User behavior directional recommendation method and system

Abstract

The invention relates to a user behavior method and system for directional recommendation, and belongs to the field of data processing, and the method comprises the steps: collecting full-platform interaction data of a target commodity, training a commodity interaction index prediction model, generating a real-time recommendation list, and carrying out the iterative optimization of the model through feedback data. In addition, the method also relates to the steps of dynamically determining a time window of data acquisition, acquiring interaction behaviors of a user on multiple terminals, extracting commodity interaction data features, training a model by using a machine learning or deep learning algorithm, generating a real-time recommendation list, and realizing dynamic alignment of recommended products and commodity interaction indexes. And pushing to the user terminal through the asynchronous message queue. Through the method, commodities can be recommended more accurately, the sales efficiency of merchants is improved, and the shopping experience of users is improved.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to a user behavior method and system for targeted recommendation. Background Art

[0002] Today, with the rapid development of the Internet economy, the number of users on e-commerce platforms is increasing day by day. However, each user has different preferences, needs, and behavior patterns. How to accurately guess the shopping needs and preferences of users to achieve personalized recommendation has always been a key issue that e-commerce platforms focus on. Usually, e-commerce platforms will collect and analyze user behavior data, model user behavior, and based on this, conduct product recommendations.

[0003] The key lies in how to accurately and efficiently collect and process user data, build a model that meets user needs, and then use this model to provide accurate product recommendations. However, existing product recommendation systems mainly make recommendations based on information such as users' browsing history and purchase history. However, this recommendation method ignores users' timeliness needs, situational needs, etc., and cannot meet users' diverse and personalized needs.

[0004] In addition, on e-commerce platforms, the behavior of each user may affect the effect of product recommendation. For example, if a user frequently clicks on a certain type of product, it may indicate that the user has a certain interest in this type of product. However, if the user has never made a purchase, it may mean that the user is just curious and does not intend to buy. Therefore, how to accurately understand and utilize this kind of user behavior data is also an important link in optimizing the product recommendation system.

[0005] In addition, in order to improve users' satisfaction with recommended products, the recommendation system also needs to be optimized according to users' feedback. For example, through data such as users' evaluations, click-through rates, and purchase behaviors of recommended products, the product recommendation system can be continuously adjusted and optimized. However, existing recommendation systems usually ignore the importance of this feedback link or do not provide an effective feedback mechanism.

[0006] The prior art CN118964745A, although it adopts a method for recommending government affairs big data, by obtaining the initial tags of the objects to be recommended, screening the first tags using annotation rules, calculating the tag weights to refine the core second tags, and constructing a multi-dimensional target product feature vector for content matching and score calculation, the specific process of its recommendation is still relatively complex, and the prediction of product interaction indicators and recommendation results may also be affected by certain subjective factors.

[0007] In summary, there are technical problems in the prior art that due to usually adopting static rules or simple statistical models and ignoring the timeliness of behavior, the capture of users' interests is not comprehensive enough and the recommendation accuracy is relatively low. Summary of the Invention

[0008] The purpose of this application is to provide a user behavior method and system for targeted recommendation, aiming to solve the technical problems in the prior art that due to the common use of static rules or simple statistical models and the neglect of the timeliness of behaviors, the capture of user interests is not comprehensive enough and the recommendation accuracy is relatively low.

[0009] In view of the above problems, this application provides a user behavior method for targeted recommendation. Collect the full-platform interaction data of the target commodity within a preset time window, and establish a commodity feature data set. The data includes category tags, price ranges, brand features, and SKU specification parameters. Extract the commodity interaction data feature set based on the commodity interaction index data set. The commodity interaction data features include the frequency of user clicks on the commodity details page, the duration of browsing a certain type of commodity, and the number of times of repeatedly viewing the same product. Train a commodity interaction index prediction model based on the commodity interaction data feature set. The model is used to predict the future behaviors of users. Generate a real-time recommendation list based on the prediction results, and dynamically adjust the recommended content in combination with the category / price association attributes. Real-time collect the feedback of commodity interaction indexes through the data embedding technology, and iteratively optimize the commodity interaction index prediction model based on the feedback data.

[0010] In one solution, the time window for data collection is dynamically determined according to business objectives and user life cycle characteristics. For the e-commerce scenario, select the past 30 days to 90 days as the core analysis period, and flexibly expand the time range according to the promotion activity cycle or seasonal demand.

[0011] In one solution, the collection of the behavior data covers the interaction behaviors of users on multiple terminals such as the mobile APP, web page, and mini-program. Through the data embedding technology, the page browsing path, commodity click events, search keyword input, add-to-cart / favorite operations, advertisement exposure and clicks, and explicit behaviors such as order payment completion are recorded in real time. At the same time, implicit feedback data is integrated, including the page stay duration, scroll depth, and video playback completion rate.

[0012] In one solution, the extraction of the commodity interaction data features includes: Statistically model and semantically associate to mine the potential interest patterns in the commodity interaction indexes. Construct a time-decaying weighted statistic for discrete events, and use the combination of the sliding window mean and standard deviation to characterize the stability for continuous behaviors.

[0013] In one solution, the commodity interaction index prediction model adopts: Training is carried out using machine learning or deep learning algorithms, including XGBoost and LSTM models. The model performance is optimized through cross-validation and hyperparameter tuning, and the model effects are evaluated using AUC, accuracy, and recall metrics.

[0014] In one solution, the generation of the real-time recommendation list includes: Combining category / price association attributes to dynamically adjust the recommended content, dynamically weighting the candidate products through an online re-ranking model, defining a time decay function to calculate the influence weight of real-time behavior on the products, and introducing a category dispersion constraint to prevent the decline of recommendation diversity.

[0015] In one solution, the feedback of the product interaction metrics is collected in real time through the data embedding technology, and the incremental update of the evaluation parameters is triggered through a periodic distribution feedback mechanism. Based on the feedback data, the product interaction metric prediction model is iteratively optimized to achieve the dynamic alignment of the recommended products and the evolution of the product interaction metrics.

[0016] In one solution, the real-time recommendation list is pushed to the user terminal through an asynchronous message queue. The push strategy follows the priority weight, and the multi-armed bandit algorithm is used to dynamically adjust the exposure probability. After receiving, the terminal performs lightweight rendering according to the device performance characteristics to ensure that the recommended loading delay is within the acceptable range.

[0017] In one aspect, a product interaction metric prediction system for targeted recommendation is used to implement the steps of the user behavior method for targeted recommendation. The product interaction metric prediction system for targeted recommendation includes: A product interaction metric data collection module, which is used to collect the behavior data of the target user within a preset historical time zone and establish a product interaction metric data set; An effective feature extraction module, which is used to extract the product interaction data feature set based on the product interaction metric data set; An evaluation mechanism construction module, which is used to construct a preset evaluation mechanism. Among them, the preset evaluation mechanism includes a behavior type evaluation mechanism and a periodic distribution feedback mechanism; A behavior prediction module, which is used to refer to the behavior type evaluation mechanism and the periodic distribution feedback mechanism, and based on the product interaction data feature set, predict the product interaction metrics, and establish a recommended product product feature vector sequence with an evaluation value greater than or equal to the preset evaluation value; A targeted recommendation module, which is used to generate recommended products based on the recommended product product feature vector sequence and send them to the terminal of the target user for targeted recommendation One or more technical solutions provided in this application have at least the following technical effects or advantages: The present invention adopts a technical solution based on the collection of commodity interaction index data and the extraction of characteristics of commodity interaction data, which solves the problems of traditional recommendation methods such as complex recommendation processes, susceptibility to subjective factors, low recommendation accuracy and efficiency, etc.

[0018] By collecting the behavior data of the target user within a preset historical time zone, establishing a commodity interaction index data set, extracting a commodity interaction data feature set based on the commodity interaction index data set, then training a commodity interaction index prediction model based on the commodity interaction data feature set, the model is used to predict the future behavior of the user, and a real-time recommendation list is generated based on the prediction result. In this way, the recommendation process becomes simple and targeted, greatly reducing the impact of subjective factors on the recommendation result.

[0019] Moreover, by using the buried point technology to collect the feedback of commodity interaction indexes in real time and iteratively optimizing the commodity interaction index prediction model based on the feedback data, the recommendation system can better understand the interests of users, continuously optimize the prediction model, so as to provide recommendation content closer to the actual needs of users, improving the accuracy and efficiency of the recommendation.

[0020] The above description is only an overview of the technical solution of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present application more obvious and understandable, the specific embodiments of the present application are specifically given below. It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present application, nor is it used to limit the scope of the present application. Other features of the present application will become easily understandable through the following description. Brief Description of the Drawings

[0021] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0022] Figure 1 It is a schematic flow chart of a user behavior method for targeted recommendation of the present application.

[0023] Figure 2 It is a schematic structural diagram of a commodity interaction index prediction system for targeted recommendation of the present application.

[0024] Description of the reference numerals: Commodity interaction index data collection module 11, Effective feature extraction module 12, Evaluation mechanism construction module 13, Behavior prediction module 14, Targeted recommendation module 15. Specific implementation manner

[0025] By providing a user behavior method and system for targeted recommendation, the present application solves the technical problems in the prior art that due to the common use of static rules or simple statistical models and the neglect of the timeliness of behaviors, the capture of user interests is not comprehensive enough and the recommendation accuracy is relatively low. By constructing a behavior type evaluation mechanism and a periodic distribution feedback mechanism, and combining with the feature extraction of the commodity interaction index dataset to predict the commodity interaction index, it is possible to more accurately identify the core interests and potential needs of users, achieving the technical effects of improving the accuracy of behavior prediction and the effectiveness of targeted recommendation.

[0026] Next, the technical solutions in the present application will be described clearly and completely with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present application. Additionally, it should be noted that for the sake of description, only the parts related to the present application are shown in the accompanying drawings rather than all of them.

[0027] Embodiment 1. Please refer to the attached Figure 1 , the present application provides a user behavior method for targeted recommendation. Among them, the user behavior method for targeted recommendation is applied to a commodity interaction index prediction system for targeted recommendation. The user behavior method for targeted recommendation specifically includes the following steps: S1. Collect the full-platform interaction data of the target commodity within a preset time window, establish a commodity feature dataset. The behavior data includes explicit behaviors and implicit feedback data, and the cross-terminal behavior trajectories are associated through a unified user ID.

[0028] First, it is necessary to dynamically determine the time window for data collection in combination with the business objectives and the characteristics of the user life cycle. For the e-commerce scenario, usually the past 30 days to 90 days are selected as the core analysis period, and at the same time, the time range is flexibly extended according to the promotion activity cycle or seasonal demands (such as holiday shopping tides) to ensure that the data can reflect both recent interests and long-term behavior patterns without omission. The data collection covers the interaction behaviors of users on multiple terminals (such as mobile APPs, web pages, mini-programs). Through the buried point technology, explicit behaviors such as page browsing paths, commodity click events, search keyword inputs, add-to-cart / favorite operations, advertisement exposures and clicks, and order payment completions are recorded in real time. At the same time, implicit feedback data (such as page stay duration, scroll depth, video playback completion rate) is integrated to form multi-dimensional original logs.

[0029] For cross - terminal users, the behavior trajectories are associated by a unified user ID (such as mobile phone number, email, or device fingerprint), and the original logs are cleaned and standardized with the help of ETL tools. Invalid sessions (such as accidental touches with page stay less than 1 second) are removed, and duplicate events (such as multiple click reports due to network jitter) are de - duplicated. Finally, the structured behavior data is stored in a distributed database (such as HBase or ClickHouse) according to the time series, forming a commodity interaction metric dataset containing fields such as user ID, behavior type, timestamp, behavior object (such as product ID, search term), and context information (such as device type, geographical location).

[0030] S2. Extract the commodity interaction data feature set based on the commodity interaction metric dataset. The commodity interaction data features include the frequency of a user clicking on the product detail page, the duration of browsing a certain type of product, and the number of times of repeatedly viewing the same product.

[0031] Conduct in - depth analysis and screening on the collected commodity interaction metric dataset to extract key features that can truly reflect the user's purchase intention or interest preference. Commodity interaction data features refer to those behavior patterns that are closely related to the target task and can significantly affect the results of the prediction model. For example, the frequency of a user clicking on the product detail page, the duration of browsing a certain type of product, and the number of times of repeatedly viewing the same product, etc., all belong to representative commodity interaction data features.

[0032] The core of extracting commodity interaction data features based on the commodity interaction metric dataset lies in statistically modeling and semantically associating to mine the potential interest patterns in the commodity interaction metrics. First, construct a time - decay weighted statistic for discrete events (such as clicks, add - to - carts, purchases) in the original behavior data. For example, the number of clicks of a user on a certain type of product is processed by an exponential decay function to obtain the recent interest intensity value: where t current is the current time, t i is the timestamp of the i - th behavior, λ is the decay factor (taking λ = 0.1 corresponds to a half - life of 7 days), and I click (i) is the indicator function (taking 1 when the i - th behavior is a click, otherwise 0). For continuous behaviors (such as page stay duration), a combination of moving window mean and standard deviation is used to characterize stability, such as the average daily page stay duration on the product detail page in the past 30 days, as follows: Secondly, map the user operation sequence to a low - dimensional dense vector through Embedding learning of the behavior sequence. Assume the commodity interaction metric sequence is {a1, a2,..., a m}, a iRepresents the combination of behavior type and object, such as "click - Product A"), and uses the Skip - gram model to maximize the co - occurrence probability of behaviors within the context window: Where is the embedding vector of behavior a t , and k is the window size. The embedding vectors generated in this process can capture the semantic associations between behaviors (the potential similarity between "purchase - headphones" and "browse - speakers"). Further, by label mapping, product categories are aligned with product interaction metrics, and the cosine similarity is used to calculate the matching degree between the behavior features and the target product's commodity feature vectors: When the similarity exceeds the preset threshold (0.75), the behavior feature is determined to be a valid feature. Finally, through feature importance ranking (based on the gain weight of XGBoost ), the TOP - K features are selected, low - contribution noise data is removed, and a user - product interaction data feature set is formed. The entire process needs to be manually calibrated in combination with business rules, and behaviors such as "high - frequency searches in the short term but not purchased" are defined as high - value features to be converted, thereby enhancing the business interpretability of the feature set.

[0033] S3. Train a product interaction metric prediction model based on the product interaction data feature set to predict the future behavior of users.

[0034] Specifically, in order to effectively predict product interaction metrics and provide targeted recommendations, constructing a preset evaluation mechanism is an important step. This mechanism assigns different evaluation values to the historical behaviors of users to quantify the recommendation value and time impact of behaviors. The preset evaluation mechanism mainly consists of two parts: the behavior type evaluation mechanism and the cycle distribution feedback mechanism.

[0035] The core of the behavior type evaluation mechanism is to set evaluation values according to the contribution degree of different behaviors to the recommendation target. For example, a user's purchase behavior is usually the most direct conversion signal and has a higher evaluation value; while the browsing behavior, although indirect, can reflect the user's potential interest, so the evaluation value is the second. Behavior types include clicking on ads, adding to the shopping cart, etc., and different evaluation values are assigned to each behavior.

[0036] When constructing the preset evaluation mechanism, it is necessary to deeply integrate the behavior type evaluation mechanism with the cycle distribution feedback mechanism to form a dynamically adjustable product interaction metric value quantification system. In the behavior type evaluation mechanism, first, the entropy weight method is combined with expert experience to assign weights to product interaction metric types. Define the purchase behavior w purchase = 0.35, the add - to - cart behavior w cart = 0.25, the search behavior w search = 0.20, page view wview = 0.15 equal weight values, and introduce the behavior quality decay factor α type = 1 - e -β·N (where N is the number of behaviors and β is the saturation coefficient), to prevent the excessive accumulation of weights for high-frequency low-value behaviors (invalid clicks). The evaluation score calculation for the single-item interaction index type is as follows: where K is the total number of behavior types, and N k is the number of occurrences of the k-th type of behavior. For the periodic distribution feedback mechanism, it is necessary to analyze the concentration and periodicity of the behavior time distribution. Define the time series {t1, t2,..., t M} as the set of timestamps of the commodity interaction index, and calculate its coefficient of variation C v = σ t / μ t , where σ t is the standard deviation of the adjacent behavior time intervals, and μ t is the mean value) to measure the dispersion degree of the behavior distribution, and combine the Fourier transform to extract the main period T dominant (the behavior peak per week or per month). The periodic distribution score consists of a concentration penalty term and a period matching term: where γ is the weight coefficient (take 0.6), η is the concentration decay factor, A p is the amplitude of the p-th period component, and T biz is the business cycle (promotion rhythm). The final preset evaluation total score is output through the coupling of the two mechanisms: This mechanism optimizes the parameters β, η, γ through the offline training set (labeling the user value level), and designs a real-time feedback loop: when the periodic distribution score is continuously lower than the threshold θ low for τ days (τ = 7), trigger the adaptive adjustment of the behavior type weights (such as increasing the search behavior weight Δw search = 0.05), to form the collaborative optimization of the evaluation mechanism and the evolution of the commodity interaction index. At the same time, set the anomaly detection rules for the behavior type and the periodic distribution (such as warning when the purchase behavior cycle deviates from the standard deviation by more than (3σ)), to ensure the robustness of the evaluation results.

[0037] S4. Generate a real-time recommendation list based on the prediction result, and dynamically adjust the recommended content in combination with the category / price association attributes.

[0038] In the prediction of product interaction metrics, comprehensively analyzing the product interaction data feature set by combining the behavior type evaluation mechanism and the cycle distribution feedback mechanism is a crucial step in constructing the product feature vectors of accurate recommendation products. This process quantifies the value and time distribution characteristics of product interaction metrics, predicts users' potential interests, and generates a sequence of product feature vectors of recommended products with evaluation values greater than or equal to the preset value, demonstrating the synergistic effect of the behavior type evaluation mechanism and the cycle distribution feedback mechanism in the prediction of product interaction metrics. By accurately calculating users' interest points and behavior values, it not only significantly improves the accuracy of product feature vectors but also provides more targeted strategic support for targeted recommendations.

[0039] In the process of predicting product interaction metrics and constructing product feature vectors of recommended products, it is necessary to integrate the product interaction data feature set, the behavior type evaluation weights, and the cycle distribution rules to establish a dynamic recommendation model. First, based on the product interaction data feature set f user = [f1, f2,..., f n (interest intensity, behavior sequence embedding vector) extracted in Step 2 and the total user evaluation value S total generated in Step 3, construct a time-aware dual-channel prediction network. The input layer maps the features to hidden states where denotes vector concatenation, and W h is the weight matrix; the time series channel models the main cycle T dominant in the behavior cycle feedback mechanism through LSTM, calculates the hidden state to capture the category / price association attribute pattern. The outputs of the two channels are fused through an attention mechanism: where V a , W a , U a are learnable parameters, and σ is the Sigmoid function. The prediction layer outputs the interaction probability of the user for the candidate product P: Here, f p is the product feature vector feature (such as category, price band, brand), and ⊙ is element-wise multiplication. The generation of the product feature vector sequence of recommended products needs to meet two constraints: 1) the prediction probability P ≥ θ pred (e.g., θ pred = 0.8); 2) the matching degree of the product feature vectors Sim(f user , f p ) ≥ θ sim , θ sim = 0.75. The final recommendation sequence is sorted by the weighted score: Score p = λ · P(y = 1||z, fp )+(1 - λ)·Sim(f user , f p ) where λ = 0.7 balances the prediction confidence and semantic matching degree. Meanwhile, the business rules are injected through the periodic distribution feedback mechanism: if the current time t satisfies |t - t last mod T dominant ≤ ΔT, where Δt is the periodic tolerance window), then the weight of the product that conforms to the periodic law is increased where A p is the historical interaction amplitude of the product within the period T dominant , and is the average amplitude of the category. Finally, the TOP-N recommended product feature vector sequence {p1, p2,..., p N} is output, and it is ensured that the comprehensive evaluation value Score p ≥ θ total (the preset threshold θ total = 85), so as to realize the joint optimization matching of the product interaction index value and the product feature vector.

[0040] S5. Real-time collect the feedback of product interaction indicators through the buried point technology, and iteratively optimize the product interaction index prediction model based on the feedback data.

[0041] In the recommended product generation and targeted push stage, it is necessary to dynamically integrate the recommended product feature vector sequence with the user terminal context to construct a real-time responsive recommendation engine. First, based on the TOP-N recommended product feature vector sequence generated in step four, combined with the category / price association attribute stream , is the timestamp, is the behavior type, is the context feature), dynamically weight the candidate products through an online re-ranking model. Define the time decay function , is the decay rate, taking 0.05 / s, and the influence weight of the real-time behavior on the product p is calculated as: where is the set of behavior types associated with the product p, is the context condition (such as the user's current geographical location, device type) matched by the product feature vector, is the indicator function. The final score of the recommended product is the linear combination of the static score in step four and the real-time weight: is the static score retention coefficient. To prevent the decline of recommendation diversity, the category dispersion constraint is introduced: When (threshold ), the category penalty item is triggered , and the score attenuation is implemented for the same category of products . At the same time, based on the periodic distribution feedback mechanism in Step 3, if the current time meets , then a gain coefficient is applied to the products that conform to the periodic law , where is the maximum historical interaction amplitude within the category.

[0042] After the recommendation list is generated, the structured data is pushed to the user terminal through the asynchronous message queue (Kafka), and the push policy follows the priority weight , and the multi-armed bandit algorithm is used to dynamically adjust the exposure probability: After receiving, the terminal performs lightweight rendering according to the device performance characteristics (such as memory, network bandwidth). If the user characteristics contain the low-end device identifier , then the feature dimensionality reduction technology is used, is the pre-trained projection matrix to ensure the recommended loading delay . The product interaction indicators (clicks, purchases, etc.) are sent back to the periodic distribution feedback mechanism in Step 3 in real time through data embedding, triggering the incremental update of the evaluation parameter : Finally, the dynamic alignment between the recommended products and the evolution of the product interaction indicators is achieved.

[0043] Embodiment 2. Based on the same inventive concept as the user behavior method for directional recommendation in the foregoing Embodiment 1, the present application also provides a product interaction index prediction system for directional recommendation. Please refer to the appendix Figure 2 , and the product interaction index prediction system for directional recommendation includes: The product interaction index data collection module 11 is used to collect the behavior data of the target user within the preset historical time zone and establish a product interaction index data set.

[0044] The effective feature extraction module 12 is used to extract the product interaction data feature set based on the product interaction index data set.

[0045] The evaluation mechanism construction module 13 is used to construct a preset evaluation mechanism, where the preset evaluation mechanism includes a behavior type evaluation mechanism and a periodic distribution feedback mechanism.

[0046] The behavior prediction module 14 is used to predict commodity interaction indicators based on the commodity interaction data feature set with reference to the behavior type evaluation mechanism and the periodic distribution feedback mechanism, and establish a sequence of recommended product commodity feature vectors with an evaluation value greater than or equal to a preset evaluation value.

[0047] The targeted recommendation module 15 is used to generate recommended products based on the sequence of recommended product commodity feature vectors and send them to the terminal of the target user for targeted recommendation.

[0048] Furthermore, the evaluation mechanism construction module 13 in the commodity interaction indicator prediction system for targeted recommendation is also used for: Establish a preset behavior type set, where the preset behavior type set includes multiple behavior types; configure multiple evaluation values for the multiple behavior types to generate the behavior type evaluation mechanism; perform cross-time zone check rate recognition for the multiple behavior types, and establish the periodic distribution feedback mechanism based on the check rate recognition results; establish the preset evaluation mechanism with the behavior type evaluation mechanism and the periodic distribution feedback mechanism.

[0049] Furthermore, the evaluation mechanism construction module 13 in the commodity interaction indicator prediction system for targeted recommendation is also used for: Traverse and combine the multiple behavior types, and collect check samples with different periodic distribution characteristics to construct multiple behavior distribution feature samples and multiple check sample sets; calculate the check rates based on the multiple check sample sets to generate multiple check rate samples; analyze based on the multiple behavior distribution feature samples and the multiple check rate samples to establish the periodic distribution feedback mechanism.

[0050] Furthermore, the evaluation mechanism construction module 13 in the commodity interaction indicator prediction system for targeted recommendation is also used for: Configure multiple feedback weights based on the multiple check rate samples, where the feedback weight is inversely proportional to the size of the check rate sample; use the multiple behavior distribution feature samples and multiple feedback weights as sample data, and train a feedback weight allocator based on a neural network model to generate the periodic distribution feedback mechanism.

[0051] Furthermore, the effective feature extraction module 12 in the commodity interaction indicator prediction system for targeted recommendation is also used for: Determine the product type labels of all products to be recommended by the target enterprise; use the product type labels as constraints to extract multiple behavior characteristics of the same type of products from the commodity interaction indicator dataset to establish the commodity interaction data feature set.

[0052] Further, the behavior prediction module 14 in the commodity interaction index prediction system for directional recommendation is further configured to: Count multiple behavior characteristics for the same type of product and the frequency of occurrence of each behavior characteristic in the commodity interaction data feature set, calculate the evaluation value according to the behavior type evaluation mechanism, and generate multiple basic evaluation values for multiple products of the same type; extract the time-period distribution characteristics for multiple behavior characteristics of the same type of product, and establish multiple time-period distribution characteristics for multiple products of the same type; analyze the multiple time-period distribution characteristics through the period distribution feedback mechanism, and perform feedback adjustment on the multiple basic evaluation values based on the feedback analysis results to generate multiple comprehensive evaluation values; determine the product types with the evaluation value greater than or equal to the preset evaluation value based on the multiple comprehensive evaluation values, and arrange them in descending order of the evaluation value to generate the recommended product commodity feature vector sequence.

[0053] Further, the directional recommendation module 15 in the commodity interaction index prediction system for directional recommendation is further configured to: Extract the first recommended product commodity feature vector from the recommended product commodity feature vector sequence, connect it with the product library of the target enterprise for matching, and obtain the first recommended product with a matching degree greater than the preset threshold; send the first recommended product to the terminal of the target user for directional recommendation.

[0054] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is the difference from other embodiments. The Figure 1 A user behavior method and specific example for directional recommendation in the first embodiment are equally applicable to the commodity interaction index prediction system for directional recommendation in this embodiment. Through the detailed description of the user behavior method for directional recommendation above, those skilled in the art can clearly know the commodity interaction index prediction system for directional recommendation in this embodiment. Therefore, for the sake of simplicity of the specification, it will not be elaborated here. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0055] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0056] Obviously, those skilled in the art can make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of this application and its equivalent technologies, this application also intends to include these changes and modifications.

Claims

1. A user behavior method for targeted recommendation, characterized in that: include: Collect the full platform interaction data of the target product within the preset time window and establish a product feature data set, which includes category tags, price segments, brand features, and SKU specification parameters; Extracting a product interaction data feature set based on the product interaction index data set. The product interaction data features include the frequency of users clicking on the product detail page, the length of time they browse a certain type of product, and the number of times they repeatedly view the same product. Training a product interaction index prediction model based on a product interaction data feature set, wherein the model is used to predict future user behavior; Generate a real-time recommendation list based on the prediction results, and dynamically adjust the recommended content based on the category / price related attribute flow; The product interaction indicator feedback is collected in real time through the tracking technology, and the product interaction indicator prediction model is iteratively optimized based on the feedback data.

2. A user behavior method for targeted recommendation according to claim 1, characterized in that: The time window for data collection is dynamically determined based on business objectives and user life cycle characteristics. For e-commerce scenarios, the past 30 to 90 days are selected as the core analysis period, and the time range is flexibly expanded based on promotional activity cycles or seasonal needs.

3. A user behavior method for targeted recommendation according to claim 1, characterized in that: The behavioral data collection covers the user's interactive behaviors on multiple terminals such as mobile APP, web pages, and mini-programs. It records the page browsing path, product click events, search keyword input, add to cart / favorite operations, advertising exposure and clicks, and order payment completion in real time through the embedding technology, and integrates implicit feedback data, including page dwell time, scrolling depth, and video playback completion rate.

4. A user behavior method for targeted recommendation according to claim 1, characterized in that: The extraction of commodity interaction data features includes: Through statistical modeling and semantic association, potential interest patterns in product interaction indicators are mined, time-decayed weighted statistics are constructed for discrete events, and the sliding window mean and standard deviation are combined to characterize stability for continuous behaviors.

5. A user behavior method for targeted recommendation according to claim 1, characterized in that: The commodity interaction index prediction model adopts: Machine learning or deep learning algorithms are trained, including XGBoost and LSTM models, and model performance is optimized through cross-validation and hyperparameter tuning. Model effects are evaluated using AUC, accuracy, and recall indicators.

6. A user behavior method for targeted recommendation according to claim 1, characterized in that: The generation of the real-time recommendation list includes: The recommended content is dynamically adjusted based on the category / price associated attribute flow, the candidate products are dynamically weighted through the online reranking model, a time decay function is defined to calculate the impact weight of real-time behavior on the product, and category dispersion constraints are introduced to prevent the reduction of recommendation diversity.

7. A user behavior method for targeted recommendation according to claim 1, characterized in that: The product interaction index feedback is collected in real time through the tracking technology, and the incremental update of the evaluation parameters is triggered through the periodic distribution feedback mechanism. The product interaction index prediction model is iteratively optimized based on the feedback data to achieve dynamic alignment of the recommended products and the evolution of the product interaction index.

8. A user behavior method for targeted recommendation according to claim 1, characterized in that: The real-time recommendation list is pushed to the user terminal through an asynchronous message queue. The push strategy follows the priority weight and uses a multi-armed bandit algorithm to dynamically adjust the exposure probability. After receiving the list, the terminal performs lightweight rendering according to the device performance characteristics to ensure that the recommendation loading delay is within an acceptable range.

9. A user behavior prediction system for targeted recommendation, characterized in that: The steps for implementing the user behavior method for targeted recommendation as claimed in any one of claims 1 to 8, wherein the commodity interaction index prediction system for targeted recommendation comprises: The commodity interaction index data collection module is used to collect the behavior data of the target users in a preset historical time zone and establish a commodity interaction index data set; An effective feature extraction module, used for extracting a commodity interaction data feature set based on the commodity interaction indicator data set; An evaluation mechanism construction module, used to construct a preset evaluation mechanism, wherein the preset evaluation mechanism includes a behavior type evaluation mechanism and a periodic distribution feedback mechanism; A behavior prediction module, used to predict user behavior based on the commodity interaction data feature set with reference to the behavior type evaluation mechanism and the periodic distribution feedback mechanism, and to establish a recommended product commodity feature vector sequence with an evaluation value greater than or equal to a preset evaluation value; The targeted recommendation module is used to generate recommended products based on the recommended product commodity feature vector sequence and send them to the terminal of the target user for targeted recommendation.

Images