the product dwell time and page scroll depth to obtain the browsing effectiveness score. The semantic relevance score and the browsing effectiveness score are weighted to obtain the interaction depth score.

[0051] The semantic relevance score aims to measure the similarity or correlation between items recently viewed by a user to reflect the user's interest in related items. The specific steps are: obtain a list of items recently viewed by the user, which can be identified by item ID or item attributes; use natural language processing (NLP) technology to analyze the descriptions, titles, category tags, and other content of these items to calculate the semantic similarity between each two items; commonly used similarity calculation methods include cosine similarity, Jaccard similarity, or similarity calculation based on word vectors (such as Word2Vec or BERT). Based on the above similarity calculation methods, a semantic relevance score is obtained between a group of items. The higher the score, the stronger the semantic correlation between the items, and the user's interest is concentrated on these related items.

[0052] The browsing effectiveness score evaluates the depth and quality of browsing based on the user's stay time and page scrolling depth on the product page; the specific steps are: obtain the stay time and page scrolling depth of each product browsing data; for the stay time, a reasonable threshold can be set, and browsing below this threshold may be considered incomplete browsing (for example, less than 10 seconds is considered invalid browsing, and the specific threshold data is set manually), and above the threshold, it indicates that the user has a high interest in the product; the page scrolling depth indicates the user's attention to the product page. Users who scroll more than a certain proportion indicate that they have a strong interest in the product content. Combining these two factors, they are converted into a unified browsing effectiveness score using the weighted average method. The weights of the stay time and page scrolling depth can be adjusted manually according to actual conditions.

[0053] The interaction depth score is obtained by combining the semantic correlation score and the browsing effectiveness score to comprehensively evaluate the interaction depth of the user; the specific steps are: normalizing the semantic correlation score and the browsing effectiveness score to ensure that the score ranges of the two are consistent; assigning weight values to the two scores, and the weight values can be set according to business requirements, for example, the semantic correlation score weight is 0.6, and the browsing effectiveness score weight is 0.4; the interaction depth score is calculated according to the weighted average method, that is, the formula: interaction depth score = (semantic correlation score x 0.6) + (browsing effectiveness score x 0.4) is used to calculate, and the specific weight value is adjusted by a person.

[0054] If the interaction depth score is lower than the set browsing threshold, the remaining commodity browsing data is divided into invalid jump data; otherwise, the remaining commodity browsing data is divided into browsing loss data; the specific value of the set browsing threshold is set by a person; the interaction depth score of each piece of remaining commodity browsing data is calculated, which combines the semantic correlation score and the browsing effectiveness score; compared with the set browsing threshold; if the interaction depth score is lower than the set browsing threshold, it indicates that the user's interest and participation in the commodity are low, and the data is divided into invalid jump data, indicating that the user does not perform effective browsing interaction and quickly leaves the page; if the interaction depth score is higher than or equal to the set browsing threshold, it indicates that the user has strong interest in the commodity, but fails to finally convert into a purchase behavior or an add-to-cart behavior, so the data is divided into browsing loss data, indicating that the user has certain interest in the commodity, but does not continue subsequent operations, resulting in loss.

[0055] Based on the transaction conversion data, an induced preferential data set corresponding to each piece of transaction conversion data is extracted; based on the induced preferential data set, user behavior analysis is performed to obtain user purchase intention features; based on the intention add-to-cart data and the browsing loss data, comparative analysis is performed to obtain user browsing intention features;

[0056] The specific steps of the user behavior analysis based on the induced preferential data set include:

[0057] From the transaction conversion data, associate each transaction conversion data with one or more inducement discount data received before the transaction, so as to construct an inducement discount data set; the inducement discount data set contains at least one acceptance discount data and zero or more rejection discount data; calculate the purchase inflection point feature vector in the inducement discount data set; the purchase inflection point feature vector specifically includes: the maximum discount amount difference between the acceptance discount data and all the rejection discount data; the time elapsed from the first receipt of the inducement discount data to the final acceptance of the discount data; the total number of rejections of the discount data before the acceptance of the inducement discount data; from the transaction conversion data, find one or more inducement discount data received before the transaction for each transaction conversion data; each transaction conversion data may be associated with multiple inducement discount data; through this step, construct an inducement discount data set, in which each record contains at least one acceptance discount data and zero or more rejection discount data, these discount data record the discount information received by the user before the transaction and his response; the acceptance discount data is the discount finally accepted by the user, and the rejection discount data records the discount that the user did not select.

[0058] Induced discount data refers to the different types of discount information provided by the platform to users before users complete the transaction conversion, as well as the users' response to such discounts; specifically, each piece of induced discount data records a certain discount received by the user and his / her selection behavior before the transaction.

[0059] For example, suppose a user receives two discount messages during the shopping process. The first is a discount of 10 yuan off for purchases over 100 yuan, and the second is a discount of 50 yuan off for purchases over 200 yuan. The user ultimately accepts the second discount, that is, the discount of 50 yuan off for purchases over 200 yuan, and rejects the first discount. In this example, the induced discount dataset will contain two records: one for the accepted discount (50 yuan off for purchases over 200 yuan) and the other for the rejected discount (10 yuan off for purchases over 100 yuan). These induced discount data help analyze how users make choices during the shopping process and which discounts can more effectively encourage users to complete purchases.

[0060] The purchase inflection point feature vector is calculated in the induced discount data set. The purchase inflection point feature vector is an indicator that measures the key behavioral changes of users in the process of accepting induced discount data; the difference between the maximum discount amount in the accepted discount data and all the rejected discount data is calculated. This difference can help identify which discounts have a greater influence on prompting users to make purchase decisions; the time from the first time the user receives the induced discount data to the final acceptance of the discount data is calculated, reflecting the time span for the user to make a choice between multiple discounts; the number of times the user has rejected the induced discount data before accepting the induced discount data is calculated, and the user's selection process and decision-making pattern for accepting the discount are evaluated. These features constitute the purchase inflection point feature vector, which is a key quantitative indicator of user behavior.

[0061] For example, when calculating the purchase inflection point eigenvector in an induced discount dataset, the difference between the maximum discount amount in the accepted discount data and the maximum discount amount in all rejected discount data is calculated. Suppose the two discounts received by the user are 10 yuan off for purchases over 100 yuan and 50 yuan off for purchases over 200 yuan, and the user finally accepts the 50 yuan off for purchases over 200 yuan, but refuses the 10 yuan off for purchases over 100 yuan. Then, the maximum discount amount difference is 50 yuan (the difference between the discount of 50 yuan off for purchases over 200 yuan and the discount of 10 yuan off for purchases over 100 yuan); calculate the time from the first time the user receives the induced discount data to the final acceptance of the discount data; the user first received the discount of 10 yuan off for purchases over 100 yuan on xx / xx / xx, and did not accept the discount of 50 yuan off for purchases over 200 yuan until xx / xx / xx, so the time span is 4 days; calculate how many times the user rejected the discount data before accepting the final discount. If the user rejected 3 different discount data during the decision-making process, the total number of times the discount data was rejected is 3 times; through these steps, the purchase inflection point feature vector will contain the maximum discount amount difference (50 yuan), time span (4 days) and number of discount rejections (3 times). These features help evaluate the behavioral patterns of users in making purchase decisions during the process of accepting discounts, and further analyze which factors have a significant impact on their purchase decisions.

[0062] Combine the purchase inflection point feature vectors corresponding to all transaction conversion data to obtain the purchase inflection point feature space; cluster the purchase inflection point feature space using an unsupervised clustering algorithm to obtain several user intention type clusters; semantically annotate each user intention type cluster to obtain the user purchase intention feature;

[0063] The purchase inflection point feature vectors corresponding to all transaction conversion data for the same user on different products are combined to form a purchase inflection point feature space. This feature space reflects the user's behavioral characteristics towards different products when accepting induced discount data. To better understand the user's behavioral patterns and purchase intentions, the feature space can be clustered using an unsupervised clustering algorithm. Common clustering algorithms include K-means or DBSCAN. In this case, clustering will divide the behavior patterns of the same user on multiple products into several different intention clusters, each cluster representing the user's specific behavioral characteristics for a certain type of product in the purchase decision. Each user's intention cluster for different products is semantically labeled. By analyzing the common behaviors and characteristics of users in the cluster under similar product categories, combined with the actual product characteristics and user purchase intentions, each cluster is given a descriptive label to represent the user's purchase intention.

[0064] For example, some clusters may represent discount-sensitive users, who are more likely to make purchasing decisions when faced with high discounts or substantial discounts, especially for higher-priced products; while other clusters may represent low-price-seeking users, who are more inclined to accept certain discounts when faced with lower-priced products with moderate discounts. In this way, the user's purchase intention characteristics are ultimately obtained, and combined with the characteristics of the product, data support is provided for subsequent accurate product recommendations, personalized recommendations, and user behavior predictions, helping the platform understand the user's purchasing preferences and decision-making process.

[0065] The specific steps for conducting a comparative analysis based on the data on purchase intent and browsing loss include:

[0066] Analyze the data on intention to add to cart items to obtain corresponding feature data for added items; analyze the data on lost browsing activity to obtain corresponding feature data for lost items; calculate the similarity between the feature data for added items and the feature data for lost items, and combine the intention to add to cart data and lost browsing activity data with a similarity greater than a product similarity threshold to obtain a paired behavior-product dataset; the product similarity threshold is manually set.

[0067] The two types of data are analyzed separately, and the corresponding product feature data is extracted. For the intention to add data, attention is paid to the product features of the product that the user adds to the shopping cart but does not complete the purchase; for example, the type, price, discount rate, brand, evaluation, etc. of the product can be used as the feature data of the added product; at the same time, for the browsing loss data, the product features of the user after browsing the product page and not continuing to interact or not generating purchase are analyzed; the features of the browsing loss product can include the attractiveness of the product (such as image quality, page loading speed), user dwell time, page scroll depth, etc. By analyzing the two types of data, add product feature data and loss product feature data are extracted respectively; the similarity between the add product feature data and the loss product feature data is calculated, and a similarity measurement method (such as cosine similarity or Euclidean distance) is used. The add data and loss data with a similarity greater than a preset product similarity threshold are combined to form a pair of behavior-product data sets.

[0068] Feature extraction is performed on the pair of behavior-product data sets to obtain a composite interaction feature vector.

[0069] The composite interaction feature vector is composed of three parts; first, the behavior signature vector, which quantifies the user's micro-interaction behavior when browsing the product, usually including information exploration depth (such as browsing page depth or dwell time), visual confirmation intensity (such as page scrolling or zooming in on the number of times), social identity seeking degree (such as viewing comments, sharing, or referring to others' ratings); second, the product static feature vector, which directly includes the inherent feature data of the product, such as product category, price, brand, discount, evaluation, etc. information; and finally, the behavior-product interaction feature, which is generated by cross-operating the dimensions in the behavior signature vector and the dimensions in the product static feature vector, describing the interaction feature of a specific behavior occurring in a specific product context. Through these features, a complete set of composite interaction feature vectors is formed.

[0070] Using the composite interaction feature vector as input and the corresponding contribution data type as target label, a product importance analysis model is constructed.

[0071] All composite interaction feature vectors are input into the product importance analysis model for analysis, and composite interaction feature vectors with high contribution are extracted; the intention rule analysis is performed on all composite interaction feature vectors with high contribution, and the user browsing intention features are obtained.

[0072] In the product importance analysis model, contribution represents the influence of each composite interaction feature vector on the prediction of the target label (such as add-to-cart, purchase, or churn). A composite interaction feature vector with a higher contribution indicates a closer relationship between the feature and the user behavior (such as add-to-cart or churn), meaning that the feature plays a greater role in explaining the user's behavioral decision-making process. Contribution is typically measured using a weight or importance score in the model, indicating the contribution of a specific feature to the prediction result. For example, feature selection methods (such as decision trees, random forests, and Lasso regression) are used to calculate the contribution of each composite interaction feature vector to the prediction of the target behavior. Feature vectors with higher contribution have greater influence in the model and can help us identify which features or behavioral patterns are most relevant to the user's ultimate decision to add to cart, purchase, or churn.

[0073] The target label is the output data of model learning, representing the user's behavioral outcome or target behavior. In product importance analysis, the target label is directly related to the user's actual behavior (such as add-to-cart, purchase, and churn). The definition of the target label depends on the type of behavior being analyzed. For example, setting the corresponding contribution data type to the add-to-cart label indicates whether the user added the item to the shopping cart after browsing. A label of 1 indicates that the user added the item to the cart, and 0 indicates that the user did not add the item to the cart.

[0074] Suppose we want to analyze a user's browsing intent when browsing two identical items of clothing. We input the composite interaction feature vector into an item importance analysis model and analyze user behavior based on different target labels (add-to-cart or churn). For example, consider two items of clothing: a red and a black. For each item, user browsing behavior may differ. For example, a user may show a strong intention to add to the cart for the red item, but may exhibit churn behavior for the black item. By analyzing user behavior through the model, we can conclude the following: For the red item, the contribution between the user's composite interaction feature vector (e.g., long dwell time, deep page scrolling, frequent review viewing, etc.) and the add-to-cart label is high, indicating that the user is highly interested in the red item and ultimately decides to add it to their cart. For the black item, although the user's behavioral characteristics are similar (e.g., long dwell time), the contribution to the churn label is higher. This may be because, despite browsing the black item for a considerable period of time, due to certain factors (e.g., color preference, low discount, etc.), the user ultimately chooses to leave the item rather than add it to the cart. From these analyses, we can infer user browsing intent: users prefer red clothing styles because they are more strongly associated with add-to-cart purchases, while black styles are more likely to lead to user churn, potentially due to color preferences or other shopping decision-making factors (such as discounts and recommendations). By combining this approach with analysis of composite interaction feature vectors and target labels, the product importance analysis model can help us identify user preferences and optimize product display, personalized recommendations, and marketing strategies. For example, the platform can provide personalized recommendations for different clothing styles based on user browsing and add-to-cart behavior, further improving conversion rates.

[0075] A user purchase interest evolution model is established based on user purchase intention characteristics and user browsing intention characteristics; products are recommended to users within the prediction period based on the user purchase interest evolution model.

[0076] The specific steps for establishing a user purchase interest evolution model based on user purchase intention characteristics and user browsing intention characteristics include:

[0077] Perform differential calculation on user browsing intention features and user purchase intention features to obtain the purchase conversion gain sequence;

[0078] Sort each user's browsing intention features and user's purchase intention features in the order of browsing timestamps to obtain the user's purchase interest evolution sequence;

[0079] The user purchase interest evolution model includes a content interest evolution channel and a discount sensitivity evolution channel; the content interest evolution channel is trained based on the user purchase interest evolution sequence, where the output is the predicted target product set in the next cycle; the discount sensitivity evolution channel is trained based on the purchase conversion gain sequence, where the output is the predicted discount sensitivity state in the next cycle.

[0080] When building a model of user purchasing interest evolution, we first perform a differential calculation on user browsing intent features and user purchase intent features to obtain a purchase conversion gain sequence. Browsing intent features reflect user interest in products while browsing, while purchase intent features reflect the user's intention to add to a cart or purchase. By differentially calculating these two types of features, we can measure the user's purchase conversion gain, that is, the degree to which a user's browsing interest in a product converts into purchase intent. For example, if a user has strong browsing interest in a product but does not express sufficient purchase intent during the subsequent conversion process, the purchase conversion gain value is low. Conversely, if the user expresses strong purchase intent during the browsing process, the gain value is high. The result of the differential calculation is a gain sequence, which helps us identify which products have a significant impact on user purchase conversions. Especially when using induced discount data, the purchase conversion gain sequence can further reflect the impact of discount strategies on conversions.

[0081] Each user's browsing and purchasing intention features are sorted by browsing timestamp to form a user purchase interest evolution sequence. This step aims to simulate the evolution of user interests and reflect the dynamic changes from initial browsing to final purchase intention. By sorting user behavior by browsing timestamp, we can better capture the evolution of user interests and intentions. For example, a user may have a strong browsing interest in a product during a certain period of time. Over time, this interest may gradually shift to purchase intention due to certain promotional activities or preferential policies. The resulting user purchase interest evolution sequence can be used to analyze the changing trends of user interests over time, thus providing a basis for subsequent personalized recommendations and marketing strategies.

[0082] The user purchasing interest evolution model includes a content interest evolution channel and a discount sensitivity evolution channel. The content interest evolution channel is trained primarily based on the user's purchasing interest evolution sequence, with the goal of predicting the set of products that the user may be interested in during the next cycle. By analyzing the evolution of a user's purchasing interests over past cycles, the model can learn which product types or features are likely to attract the user again in the future. For example, if a user has historically shown a strong interest in athletic shoes, the model can predict that the user is likely to continue to be interested in similar athletic shoes in future cycles; therefore, the content interest evolution channel can help the platform make more accurate product recommendations in future cycles.

[0083] The Offer Sensitivity Evolution Channel is trained based on purchase conversion gain sequences, aiming to predict a user's sensitivity to offers in the next cycle. By analyzing a user's purchase conversion gain sequences in past cycles, the model can identify which offers played a key role in their purchasing decisions. For example, some users may be very sensitive to deep discounts, while others may be less sensitive to offers during promotional events. Based on these gain sequences, the Offer Sensitivity Evolution Channel can predict user reactions to offers in future cycles, providing a basis for developing precise marketing strategies, such as adjusting the discount size, type of offer, or delivery method.

[0084] The constructed user purchasing interest evolution model can simultaneously focus on the dynamic changes in users' content preferences and discount sensitivity, helping the platform to more accurately predict users' product interests and purchasing intentions in future cycles, thereby optimizing recommendation algorithms and marketing strategies, and improving user conversion rates and loyalty.

[0085] The specific steps for recommending products to users within the prediction period based on the user's purchase interest evolution sequence include:

[0086] Obtain the user's behavior sequence within the prediction period; input the behavior sequence into the user's purchase interest evolution model for dual-channel analysis to obtain a list of recommended products and a matching personalized inducement discount strategy;

[0087] Recommend products to users using a list of recommended products and a matching personalized discount strategy.

[0088] When recommending products to users within a prediction period based on their purchase interest evolution sequence, the first step is to obtain the user's behavior sequence during the prediction period. This process involves collecting and organizing all user interaction data during the prediction period, typically including browsing history on product pages, add-to-cart behavior, purchase behavior, and interaction with promotional events. By tracking user interactions with products, behavior sequences provide detailed data on user interest changes and preferences, helping the model understand users' current needs and potential purchase intentions. These behavior sequences provide key information input for subsequent recommendations. User behavior sequences are then input into the user purchase interest evolution model for dual-channel analysis; the user purchase interest evolution model consists of two channels: the content interest evolution channel and the offer sensitivity evolution channel.

[0089] In the content interest evolution channel, the model analyzes user interest evolution trends and, based on past behavioral trajectories, predicts the types or specific products that users are most likely to be interested in during the forecast period. This process, based on users' historical browsing, add-to-cart, and purchase records, as well as changes in user preferences, helps the model determine which products are likely to pique users' interest and drive purchases. Meanwhile, in the offer sensitivity evolution channel, the model analyzes users' responses to promotional activities over past periods, such as discounts and purchase discounts, to predict which promotional strategies users are likely to be most sensitive to during the forecast period. The combined analysis of these two channels not only recommends products to users but also matches each recommended product with a corresponding personalized promotional strategy.

[0090] Utilizing the recommended product list derived from dual-channel analysis and the matching personalized incentive discount strategies, we provide users with precise product recommendations. When recommending products, we prioritize products based on the user's current interests and sensitivity to discounts, along with personalized discount strategies. This approach ensures that the recommended products not only align with the user's purchasing interests but also stimulate their purchasing desire through appropriate discount strategies, thereby increasing conversion rates.

[0091] For example, a brand of sneakers a user recently viewed might be recommended, along with exclusive discounts or coupons from that brand, to incentivize the user to complete the purchase. Through these steps, highly personalized product recommendations and incentive discount strategies can be provided to each user within the forecast period, improving the user experience while effectively boosting conversions and sales on the platform.

[0092] In this example, we assume that personalized product recommendations and promotional offers are being pushed to users of an e-commerce platform. During the prediction cycle, the user's behavior sequence is first obtained, including browsing history, add-to-cart behavior, and purchase behavior over the past 30 days. For example, a user browsed 100 items, added 20 items to their cart, and ultimately purchased 5 items over the past 30 days. These behavior sequences are then fed into a user purchase interest evolution model for dual-channel analysis. In the content interest evolution channel, the model analyzes the user's browsing and purchase trajectory, predicting that the user will have a high interest in "sports shoes" and "fitness equipment" in the future cycle and recommending 10 products that match these interests. In the promotion sensitivity evolution channel, the model identifies that the user is sensitive to promotions offering ¥50 off purchases over ¥200 and therefore matches this promotional offer to the recommended products. Finally, the recommendation system provides the user with a list of 10 recommended products, each with a personalized promotional offer, such as "¥50 off purchases over ¥200 for sports shoes," to stimulate the user's purchasing desire. During this process, the composite interaction feature vector helps analyze changes in user interests, and the purchase conversion gain sequence further optimizes the matching of products and offers, ensuring the accuracy and effectiveness of recommendations.

[0093] Example 2, a product intelligent recommendation system based on user behavior analysis, see Figure 1 Shown, including:

[0094] The behavior analysis module includes a purchase behavior analysis unit and a browsing behavior analysis unit. The purchase behavior analysis unit is used to obtain a number of historical product browsing data of users in the previous cycle; the product browsing data is classified and screened to obtain transaction conversion data, intention to purchase data, browsing loss data, and invalid jump data; the browsing behavior analysis unit is used to analyze based on the transaction conversion data and extract the inducement discount data set corresponding to each transaction conversion data; user behavior analysis is performed based on the inducement discount data set to obtain user purchase intention characteristics; based on the intention to purchase data and browsing loss data, comparative analysis is performed to obtain user browsing intention characteristics;

[0095] The product recommendation module includes an intelligent recommendation unit, which is used to establish a user purchase interest evolution model based on user purchase intention characteristics and user browsing intention characteristics; and recommend products to users within the prediction period based on the user purchase interest evolution model.

[0096] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. A method for intelligent product recommendation based on user behavior analysis, characterized in that: The following steps are involved: Obtain several items of historical product browsing data from the user in the previous cycle; classify and filter the product browsing data to obtain transaction conversion data, purchase intention data, browsing loss data, and invalid jump data; Analyze transaction conversion data to extract the corresponding preferential treatment dataset for each transaction conversion data; analyze user behavior based on the preferential treatment dataset to obtain user purchase intention characteristics; Compare and analyze the data on purchase intention and browsing loss to obtain the characteristics of user browsing intention; Establish a user purchase interest evolution model based on user purchase intention characteristics and user browsing intention characteristics; recommend products to users within the prediction period based on the user purchase interest evolution model; The specific steps for classifying and filtering product browsing data include: For each item browsing data, obtain the corresponding raw behavior data and user interaction characteristics; raw behavior data includes user-item identifier, browsing timestamp, and subsequent behavior type; user interaction characteristics include product dwell time and page scroll depth; Based on the subsequent behavior type, the corresponding product browsing data is directly divided into transaction conversion data, purchase intention data, and remaining product browsing data; For the remaining product browsing data, the semantic relevance scores between the user's recently browsed products are obtained. At the same time, a quantitative score is performed based on the product dwell time and page scroll depth to obtain the browsing effectiveness score. The semantic relevance score and the browsing effectiveness score are weighted to obtain the interaction depth score. If the interaction depth score is lower than the set browsing threshold, the remaining product browsing data will be classified as invalid jump data; otherwise, the remaining product browsing data will be classified as browsing loss data; The specific steps for analyzing user behavior based on the induced discount dataset include: From the transaction conversion data, associate each transaction conversion data item with one or more inducement discount data items received before the transaction is completed, thereby constructing an inducement discount data set; the inducement discount data set includes at least one offer acceptance data item and zero or more offer rejection data items; calculate the purchase inflection point feature vector in the inducement discount data set; the purchase inflection point feature vector specifically includes: the maximum discount amount difference between the offer acceptance data item and all the offer rejection data items; the time elapsed from the first receipt of the inducement discount data item to the final acceptance of the offer data item; and the total number of offer rejections before the offer data item is accepted; Combine the purchase inflection point feature vectors corresponding to all transaction conversion data to obtain the purchase inflection point feature space; cluster the purchase inflection point feature space using an unsupervised clustering algorithm to obtain several user intention type clusters; semantically annotate each user intention type cluster to obtain the user purchase intention feature; The specific steps for conducting a comparative analysis based on the data on purchase intent and browsing loss include: Analyze the intended add-to-cart data to obtain the corresponding feature data of the added-to-cart items; analyze the browse loss data to obtain the corresponding feature data of the lost items; calculate the similarity between the feature data of the added-to-cart items and the feature data of the lost items, and combine the intended add-to-cart data and browse loss data with similarity greater than the item similarity threshold to obtain the dual behavior-item dataset; Perform feature extraction on the dual behavior-product dataset to obtain a composite interaction feature vector; A product importance analysis model is constructed using the composite interaction feature vector as input and the corresponding contribution data type as the target label; All composite interaction feature vectors are input into the product importance analysis model for analysis, and composite interaction feature vectors with higher contribution are extracted. Intention rule analysis is performed on all composite interaction feature vectors with higher contribution to obtain user browsing intention features. The specific steps for establishing a user purchase interest evolution model based on user purchase intention characteristics and user browsing intention characteristics include: Perform differential calculation on user browsing intention features and user purchase intention features to obtain the purchase conversion gain sequence; Sort each user's browsing intention features and user's purchase intention features in the order of browsing timestamps to obtain the user's purchase interest evolution sequence; The user purchase interest evolution model includes a content interest evolution channel and a discount sensitivity evolution channel; the content interest evolution channel is trained based on the user purchase interest evolution sequence, where the output is the predicted target product set in the next cycle; the discount sensitivity evolution channel is trained based on the purchase conversion gain sequence, where the output is the predicted discount sensitivity state in the next cycle.

2. The method for intelligent product recommendation based on user behavior analysis according to claim 1, characterized in that: The specific steps for recommending products to users within the prediction period based on the user's purchase interest evolution sequence include: Obtain the user's behavior sequence within the prediction period; input the behavior sequence into the user's purchase interest evolution model for dual-channel analysis to obtain a list of recommended products and a matching personalized inducement discount strategy; Recommend products to users using a list of recommended products and a matching personalized discount strategy.

3. A product intelligent recommendation system based on user behavior analysis, characterized in that: The system applies the method for intelligent product recommendation based on user behavior analysis according to any one of claims 1 to 2, comprising: The behavior analysis module includes a purchase behavior analysis unit and a browsing behavior analysis unit. The purchase behavior analysis unit is used to obtain a number of historical product browsing data of users in the previous cycle; the product browsing data is classified and screened to obtain transaction conversion data, intention to purchase data, browsing loss data, and invalid jump data; the browsing behavior analysis unit is used to analyze based on the transaction conversion data and extract the inducement discount data set corresponding to each transaction conversion data; user behavior analysis is performed based on the inducement discount data set to obtain user purchase intention characteristics; based on the intention to purchase data and browsing loss data, comparative analysis is performed to obtain user browsing intention characteristics; The product recommendation module includes an intelligent recommendation unit, which is used to establish a user purchase interest evolution model based on user purchase intention characteristics and user browsing intention characteristics; and recommend products to users within the prediction period based on the user purchase interest evolution model.

Images