Information recommendation method and device in scientific instrument field

Abstract

The invention provides an information recommendation method and device in the field of scientific instruments, and belongs to the technical field of data processing.The method comprises the steps that a search request is received, the search request comprises first user identity information and first search information, the first user identity information corresponds to a first user identity feature, and the first search information corresponds to a second user identity feature; the first user identity feature comprises at least one of a city where the user is located, a gender of the user, an education background of the user, a user type and a user basic portrait; obtaining a prediction probability of each browsing direction based on the first user identity information and the first search information; and outputting a preset number of browsing directions with the highest prediction probability. According to the method, the discrimination capability of the prediction model on the user intention is improved by integrating the user identity features and the search information, the recommendation accuracy is improved, and the user experience is optimized.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an information recommendation method and apparatus in the field of scientific instruments. Background Technology

[0002] The Instrument Information Network platform is a professional service platform focusing on the field of scientific instruments, primarily providing services such as industry news, product procurement, and technical exchange. When users conduct text searches on the Instrument Information Network platform, the platform will output browsing direction recommendations, such as information matching the search text in the industry news, product procurement, and technical exchange browsing directions.

[0003] Existing scientific instrument information recommendation methods on instrument information network platforms rely solely on search text for recommendations, neglecting personalized user characteristics. This results in low accuracy and low user satisfaction. Summary of the Invention

[0004] This application provides an information recommendation method and apparatus in the field of scientific instruments, aiming to solve the problem that existing recommendation methods ignore user personalized features, resulting in low accuracy of recommendation results.

[0005] Firstly, this application provides an information recommendation method in the field of scientific instruments, including: Receive a search request, which includes first user identity information and first search information. The first user identity information corresponds to first user identity features, which include at least one of the following: user's city, user's gender, user's education level, user type, and user basic profile. The predicted probabilities for each browsing direction are obtained based on the first user's identity information and the first search information; Output the preset number of browsing directions with the highest predicted probability.

[0006] In one embodiment, after receiving a search request, the method further includes: Based on the first user's identity information, extract the first user's behavior data within a preset time period before the current time, and determine the first user's behavior characteristics based on the first user's behavior data; Furthermore, the predicted probability of each browsing direction is obtained based on the first user's identity information, the first search information, and the first user's behavioral characteristics.

[0007] In one embodiment, determining the first user behavior characteristics based on the first user behavior data specifically includes: The first user behavior data is categorized according to the actual browsing direction clicked in each first historical behavior; each browsing direction contains multiple events. For each browsing direction, the historical behavior score of the browsing direction is determined based on the behavior time of all first historical behaviors of the browsing direction and the actual click events corresponding to all first historical behaviors. The historical behavior scores from all browsing directions constitute the first user behavior characteristic.

[0008] In one embodiment, the predicted probability of each browsing direction is obtained based on the first user identity information, the first search information, and the first user behavior characteristics, specifically including: The first user identity information, the first search information, and the first user behavior features are respectively transformed into first feature vectors; The prediction model is input with the first user identity information, the first search information, and the first feature vector corresponding to the first user behavior features to obtain the prediction probability for each browsing direction.

[0009] In one embodiment, training the prediction model includes: Obtain the basic dataset. In the basic dataset, each piece of basic data includes the second user behavior characteristics based on the same second historical behavior, the second historical behavior information, the actual browsing direction of the click in the second historical behavior, and the second user identity characteristics. All second user behavior features, all second historical behavior information, and all second user identity features are converted into corresponding second feature vectors, and all browsing directions are converted into corresponding tag vectors. The second feature vectors corresponding to the second user behavior feature, second historical behavior information, and second user identity feature of each piece of basic data are weighted and concatenated, and then concatenated with the label vector of the actual browsing direction clicked in the basic data to form a sample corresponding to the basic data. All samples are then combined to form a training dataset. In the training dataset, the weights of the second user behavior feature, second historical behavior information, and each second user identity feature are the initial weights. The prediction model is iteratively trained using the training dataset. After training, the final weights of user behavior features, search information, and user identity features are obtained.

[0010] In one embodiment, the prediction model is a limit gradient boosting model, and the final weights of user behavior features, search information, and user identity features are determined by the total number of times user behavior features, search information, and user identity features are used as split points in all trees when training is complete.

[0011] Secondly, this application also provides an information recommendation device in the field of scientific instruments, comprising: The receiving module is used to receive a search request. The search request includes first user identity information and first search information. The first user identity information corresponds to first user identity features. The first user identity features include at least one of the following: the user's city, the user's gender, the user's education level, the user type, and the user's basic profile. The prediction module is used to obtain the predicted probability of each browsing direction based on the first user identity information and the first search information; The output module is used to output the preset number of browsing directions with the highest predicted probability.

[0012] Thirdly, this application also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the information recommendation method in any of the above-mentioned scientific instrument fields.

[0013] Fourthly, this application also provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the information recommendation method in any of the above-mentioned scientific instrument fields.

[0014] Fifthly, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the information recommendation method in any of the above-mentioned scientific instrument fields. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is one of the flowcharts illustrating the information recommendation method in the field of scientific instruments provided in this application; Figure 2 This is the second flowchart illustrating the information recommendation method in the field of scientific instruments provided in this application; Figure 3 This is one of the flowcharts provided in this application for determining the characteristics of a first user behavior based on first user behavior data; Figure 4 This is one of the flowcharts illustrating the training process for the prediction model provided in this application; Figure 5 This is an example of the final list of the top 20 features by weight provided in this application; Figure 6This is a graph showing the process loss during training provided in this application; Figure 7 This is a graph showing the change in accuracy during training, as provided in this application. Figure 8 This is a schematic diagram of the structure of the information recommendation device in the field of scientific instruments provided in this application; Figure 9 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0019] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects have an "or" relationship.

[0020] The following is combined Figures 1 to 9 This application describes the information recommendation method and apparatus in the field of scientific instruments.

[0021] It should be noted that the information recommendation method in the field of scientific instruments provided in this application is implemented based on the information recommendation device in the field of scientific instruments. The information recommendation method in the field of scientific instruments recommends browsing directions based on user personalized features and multimodal data of search information, thereby improving the matching degree between the recommendation results and the user.

[0022] This application describes an information recommendation method in the field of scientific instruments, using an information recommendation device in the field of scientific instruments as an example.

[0023] It should be noted that this application is applicable to various information recommendation platforms, such as the Instrument Information Network platform. The following detailed explanation uses the Instrument Information Network platform as an example.

[0024] In one possible implementation, the instrument information network platform includes multiple recommended browsing directions, including communities, resources, instruments, manufacturers, news, consumables, reagents, and industry applications.

[0025] Figure 1 This is one of the flowcharts illustrating the information recommendation method in the field of scientific instruments provided in this application. For example... Figure 1 As shown, the information recommendation method in the field of scientific instruments provided in this application includes: S110: Receive a search request. The search request includes first user identity information and first search information.

[0026] Among them, user identity information corresponds to user identity characteristics, and users with different user identity characteristics have different preferences for different browsing directions on the information recommendation platform.

[0027] In an embodiment of the Instrument Information Network platform, user identity characteristics may include at least one of the following: user's city (discrete value), user's gender (discrete value), user's education level (discrete value), user type (discrete value), and user basic profile. The user basic profile may include the nature of the organization, principle classification 1, principle classification 2, principle classification 3, industry classification 1, and industry classification 2. The classification methods for the three principle classifications are different, and the classification methods for the two industry classifications are also different.

[0028] S120: Based on the first user identity information and the first search information, obtain the predicted probability of each browsing direction, so as to obtain the user's preference probability for each browsing direction.

[0029] Specifically, in the prediction process, the user identity features corresponding to the first user identity information and the semantic features corresponding to the search information are fused together. The preference information provided by the user identity features is used, combined with the semantic features, to obtain a browsing direction that is close to the user's preferences and meets the current search needs.

[0030] In this embodiment, the prediction model learns the importance of user identity features and text features for browsing direction. In the trained prediction model, different browsing directions have varying dependencies on different user identity features and text features.

[0031] Among these, user identity characteristics are crucial for distinguishing users' professional field intentions. For example, in the aforementioned embodiment of the Instrument Information Network platform, the recommendations for browsing "manufacturers" and "instruments" rely more on the user's "organizational nature" (such as universities, research institutes, and enterprises) and "industry classification"; while the recommendations for browsing "communities" and "information" are more correlated with the user's "education level" and "user type".

[0032] The semantic features of search information directly capture the user's current immediate intent, which is the foundation for ensuring the relevance of recommendations. For example, the semantic features of the search term "liquid chromatography principle" are closer to the text content of the "data" browsing direction in the vector space, while "Agilent 1200 price quote" is closer to the "instrument" or "manufacturer" browsing direction.

[0033] Therefore, based on user identity characteristics and semantic features of search information, we can recommend browsing directions with a higher accuracy rate to users.

[0034] S130: Outputs the preset number of browsing directions with the highest predicted probability.

[0035] Specifically, the predicted probabilities of each browsing direction are sorted in descending order, and the top-ranked browsing directions (preset number) are output as a list of recommended browsing directions.

[0036] It should be noted that in this step, after obtaining a preset number of browsing directions, the system also needs to filter at least one piece of data that matches the search information from multiple data entries for each browsing direction, and then output the filtered data to the user. The data filtering process can employ existing filtering methods, and this application does not impose any restrictions on it.

[0037] This application embodiment improves the predictive model's ability to discern user intent by integrating user identity features and search information, thereby increasing recommendation accuracy and optimizing user experience.

[0038] Traditional recommendation methods mostly rely on manually defined semantic templates, making recommendations based solely on the search text. When a user enters search text, the recommendation system matches the search text with the browsing direction based on predefined semantic rules to obtain recommendations. For example, based on semantic rules, the search text "spectrometer" corresponds to the browsing direction of "instrument," so when a user enters the search text "spectrometer," the output recommendation information will be related to the browsing direction of "instrument." Manually defined semantic templates cannot dynamically adapt to changes in user preferences, resulting in low accuracy of recommendations, such as when facing new users or changes in user behavior patterns.

[0039] Based on the above considerations, such as Figure 2As shown, the information recommendation method in the field of scientific instruments provided in this application includes: S210: Receive search request. Please refer to step S110.

[0040] S220: Based on the first user's identity information, extract the first user's behavior data within a preset time period (e.g., 30 days) prior to the current time, and determine the first user's behavior characteristics based on the first user's behavior data.

[0041] User behavior data refers to user behavior data within the Instrument Information Network platform's search and eight sections. This includes the user's search terms, the browsing direction clicked after entering search terms, and the browsing direction clicked directly without entering search terms. Recent historical user behavior data directly reflects users' long-term interests and preferences, serving as crucial foundational data for personalized recommendations.

[0042] Specifically, user search logs are obtained from information recommendation platforms (such as the Instrument Information Network platform) as the first user behavior data, including user identity information, search text, click behavior, and timestamps of user behavior. Subsequently, the corresponding first user behavior features are obtained based on the first user behavior data.

[0043] S230: Obtain the predicted probabilities for each browsing direction based on the first user's identity information, the first search information, and the first user's behavioral characteristics. Please refer to step S120.

[0044] Building upon step S120 above, it's important to further explain that the predictive model also learns the importance of user behavior characteristics for browsing direction. For example, a user who consistently exhibits high activity in the "Documents" browsing direction is significantly more likely to have their new search requests recommended to the "Documents" browsing direction.

[0045] During training, user behavior features for each browsing direction are input as continuous features into the prediction model. The model can automatically learn the complex nonlinear relationship between these user behavior features and the final browsing direction clicked, achieving an intelligent trade-off between long-term interests and immediate intentions. For example, even if a user's behavior features are high in the "instruments" browsing direction, but the search text for this search is highly relevant to the "data" browsing direction, the model may still recommend "data".

[0046] On the other hand, the role of users' historical behavioral characteristics in recommendation results is as follows: for new users, since there are no historical behavioral characteristics, the recommendations for new users rely more on user identity characteristics; for old users, since there are historical behavioral characteristics, the recommendations are closer to user preferences.

[0047] S240: Output the browsing direction with the highest predicted probability (preset number). Please refer to step S130.

[0048] This application's embodiments improve the predictive model's ability to discern user intent by integrating user identity features, historical user behavior features, and search information, making the recommendation results more aligned with user preferences, achieving dynamic personalized recommendations, and improving recommendation accuracy.

[0049] Because traditional recommendation methods have low accuracy (see the above explanation), user satisfaction metrics (such as click-through rate) of information recommendation platforms are poor, failing to meet the platform's need for high-precision recommendations.

[0050] Based on these considerations, in one possible implementation, such as Figure 3 As shown, in step S220, determining the first user behavior characteristics based on the first user behavior data specifically includes: S310: Classify the first user behavior data according to the browsing direction of the actual click in each first historical behavior in the first user behavior data.

[0051] Each browsing direction contains multiple events. For example, for the "Instruments" browsing direction, events include "Instrument Selection - Click to Apply for Trial", "Instrument Selection - Click to Leave a Message", and "Instrument Selection - Message Successful". For the "Manufacturers" browsing direction, events include "Find a Manufacturer - Phone Inquiry", "Find a Manufacturer - Click to Leave a Message", and "Manufacturer - Login Successful".

[0052] For each browsing direction, user history includes both historical actions where the user clicked on an event in that browsing direction after entering search terms, and historical actions where the user clicked on an event in that browsing direction without entering search terms. The first user behavior data is categorized based on the browsing direction to which the user's actual click event belongs, thus obtaining the first historical behavior for each browsing direction.

[0053] S320: For each browsing direction, determine the historical behavior score of the browsing direction based on the behavior time of all first historical behaviors of that browsing direction and the actual click events corresponding to all first historical behaviors.

[0054] Specifically, the historical behavior score for each browsing direction is determined based on all first historical behaviors under that browsing direction.

[0055] In one possible implementation, an exponential time decay algorithm is used to calculate the historical behavior score for each browsing direction, as shown in the following formula: (1); in, It is the first iHistorical behavior scores for each browsing direction; It is the first i The number of events in each browsing direction, It is a positive integer; It is the first i The first browsing direction j The weight of each event; It is a natural constant e (Euler number), approximately equal to 2.71828; The time decay factor, For the occurrence of the first i The first browsing direction j The time of each event; Record the time for the sample data (e.g., the time unit is days).

[0056] Here are a few examples of historical behavior ratings: Example 1: The following are the user behaviors recorded in the "Instruments" browsing direction over the past six months: 1) Time: =2025-06-10; Event: Browsing Instrument Preferred; Event Weight: 1.

[0057] 2) Time: =2025-05-15; Event: Browsing preferred instruments; Event weight: 1.

[0058] =0.01, =2025-07-14, then the historical behavior score is calculated as follows: 1* + 1* = 0.71 + 0.55 = 1.26.

[0059] Example 2: Users who have not used "consumables" in the past six months and whose browsing behavior is in the "data" direction include: 1) =2025-07-01; Event: Browsing the Data Center; Event Weight: 1.

[0060] =0.01, =2025-07-14, then the historical behavior score = 1* =0.88.

[0061] In one possible implementation, weights Manual prior assignment of values ​​based on the value of different types of events.

[0062] In one possible implementation, There are multiple levels, each corresponding to a different weight value. For example, the weight value for "telephone inquiry" events is 5, and the weight value for "browsing" events is 1.

[0063] Time decay factor The time decay factor determines the rate of interest decay; a higher value indicates faster decay, suggesting a smaller weighting of long-term behavioral factors and a greater bias in ratings towards recent behavior. (Time decay factor) The settings make recent behavior have a much greater impact on ratings than long-term behavior, thus dynamically and continuously depicting the changes in user interests, rather than binary or static statistics.

[0064] S330: Historical behavior scores across all browsing directions constitute the primary user behavior characteristic.

[0065] This application's embodiments determine the weight of each type of event based on its value, thereby determining the historical behavior score for each browsing direction. This allows the prediction model to learn the contribution of each type of event to each browsing direction during training, and the model can predict the probability of each browsing direction being clicked at a finer granular level during the prediction process, thereby improving user satisfaction metrics.

[0066] In one possible implementation, step S230 involves obtaining the predicted probabilities of each browsing direction based on the first user identity information, the first search information, and the first user behavior characteristics. Specifically, this includes: S2301: Convert the first user identity information, the first search information, and the first user behavior features into first feature vectors respectively.

[0067] Understandably, data cleaning is performed before feature transformation, including handling missing values, outliers, duplicate values, etc.

[0068] In one possible implementation, missing discrete values ​​are filled with -100, treating them as a category. Missing continuous values ​​are filled with the mean. If a user's identity feature is null, it is filled with "unknown".

[0069] In one possible implementation, for abnormal user behavior features that exceed the data range of user behavior characteristics, the abnormal user behavior features are replaced with the boundary value of the data range. For events whose timestamp of the behavior time exceeds the current time, the timestamp of the behavior time is set to the current time.

[0070] In one possible implementation, duplicate values ​​are processed based on the time field. If the time interval between two duplicate values ​​is within a preset time period (e.g., five minutes), one of them is retained; if the time interval exceeds the preset time period, both duplicate values ​​are retained.

[0071] In one possible implementation, the user identity features corresponding to the user identity information are one-hot encoded or labeled to obtain the feature vector corresponding to the user identity features.

[0072] In one possible implementation, user behavior features are standardized and normalized to obtain feature vectors corresponding to the user behavior features.

[0073] In one possible implementation, word embedding is performed on the search information to obtain the feature vector corresponding to the search information.

[0074] In one possible implementation, the search information (e.g., search text) is first preprocessed, including stop word removal and standardization. Then, the search text is embedded using Tencent's Chinese text pre-trained word vector model, outputting a 200-dimensional vector.

[0075] S2302: Input the first user identity information, the first search information, and the first feature vector corresponding to the first user behavior features into the prediction model to obtain the prediction probability of each browsing direction.

[0076] This application embodiment improves the model's running speed by converting user identity information, search information, and user behavior characteristics into feature vectors that the model can recognize, and then inputting them into the prediction model.

[0077] In one possible implementation, the prediction model is trained using supervised training. In one possible implementation, such as... Figure 4 As shown, training the prediction model includes: S410: Obtain the basic dataset.

[0078] In the basic dataset, each piece of basic data includes second user behavior features based on the same second historical behavior, second historical behavior information, the actual browsing direction clicked in the second historical behavior, and second user identity features. Please refer to step S220.

[0079] During the training phase, the actual browsing direction clicked in the second historical behavior is the label of the second historical behavior.

[0080] S420: Convert all second user behavior features, all second historical behavior information, and all second user identity features into corresponding second feature vectors, and convert all browsing directions into corresponding tag vectors. Please refer to step S2301.

[0081] S430: For each piece of basic data, the weighted concatenation of the second feature vector corresponding to the second user behavior feature, second historical behavior information, and second user identity feature is performed, and then concatenated with the label vector of the actual browsing direction clicked in the basic data. This concatenation forms a sample corresponding to the basic data, and all samples are combined to form the training dataset. In the training dataset, the weights of the second user behavior feature, second historical behavior information, and each second user identity feature are the initial weights.

[0082] This application embodiment uses a feature-weighted concatenation method to fuse multi-source data together, and concatenates three types of features into a high-dimensional joint feature vector.

[0083] In the early stages of model training, the importance of user behavior features, historical behavior information, and user identity features is analyzed to assign initial weights to them.

[0084] For example, analysis of the basic dataset revealed that historical behavioral features and user identity features contributed significantly, and their initial weights were higher than those of search information. This allowed the model to learn user-specific behavioral patterns more effectively. Compared to indiscriminate merging, this domain-knowledge-based weighted fusion method more effectively guides the model to focus on key information.

[0085] It should be noted that each user behavior feature for each browsing direction, each historical behavior information, and each user identity feature (user's city, user's gender, user's education level, user type, and user's basic profile) corresponds to an initial weight.

[0086] S440: Iteratively train the prediction model using the training dataset.

[0087] During the training of the prediction model, the weights of user behavior features, historical behavior information, and user identity features in each browsing direction are adjusted based on changes in the loss function and accuracy in order to improve the model's recommendation accuracy.

[0088] After training, the final weights of user behavior features, search information, and user identity features are obtained. This combination of weights maximizes the model's recommendation accuracy.

[0089] Figure 5An example of the list of the top 20 features ranked by the final weights is given. The features from top to bottom are as follows: the historical behavior score in the community browsing direction (shequ), the historical behavior score in the profile browsing direction (ziliao), the historical behavior score in the instrument browsing direction (yiqi), the historical behavior score in the information browsing direction (zixun), the historical behavior score in the manufacturer browsing direction (changshang), the historical behavior score in the consumable browsing direction (haocai), the historical behavior score in the industry application browsing direction (hangyeyingyong), the historical behavior score in the reagent browsing direction (shiji), principle classification 2 (yl_2_encoder), industry classification 2 (hy_2_encoder), principle classification 3 (yl_3_encoder), city (city_encoder), industry classification 3 (hy_1_encoder), principle classification 1 (yl_1_encoder), the 138th, 156th, 28th, 80th, 112th, and 144th features in the 200-dimensional text feature vector (vec_138, vec_156, vec_28, vec_80, vec_112, vec_144).

[0090] In the embodiment of this application, on the basis of integrating multi-source features, the weights of various features are adjusted through model training, which can ensure that the recommendation accuracy reaches the best.

[0091] For the training data set with a large difference in the number of samples in different browsing directions (for example, there are more samples in the "information" browsing direction and fewer samples in the "reagent" browsing direction), the training process is likely to lead to unfair recommendation results due to class imbalance.

[0092] Based on such considerations, in a possible implementation manner, in step S440, the prediction model is iteratively trained using the training data set, which specifically includes: Obtaining a loss function based on the class weights of each browsing direction; Among them, the class weights are inversely proportional to the number of samples in the browsing direction in the training data set, and the loss function is proportional to the class weights.

[0093] In a possible implementation manner, the class weights are calculated as follows: (2); Among them, is the total number of samples, is the total number of classes (i.e., the total number of browsing directions), is the number of samples in the j th browsing direction.

[0094] The loss function The calculation formula is as follows: (3) in, It is a true classification, if the first i The sample belongs to the first j Each category, then =1; otherwise, =0; log is the natural logarithm; It is the first one predicted by the model. i The sample belongs to the first j The probability of each category.

[0095] As shown in the above formula, the category weight of each browsing direction is inversely proportional to the number of samples in that browsing direction. This increases the category weight of browsing directions with fewer samples, which in turn increases the proportion of browsing directions with fewer samples in the loss function. This causes the model to impose a greater penalty on the prediction errors of categories with fewer samples during the training process, thereby forcing the model to learn the features of categories with fewer samples better and effectively alleviating the problem that the model tends to recommend categories with more samples.

[0096] This application's embodiments use category weights to eliminate the impact of category imbalance, ensuring fair recommendation results.

[0097] Based on the above, in one possible implementation, the prediction model is the eXtreme Gradient Boosting (XGBoost) model. In this embodiment, the final weights of user behavior features, search information, and user identity features are determined by the total number of times user behavior features, search information, and user identity features are used as split points in all trees at the end of training.

[0098] The XGBoost model boasts fast inference speed, making it suitable for real-time recommendation scenarios. Furthermore, user identity features, after feature encoding, generate a large number of sparse features. The XGBoost algorithm itself possesses sparsity awareness, enabling efficient processing of such features. Specifically, this embodiment uses the Histogram algorithm for node splitting and leverages its default sparse data processing strategy to automatically learn the splitting direction of sparse features, significantly improving training efficiency and accuracy.

[0099] In one possible implementation, during model training, a grid search is used for parameter tuning, with a decision tree depth ranging from 5 to 10 layers, a learning rate ranging from 0.001 to 0.3, and the number of decision trees ranging from 100 to 1000. The training process loss (Loss History) and accuracy history are as follows: Figure 6 and Figure 7 As shown.

[0100] In one example, the optimal parameter combination determined through grid search is: a tree depth of 6 layers and a tree size of 500 trees. The relatively shallow tree depth (6 layers) effectively prevents overfitting and improves the model's generalization ability. The size of 500 trees controls computational costs while ensuring model capacity.

[0101] The embodiments of this application employ an early stopping strategy for model training to achieve convergence: the number of early stopping rounds is set (e.g., 50 rounds), and training is automatically terminated when the F1 score on the validation set no longer improves for several consecutive rounds, with a maximum of 1000 iterations. This not only avoids overfitting but also saves training resources.

[0102] To verify the effectiveness of the embodiments of this application, an evaluation was performed on a test set. The evaluation results are shown in Table 1 below: Table 1: Performance Evaluation Report of XGBoost Multi-Class Classification Model on the Test Set Browsing direction name Accuracy Recall rate F1 score Number of supporters instrument 0.97 0.96 0.96 19786 Manufacturers 0.95 0.96 0.96 7437 Community 0.93 0.94 0.94 15713 Consumables 0.94 0.96 0.95 2170 Industry Applications 0.90 0.94 0.92 883 reagents 0.97 0.97 0.97 1017 material 0.94 0.93 0.94 16874 Information 0.93 0.94 0.94 6971 accuracy 0.95 70851 The evaluation results in the table above show that the embodiments of this application have achieved excellent and balanced performance in all browsing directions (F1 scores are all above 0.92). Even in the "industry application" browsing direction with the smallest sample size, an F1 score of 0.92 has been achieved. This indicates that the feature processing and model training strategies of the embodiments of this application have effectively solved the data imbalance problem, achieved high-quality personalized recommendations, and demonstrated the high accuracy and reliability of the solution.

[0103] Based on the XGBoost model trained above, several inference examples are given below: Example 1: User searches for "spectrometer": Input: User identity information: U1001, Search text: "Spectrometer".

[0104] Feature calculation: User identity characteristics: City = "Beijing" (encoded value 1), Gender = "Male" (encoded value 1), Education = "Bachelor's Degree" (encoded value 3), User Type = "Enterprise User" (encoded value 2), User Basic Profile: Unit Nature = "Production" (encoded value 1), Principle Classification 1 = 1 (Yes), Industry Classification 1 = 1 (Yes).

[0105] User historical behavior characteristics: Behavior records within the past six months: User behavior in the "Instrument" browsing direction is as follows: 1) Time: =2025-06-10; Event: Browsing Instrument Preferred; Event Weight: 1. 2) Time: =2025-05-15; Event: Browsing preferred instruments; Event weight: 1.

[0106] Search information: word vectors for "spectrometer" (200-dimensional vectors, e.g., [0.12, -0.05, ..., 0.08]).

[0107] Model inference: The feature vectors of the above features are input into the XGBoost model, and the output preference probabilities are: Instrument = 0.85, Manufacturer = 0.10, Other browsing directions < 0.05.

[0108] Recommended output: Top-3 browsing directions after sorting: Instruments, Manufacturers, Documents.

[0109] Example 2: User searches for "laboratory consumables": Input: User identity information: U2002, search text: "laboratory consumables".

[0110] Feature calculation: User identity characteristics: City = "Shanghai" (encoded value 2), Gender = "Female" (encoded value 0), Education = "Master" (encoded value 4), User type = "Research institution" (encoded value 3), User basic profile: Institution nature = "Research-oriented" (encoded value 2), Principle classification 2 = 1 (yes).

[0111] User historical behavior characteristics: No "consumables" behavior in the past six months; user behavior in the "data" browsing direction includes: 1) =2025-07-01; Event: Browsing the Data Center; Event Weight: 1.

[0112] Search information: word vector for "laboratory consumables" (200-dimensional vector).

[0113] Model inference: Output preference probability: Consumables = 0.75, Data = 0.20, Other browsing directions < 0.05.

[0114] Recommended output: Top-3 browsing categories: consumables, documents, reagents.

[0115] Example 3: User searches for "industry application cases": Input: User identity information: U3003, Search text: "Industry application cases".

[0116] Feature calculation: User identity characteristics: City = "Guangzhou" (encoded value 3), Gender = "Male" (encoded value 1), Education = "Doctor" (encoded value 5), User type = "Educational institution" (encoded value 4).

[0117] User historical behavior characteristics: No historical behavior within the past six months Search text features: word vectors for “industry application cases”.

[0118] Model Inference: Output Preference Probabilities: Industry Applications = 0.90, Information = 0.08, Other Browsing Directions < 0.02.

[0119] Recommended browsing directions: Top-3 browsing areas: industry applications, news, and communities.

[0120] This application's embodiments adjust the final weights of each feature by adjusting the total number of times each feature is used as a split point in all trees, thereby enabling the XGBoost model to be better integrated into the information recommendation field and improving recommendation accuracy.

[0121] This application successfully solves the problem of personalized and high-precision recommendation in vertical domain comprehensive search by using a carefully designed feature fusion method, a targeted model optimization strategy, and a dynamic user interest quantification algorithm, which has significant technological progress and application value.

[0122] Based on the above, this application also provides an information recommendation device in the field of scientific instruments. The information recommendation device in the field of scientific instruments can be referred to in correspondence with the aforementioned information recommendation method in the field of scientific instruments.

[0123] As an example, such as Figure 8 As shown, information recommendation devices in the field of scientific instruments include: The receiving module 810 is used to receive a search request. The search request includes first user identity information and first search information. The first user identity information corresponds to first user identity features. The first user identity features include at least one of the following: the user's city, the user's gender, the user's education level, the user type, and the user's basic profile. The prediction module 820 is used to obtain the predicted probability of each browsing direction based on the first user identity information and the first search information; Output module 830 is used to output the preset number of browsing directions with the highest predicted probability.

[0124] This application embodiment improves the predictive model's ability to discern user intent by integrating user identity features and search information, thereby increasing recommendation accuracy and optimizing user experience.

[0125] In one possible implementation, the information recommendation device in the field of scientific instruments also includes: The behavior feature determination module 840 is used to extract the behavior data of the first user within a preset time period before the current time based on the first user's identity information, and to determine the behavior features of the first user based on the behavior data of the first user. Based on this, the prediction module is used to obtain the predicted probability of each browsing direction based on the first user's identity information, the first search information, and the first user's behavioral characteristics.

[0126] This application's embodiments improve the predictive model's ability to discern user intent by integrating user identity features, historical user behavior features, and search information, making the recommendation results more aligned with user preferences, achieving dynamic personalized recommendations, and improving recommendation accuracy.

[0127] In one possible implementation, the behavior feature determination module 840 is specifically used for: The first user behavior data is categorized according to the actual browsing direction clicked in each first historical behavior; each browsing direction contains multiple events. For each browsing direction, the historical behavior score of the browsing direction is determined based on the behavior time of all first historical behaviors of the browsing direction and the actual click events corresponding to all first historical behaviors. The historical behavior scores from all browsing directions constitute the first user behavior characteristic.

[0128] This application's embodiments determine the weight of each type of event based on its value, thereby determining the historical behavior score for each browsing direction. This allows the prediction model to learn the contribution of each type of event to each browsing direction. During the prediction process, the model predicts the probability of each browsing direction being clicked at a finer granular level, thereby improving user satisfaction metrics.

[0129] In one possible implementation, the prediction module 820 is specifically used for: The first user identity information, the first search information, and the first user behavior features are respectively transformed into first feature vectors; The prediction model is input with the first user identity information, the first search information, and the first feature vector corresponding to the first user behavior features to obtain the prediction probability for each browsing direction.

[0130] This application embodiment improves the model's running speed by converting user identity information, search information, and user behavior characteristics into feature vectors that the model can recognize, and then inputting them into the prediction model.

[0131] In one possible implementation, the information recommendation device in the field of scientific instruments further includes a training module 850, which is used for: Obtain the basic dataset. In the basic dataset, each piece of basic data includes the second user behavior characteristics based on the same second historical behavior, the second historical behavior information, the actual browsing direction of the click in the second historical behavior, and the second user identity characteristics. All second user behavior features, all second historical behavior information, and all second user identity features are converted into corresponding second feature vectors, and all browsing directions are converted into corresponding tag vectors. The second feature vectors corresponding to the second user behavior feature, second historical behavior information, and second user identity feature of each piece of basic data are weighted and concatenated, and then concatenated with the label vector of the actual browsing direction clicked in the basic data to form a sample corresponding to the basic data. All samples are then combined to form a training dataset. In the training dataset, the weights of the second user behavior feature, second historical behavior information, and each second user identity feature are the initial weights. The prediction model is iteratively trained using the training dataset. After training, the final weights of user behavior features, search information, and user identity features are obtained.

[0132] This application embodiment, by integrating multi-source features and adjusting the weights of various features through model training, can ensure that the recommendation accuracy reaches the optimal level.

[0133] Figure 9 This is a schematic diagram of the structure of the electronic device provided in this application, such as... Figure 9 As shown, the electronic device may include: a processor 910, a communications interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communications interface 920, and the memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions in the memory 930 to execute an information recommendation method in the field of scientific instruments, the method including: Receive a search request, which includes first user identity information and first search information. The first user identity information corresponds to first user identity features, which include at least one of the following: user's city, user's gender, user's education level, user type, and user basic profile. The predicted probabilities for each browsing direction are obtained based on the first user's identity information and the first search information; Output the preset number of browsing directions with the highest predicted probability.

[0134] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] On the other hand, this application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer is able to execute the information recommendation method in the field of scientific instruments provided in the above embodiments. The method includes: Receive a search request, which includes first user identity information and first search information. The first user identity information corresponds to first user identity features, which include at least one of the following: user's city, user's gender, user's education level, user type, and user basic profile. The predicted probabilities for each browsing direction are obtained based on the first user's identity information and the first search information; Output the preset number of browsing directions with the highest predicted probability.

[0136] In another aspect, this application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the information recommendation method in the field of scientific instruments provided in the above embodiments, the method comprising: Receive a search request, which includes first user identity information and first search information. The first user identity information corresponds to first user identity features, which include at least one of the following: user's city, user's gender, user's education level, user type, and user basic profile. The predicted probabilities for each browsing direction are obtained based on the first user's identity information and the first search information; Output the preset number of browsing directions with the highest predicted probability.

[0137] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0138] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An information recommendation method in the field of scientific instruments, characterized by, The method comprises the following steps: receiving a search request, the search request comprising first user identity information and first search information, the first user identity information corresponding to first user identity features, the first user identity features comprising at least one of the following: city where the user is located, gender of the user, education of the user, user type, and user basic portrait; obtaining a prediction probability of each browsing direction based on the first user identity information and the first search information; outputting a preset number of browsing directions with the highest prediction probability.

2. The information recommendation method according to claim 1, wherein After receiving the search request, the method further comprises the following steps: extracting first user behavior data in a preset time period before the current time based on the first user identity information, and determining first user behavior features based on the first user behavior data; and obtaining a prediction probability of each browsing direction based on the first user identity information, the first search information, and the first user behavior features.

3. The information recommendation method according to claim 2, wherein The step of determining the first user behavior features based on the first user behavior data comprises the following steps: classifying the first user behavior data according to the browsing direction actually clicked in each first historical behavior in the first user behavior data; wherein a plurality of events are provided in each browsing direction; for each browsing direction, determining a historical behavior score of the browsing direction based on the behavior time of all first historical behaviors of the browsing direction and the actual click events corresponding to all first historical behaviors; the historical behavior scores of all browsing directions form the first user behavior features.

4. The information recommendation method according to claim 2, wherein The step of obtaining a prediction probability of each browsing direction based on the first user identity information, the first search information, and the first user behavior features comprises the following steps: converting the first user identity information, the first search information, and the first user behavior features into first feature vectors respectively; inputting the first feature vectors corresponding to the first user identity information, the first search information, and the first user behavior features into a prediction model to obtain a prediction probability of each browsing direction.

5. The information recommendation method according to claim 4, wherein The step of training the prediction model comprises the following steps: obtaining a basic data set, wherein each piece of basic data in the basic data set comprises second user behavior features, second historical behavior information, a browsing direction actually clicked in the second historical behavior, and second user identity features based on the same second historical behavior; converting all second user behavior features, all second historical behavior information, and all second user identity features into corresponding second feature vectors, and converting all browsing directions into corresponding label vectors; weighting and splicing the second feature vectors corresponding to the second user behavior features, the second historical behavior information, and the second user identity features of each piece of basic data, and splicing the label vector of the browsing direction actually clicked in the basic data to form a sample corresponding to the basic data, and grouping all samples into a training data set; wherein the weights of the second user behavior features, the second historical behavior information, and each second user identity feature in the training data set are initial weights. The prediction model is iteratively trained using the training data set, and after the training is completed, the final weight of the user behavior feature, the final weight of the search information, and the final weight of the user identity feature are obtained.

6. The information recommendation method according to claim 5, wherein The prediction model is an extreme gradient boosting model, and the final weight of the user behavior feature, the final weight of the search information, and the final weight of the user identity feature are respectively determined by the total number of times that the user behavior feature, the search information, and the user identity feature are used as a split point in all trees when the training is completed.

7. An information recommendation device in the field of scientific instruments, characterized by comprising: Comprise: The receiving module is configured to receive a search request, wherein the search request comprises first user identity information and first search information, the first user identity information corresponds to a first user identity feature, and the first user identity feature comprises at least one of a city where a user is located, a gender of the user, an education background of the user, a user type, and a user basic portrait. The prediction module is configured to obtain a prediction probability of each browsing direction based on the first user identity information and the first search information. The output module is configured to output a preset number of browsing directions with the highest prediction probability.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein The computer program is executed by the processor to implement the information recommendation method in the field of scientific instruments according to any one of claims 1 to 6. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, wherein, The computer program is executed by the processor to implement the information recommendation method in the field of scientific instruments according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the information recommendation method in the field of scientific instruments according to any one of claims 1 to 6.

Images