Project recommendation method applied to automatic artificial intelligence

Abstract

The invention discloses a project recommendation method applied to automatic artificial intelligence, and the method comprises the following steps: collecting multi-dimensional user data, and constructing a user interest hierarchy model; a reinforcement learning model is introduced based on a user interest hierarchy model, user interaction feedback is used as a reward signal, and a recommendation strategy is continuously optimized; and in combination with a user click behavior under the influence of a project recommendation strategy, mining a hidden demand of the user for a potential project, and according to a mining result of the hidden demand of the user, adjusting a hierarchy weight in the interest hierarchy model in real time. The project recommendation strategy is adjusted through an automatic feedback mechanism, user interaction data generated by a new recommendation strategy is used as input, and hidden interest mining and hierarchical weight adjustment are executed circularly. According to the method, the invisible demands of the user are accurately mined through hierarchical clustering and knowledge graph verification, and the recommendation accuracy can be remarkably improved and the information cocoon room can be effectively broken through in combination with dynamic hierarchical weight adjustment and reinforcement learning multi-strategy adaptive mechanism.

Description

Technical Field

[0001] The present invention relates to the field of computer software technology, and in particular to a project recommendation method applied to automatic artificial intelligence. Background Art

[0002] In the current AI-driven project recommendation landscape, traditional recommendation systems primarily rely on explicit user behavior data (such as ratings and clicks) to build interest models. However, this approach has significant limitations. Specifically, a large number of users' potential needs exist in an implicit form, and traditional collaborative filtering algorithms, relying on explicit interaction data, struggle to capture these needs. This leads to homogenized recommendation results and exacerbates the "information cocoon" effect. Furthermore, existing dynamic interest modeling methods, which often employ linear decay or simple sliding window mechanisms, have inherent flaws in balancing short-term preferences with long-term stable interests. Short-term noise (such as accidental clicks) can easily interfere with long-term interest representations, while fixed weight allocation mechanisms cannot adapt to the dynamic nature of interest drift, ultimately resulting in reduced recommendation accuracy. Summary of the Invention

[0003] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of this application to avoid obscuring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions should not be used to limit the scope of the present invention.

[0004] To solve the above technical problems, the present invention provides the following technical solution: a project recommendation method applied to automatic artificial intelligence, comprising the following steps:

[0005] Collect multi-dimensional user data, build a user interest hierarchy model, and generate an initial recommendation strategy based on the user interest hierarchy model;

[0006] Based on the user interest hierarchy model, a reinforcement learning model is introduced, user interaction feedback is used as a reward signal, and a multi-strategy adaptive reinforcement learning method is combined to continuously optimize the recommendation strategy;

[0007] Combined with the user click behavior under the influence of the project recommendation strategy, the implicit demand of users for potential projects is explored. Based on the mining results of users' implicit demand, the hierarchical weights in the interest hierarchy model are adjusted in real time.

[0008] The project recommendation strategy is adjusted through an automated feedback mechanism, and the user interaction data generated by the new recommendation strategy is used as input to cyclically perform implicit interest mining and hierarchical weight adjustment to form a self-evolving recommendation system.

[0009] As a preferred solution of the project recommendation method applied to automatic artificial intelligence described in the present invention, wherein: the construction of the user interest hierarchy model includes dividing the user interest hierarchy based on the collected multi-dimensional user data, dividing the user interest hierarchy into short-term project preferences, long-term project preferences, professional fields to which the projects belong, and project interaction behavior patterns, and assigning initial hierarchy weights to each hierarchy.

[0010] As a preferred solution of the project recommendation method applied to automatic artificial intelligence of the present invention, a hierarchical clustering method is used to mine the user's implicit demand for potential projects, specifically comprising the following steps:

[0011] The user's interaction frequency, stay time, and cross-category jump paths for the project are used as feature vectors;

[0012] The Ward variance minimization criterion was used to calculate the inter-class distance to perform clustering;

[0013] When the silhouette coefficient of the clustering result is greater than the first set threshold, the Calinski-Harabasz index improvement relative to the previous iteration is less than the second threshold, and the number of iterations reaches the upper limit of 100, clustering is stopped and the clustering result is output;

[0014] Through knowledge graph verification, accurate identification of implicit interests is obtained, and the obtained implicit interests are injected into the corresponding level according to the similarity ratio.

[0015] As a preferred solution of the project recommendation method applied to automatic artificial intelligence of the present invention, the knowledge graph verification includes the following steps:

[0016] Build a knowledge graph containing users, projects, categories and their relationships based on users' historical interaction data, project information and external knowledge sources;

[0017] For uninteracted items, calculate their relevance to the user's interests at all levels in the knowledge graph;

[0018] According to the calculated correlation, uninteracted items with high correlation with user interests are screened out as potential implicit interest items;

[0019] Based on the calculated correlation, the non-interacted items with high correlation with the user's interests are screened out through the screening mechanism as potential implicit interest items. The potential implicit interest items are injected into the corresponding level, and the weights of each level in the user interest hierarchy model are dynamically updated.

[0020] As a preferred solution of the project recommendation method applied to automatic artificial intelligence of the present invention, the correlation calculation includes the following steps:

[0021] Calculate the cosine similarity between users and items:

[0022]

[0023] Introduce the hierarchical correction coefficient to calculate the correlation between users and projects:

[0024]

[0025] in:

[0026] Represents the user's interest characteristics;

[0027] A feature vector representing an item;

[0028] Represents the user's short-term interest vector, which represents the user's recent interest preferences;

[0029] Represents the user's professional field interest vector, which represents the user's long-term and stable interest in a certain professional direction;

[0030] Represents the cosine similarity between users' short-term interests and items;

[0031] Represents the cosine similarity between the user's professional interests and the project;

[0032] α represents the hierarchical correction coefficient of short-term interest, which indicates the weight of the impact of short-term interest on matching degree;

[0033] β is the hierarchical correction coefficient of professional field interest, which represents the influence weight of professional field interest on matching degree;

[0034] max(·) means taking the maximum value of the short-term interest matching and the professional field interest matching, so as to adjust the corresponding level when updating the user interest hierarchy model in the future;

[0035] if Higher than This indicates that implicit needs are closer to the user's short-term interest layer; otherwise, they are closer to the professional field interest layer.

[0036] As a preferred solution of the project recommendation method applied to automatic artificial intelligence of the present invention, the screening mechanism includes the following rules:

[0037] when When , the uninteracted items are injected into the short-term interest layer and trigger the short-term weight update;

[0038] when When the interaction is complete, the non-interaction items will be added to the corresponding categories of the professional field layer, and the professional field layer weight adjustment will be triggered; the layer weight adjustment will be triggered;

[0039] For non-interactive items that meet both conditions, priority is given to The corresponding layer and trigger weight update.

[0040] As a preferred solution of the project recommendation method applied to automatic artificial intelligence of the present invention, the weight update formula is:

[0041] When the invisible interest item i is satisfied When , the short-term interest weight update formula is:

[0042]

[0043]

[0044] Where α is the short-term correction coefficient; The correlation between the hidden interest items and the short-term interest layer; active is the number of active days of the sliding window of the invisible interest item; T decay Decay period; ΔW short is the short-term layer weight increment; is the weight of the short-term interest layer before adjustment; is the adjusted short-term interest layer weight;

[0045] When implicit interest item i satisfies When , the weight update formula of professional field layer is:

[0046]

[0047] Where ΔW domain is the weight increment of the professional field layer; β is the professional field correction coefficient; Indicates category c i The number of hidden interest items; N total Indicates the total number of projects at the professional level; is the weight of the professional field layer before adjustment; is the adjusted professional field layer weight;

[0048] The long-term interest weight update formula is:

[0049]

[0050] Where ΔW long is the long-term layer weight increment; k is the domain association factor, ranging from [0,1], and the k value is calculated through the semantic distance of the knowledge graph. d is the shortest path length of the knowledge graph. When d ≥ the set value, k = 0 to block the transfer of irrelevant interests. τ is the smoothing coefficient, which is used to control the smoothness of the transfer process. is the long-term interest weight before adjustment; is the adjusted long-term interest layer weight.

[0051] As a preferred solution of the project recommendation method applied to automatic artificial intelligence of the present invention, after completing the update of the weights of each level, normalizing the weights of all levels is performed, which specifically includes the following steps:

[0052] Calculate the sum of the updated weights of each layer;

[0053] The weight value of each level is divided by the sum to obtain a normalized weight value so that the sum of the weights of all levels is equal to 1.

[0054] As a preferred solution of the project recommendation method applied to automatic artificial intelligence described in the present invention, the values of the short-term correction coefficient α and the long-term correction coefficient β are determined by grid search, wherein the value range of α is 0.6-0.8, the value range of β is 0.8-1.0, and the constraint condition α+β≤1.5 is satisfied.

[0055] The present invention also provides a project recommendation system applied to automatic artificial intelligence, which is applied to a project recommendation method of automatic artificial intelligence, including:

[0056] The data collection and preprocessing module is responsible for collecting multi-dimensional user data and preprocessing the data to provide a basis for subsequent modeling;

[0057] The user interest hierarchical modeling module builds a multi-level user interest model based on the collected multi-dimensional user data. It divides user interests into short-term project preference layer, long-term project preference layer, project professional field layer, and project interaction behavior pattern layer, and sets initial weight values for each layer to accurately characterize the user interest structure.

[0058] The recommendation strategy generation module builds an item recommendation strategy that matches the user's interest level based on the user's interest level model;

[0059] The implicit demand mining module combines the user's interactive behavior under the influence of the recommendation strategy and uses a hierarchical clustering method to identify the user's potential needs, extract the user's implicit interest preferences for uninteracted items, and verify them through the knowledge graph;

[0060] The hierarchical weight dynamic adjustment module dynamically updates the weights in the user's interest hierarchy in real time based on the similarity matching results between implicit interest items and each interest hierarchy;

[0061] The recommendation feedback loop mechanism module uses the user interaction data generated by the new recommendation strategy as new input and feeds it back to the implicit demand mining module and the hierarchical weight adjustment module, thus realizing a closed-loop structure of recommendation strategy-user feedback-model update, forming an intelligent recommendation system with self-evolution capabilities.

[0062] The reinforcement learning recommendation optimization module introduces a reinforcement learning model and uses user interaction feedback as a reward signal. It combines a multi-strategy adaptive reinforcement learning algorithm to dynamically optimize the recommendation strategy and form a recommendation behavior that can be continuously learned and evolved.

[0063] Beneficial effects of the present invention:

[0064] 1. Through hierarchical clustering and knowledge graph verification, the present invention can accurately explore users' hidden needs, significantly improve the coverage rate of non-interacted items, effectively break through the information cocoon, and realize the precise dynamic expansion of hidden interests.

[0065] 2. By dynamically adjusting the weights of user interest levels and introducing a sliding window attenuation mechanism and semantic correlation factors, the present invention can accurately balance short-term preferences with long-term stable interests, effectively improving user satisfaction and reducing invalid repeated recommendations.

[0066] 3. The present invention can significantly improve the recommendation accuracy by combining reinforcement learning with a multi-strategy adaptive mechanism and dynamically adjusting the recommendation strategy using real-time interactive feedback from users. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0068] Figure 1 This is a flowchart of a project recommendation method applied to automatic artificial intelligence of the present invention. DETAILED DESCRIPTION

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0070] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0071] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0072] Furthermore, the present invention is described in detail with reference to schematic diagrams. For ease of illustration, when describing the embodiments of the present invention, cross-sectional views illustrating device structures may be partially enlarged and not to scale. Furthermore, the schematic diagrams are merely illustrative and should not limit the scope of protection of the present invention. Furthermore, in actual production, the three-dimensional dimensions of length, width, and depth should be included.

[0073] Example 1

[0074] Reference Figure 1 , as an embodiment of the present invention, provides a project recommendation method applied to automatic artificial intelligence, comprising the following steps:

[0075] S1: Collect multi-dimensional user data, build a user interest hierarchy model, and generate an initial recommendation strategy based on the user interest hierarchy model.

[0076] Specifically, multi-dimensional user data includes user interest in historically recommended items, item interaction records (browsing, clicking, favorites, etc.), item browsing behavior habits (such as browsing time, browsing duration, browsing frequency, etc.), item click-through rate, and external environmental data associated with the items (such as holidays, social hot spots, regional events, etc.). After collecting multi-dimensional data, the data is pre-processed, including but not limited to data cleaning, denoising, and normalization, to ensure data quality. Feature engineering is also performed on different data types, including but not limited to normalization, dimensionality scaling, and feature selection, to extract the most representative and predictive features, laying a solid foundation for the subsequent construction of an accurate user interest hierarchy model.

[0077] Specifically, a user interest hierarchy model is constructed, including dividing the user interest hierarchy based on the collected multi-dimensional user data, and dividing the user interest hierarchy into short-term project preferences (interests with greater recent interaction impact), long-term project preferences (long-term stable interest preferences), project professional fields (users' in-depth attention to certain specific professional directions) and project interaction behavior patterns (users' decision-making methods, active periods, etc.), and assigning corresponding hierarchical weights to each level.

[0078] It's important to note that the short-term item preference layer captures a user's most recent N (N=20) item interactions, weighted by a time-decay coefficient to reflect immediate changes in interest. The long-term item preference layer calculates the category distribution of all historical items a user interacts with, smoothing it to create a steady-state interest profile. The professional domain layer identifies verticals where user interaction frequency exceeds 30% of total interactions and average dwell time exceeds the 90th percentile of users in the same field on the platform. The behavioral pattern layer analyzes user decision-making characteristics, covering active time distribution (e.g., nighttime activity) and interaction path patterns (e.g., the "browse-save-buy" sequence). The initial weighting of each layer is 4:3:2:1 for short-term:long-term:professional:behavioral.

[0079] S2: Based on the user interest hierarchy model, a reinforcement learning model is introduced, user interaction feedback is used as the reward signal, and a multi-strategy adaptive reinforcement learning method is combined to continuously optimize the recommendation strategy.

[0080] S3: Combined with the user's click behavior under the influence of the project recommendation strategy, the user's implicit demand for potential projects is explored. According to the mining results of the user's implicit demand, the hierarchical weights in the interest hierarchy model are adjusted in real time.

[0081] Specifically, the hierarchical clustering method is used to mine users' implicit needs for potential projects, including the following steps:

[0082] S31: The user's interaction frequency, stay time, and cross-category jump path are used as feature vectors;

[0083] S32: The distance between classes is calculated using the Ward variance minimization criterion to perform clustering. The calculation formula is as follows:

[0084] Among them, D(A,B) is the inter-class distance, which is used to measure the similarity between two classes (clusters) A and B; is the center vector (center of gravity) of cluster A, is the center vector of cluster B, and is the mean vector of all data points in the cluster; The square of the Euclidean distance is used to calculate the distance between the center points of cluster A and cluster B, which means the similarity of the center points between the two clusters. The smaller the distance, the more similar they are. -1 +B -1 Used to normalize the distance between classes to avoid deviations caused by uneven cluster sizes.

[0085] The above formula is used to measure the increase in intra-class variance after cluster A and cluster B are merged. The specific example is as follows:

[0086] Suppose we have two clusters, A and B. Cluster A represents user interaction data for technology projects, while cluster B represents user interaction data for lifestyle projects. If the center vector of cluster A (i.e., the mean vector of all interaction data for technology projects) and the center vector of cluster B (i.e., the mean vector of all interaction data for lifestyle projects) are very close in feature space—for example, if their values for the three features of interaction frequency, dwell time, and cross-category jump paths are similar—then we can say that clusters A and B have similar centroids. In this case, according to the Ward variance minimization criterion, merging clusters A and B will result in a relatively small increase in the intra-cluster variance, indicating strong similarity in the interaction behavior patterns of clusters A and B. By identifying similarities in user interaction patterns, users with similar behavior patterns can be grouped together, enabling more accurate recommendations. For example, if users in clusters A and B have similar interaction behaviors, successful recommendation strategies for cluster A users can also be applied to cluster B users, improving recommendation accuracy and user satisfaction.

[0087] Normalization factor A -1 +B -1 We further adjust for the effects of uneven cluster sizes to make distance calculations more fair. This allows us to more accurately measure the similarities between clusters, leading to more effective hierarchical clustering.

[0088] S33: When the silhouette coefficient of the clustering result is greater than the first set threshold (0.65, determined based on the grid search of the MovieLens dataset), the Calinski-Harabasz index improvement relative to the previous iteration is less than the second threshold (5%), and the number of iterations reaches the upper limit of 100, clustering is stopped and the clustering result is output.

[0089] It should be noted that the silhouette coefficient is used to assess clustering quality, with higher values indicating closer clustering. As a metric for evaluating clustering effectiveness, the silhouette coefficient is a well-established method in the art and widely used in fields such as data mining and machine learning. We will not discuss this in detail here. The Calinski-Harabasz index is also an internal metric used to assess clustering effectiveness. It measures clustering quality by calculating the ratio of the between-cluster variance to the within-cluster variance. Many programming languages and data analysis tools provide functions for calculating the Calinski-Harabasz index, allowing users to quickly assess the quality of clustering results. We will not discuss this in detail here.

[0090] S34: Through knowledge graph verification, accurate identification of implicit interests is obtained, and the obtained implicit interests are injected into the corresponding level of S11 according to the similarity ratio (if the professional field similarity is high, the professional field layer is updated; if the short-term similarity is high, the short-term interest layer is updated), and user interests are grouped and classified.

[0091] Specifically, knowledge graph verification includes the following steps:

[0092] S34-1: Build a knowledge graph containing users, projects, categories, and their relationships based on historical user interaction data, project information, and external knowledge sources;

[0093] S34-2: For uninteracted items (no browsing, clicking, favorites, purchases, ratings, etc.), calculate the correlation between them and the user's interests at each level (short-term project preferences, long-term project preferences, and the professional fields to which the projects belong) in the knowledge graph. Specifically, vectors are used to represent users and items, and the correlation is calculated using methods such as cosine similarity. At the same time, hierarchical correction coefficients (such as the short-term project preference correction coefficient α) and the professional field correction coefficient β are introduced to adjust the influence weight of different levels of interests;

[0094] S34-3: Based on the calculated correlation, select the non-interacted items that are highly correlated with the user's interests as potential implicit interest items.

[0095] It should be noted that the knowledge graph is a semantic network that uses nodes to represent entities (such as users, projects, and categories) and edges to represent relationships between entities (such as the interaction between users and projects, and the subordinate relationship between projects and categories). By building a knowledge graph, scattered data can be integrated to form a structured knowledge representation, providing a foundation for subsequent interest mining and recommendation. When calculating the relevance, for example:

[0096] Assume that the user vector is represented as (representing the user's interest characteristics), the item vector is represented as (representing the feature vector of the item), the cosine similarity between the user and the item is expressed as:

[0097]

[0098] After introducing the hierarchical correction coefficient, the formula for calculating the association between users and projects is:

[0099]

[0100] if Higher than This indicates that implicit needs are closer to the user's short-term interest layer; otherwise, they are closer to the professional field interest layer.

[0101] in, Represents the user's short-term interest vector, which represents the user's recent interest preferences; Represents the user's professional field interest vector, which represents the user's long-term and stable interest in a certain professional direction; Represents the cosine similarity between users' short-term interests and items; represents the cosine similarity between the user's professional field interests and the project; α is the hierarchical correction coefficient of short-term interests, which indicates the weight of the impact of short-term interests on the matching degree; β is the hierarchical correction coefficient of professional field interests, which indicates the weight of the impact of professional field interests on the matching degree; max(·) means taking the maximum value of the short-term interest matching degree and the professional field interest matching degree, so as to adjust the corresponding level when the user interest hierarchy model is subsequently updated.

[0102] More specifically, the values of the short-term correction coefficient α and the long-term correction coefficient β were determined through grid search, with α ranging from 0.6 to 0.8 and β ranging from 0.8 to 1.0, satisfying the constraint α + β ≤ 1.5. This setting has the advantage of preventing the model from being overly sensitive to noisy data due to excessive weight adjustments, thereby maintaining the model's generalization ability. Furthermore, it ensures that weight adjustments for the short-term interest layer and the professional domain layer do not excessively squeeze the weight space of other layers (such as the long-term interest layer and the behavioral pattern layer), maintaining the rationality of the four-layer weight structure.

[0103] S34-4: Based on the calculated correlation, the non-interacted items (hidden interests) with high correlation with the user's interests are screened out through the screening mechanism as potential implicit interest items, the potential implicit interest items are injected into the corresponding level, and the weights of each level in the user interest level model are dynamically updated.

[0104] Specifically, the screening mechanism includes the following rules:

[0105] when When (θ = 0.65, the value is determined by AB testing), the uninteracted items are injected into the short-term interest layer and trigger the short-term weight update;

[0106] when When (θ=0.8), the non-interacted items are added to the corresponding categories of the professional field layer, and the professional field layer weight adjustment is triggered; the layer weight adjustment is triggered;

[0107] For non-interaction items that meet both conditions, that is, when θ>0.8, priority is given to The corresponding layer and trigger weight update.

[0108] The weight update formula is:

[0109] When the invisible interest item i is satisfied When , the short-term interest weight adjustment formula is:

[0110]

[0111] Where: α is the short-term correction coefficient (default is 0.7); The correlation between the hidden interest items and the short-term interest layer; active is the number of active days of the sliding window of the invisible interest item; T decay Decay period (default 7 days); ΔW short is the short-term layer weight increment; The weight of the short-term interest layer before adjustment, with an initial value of 0.4 (accounting for 40%); The adjusted short-term interest layer weight is capped at 0.5 (accounting for ≤50%).

[0112] Therefore, if ΔW short =0.2→new weight min(0.5,0.4×1.2)=0.48.

[0113] When implicit interest item i satisfies When , the professional field layer weight adjustment formula is:

[0114]

[0115] Where ΔW domain is the weight increment of the professional field layer; β is the professional field correction coefficient (default is 0.9); Indicates category c i The number of hidden interest items; N total Indicates the total number of projects at the professional level; The weight of the professional field layer before adjustment, with an initial value of 0.2 (accounting for 20%); is the adjusted weight of the professional field layer.

[0116] For example, user A's professional domain layer includes the following categories:

[0117] Coffee utensils (subcategories: hand brewer, moka pot, bean grinder), then

[0118] Outdoor equipment (sub-categories: tents, hiking shoes), then

[0119] If the newly added implicit interest item "coffee filter paper" has a correlation of 0.85 with "coffee utensils", then:

[0120]

[0121] New weights:

[0122] The long-term interest weight update formula is:

[0123]

[0124] in:

[0125] ΔW long is the long-term layer weight increment; k is the domain association factor, ranging from [0,1]. If the user frequently browses "drones" in the short term and has "photographic equipment" in the long term, the k value will be higher (such as 0.8). The k value is calculated through the semantic distance of the knowledge graph. d is the shortest path length in the knowledge graph. When d ≥ the set value, k = 0 to block irrelevant interest transfers. For example, the short-term interest is a moka pot; the long-term interest is espresso (historical purchases). The knowledge graph path is: moka pot → espresso utensil → espresso (d = 2), so k = 1 / (1 + 2) ≈ 0.33.

[0126] τ is the smoothing coefficient, which is used to control the smoothness of the transfer process. The value of τ is determined by grid search, and the optimal value is 0.5; The long-term interest weight before adjustment is 0.3 (30%). is the adjusted long-term interest layer weight.

[0127] Furthermore, after completing the weight update of each level, the weights of all levels are normalized, which specifically includes the following steps:

[0128] Calculate the sum of the updated weights of each layer;

[0129] The weight value of each layer is divided by the total to obtain the normalized weight value so that the sum of the weights of all layers is equal to 1.

[0130] S4: Adjust the project recommendation strategy through an automated feedback mechanism, and use the user interaction data generated by the new recommendation strategy as input to cyclically execute the implicit interest mining and hierarchical weight adjustment of S3 to form a self-evolving recommendation system.

[0131] Specifically, the automated feedback mechanism includes the following steps:

[0132] After the recommendation strategy is generated, actual user interaction data on the recommended items is continuously collected, including but not limited to click behavior, browsing time, bounce rate, item collection / like behavior, and cross-category access paths. Based on this new user interaction data, the system repeatedly executes the implicit interest mining process. This includes updating the interaction feature vector, performing hierarchical clustering, dynamically evaluating the silhouette coefficient and Calinski-Harabasz index to see if they meet the set termination criteria, and verifying the relevance and stability of the mined potential interests through the knowledge graph.

[0133] The mined implicit interests are matched based on their similarity with the interest layers in the existing hierarchical model. The weights of the short-term interest layer, professional field layer, and long-term interest layer are incrementally adjusted according to the corresponding weight update formulas. At the same time, various correction factors (such as the short-term correction coefficient, professional field correction coefficient, and field correlation factor) are updated to ensure the model's sensitivity to new interests and the continued maintenance of stable interests.

[0134] After all weight adjustments are completed, the system normalizes the weights of the user interest hierarchy model so that the sum of the weights at each level is 1, thereby maintaining the model's numerical stability and structural consistency. Based on the newly updated interest model, the recommendation strategy will be updated in real time, selecting the optimal action through a reinforcement learning model to optimize subsequent recommendation results.

[0135] The user interaction data generated by the new recommendation strategy is automatically used as the input for the next round of implicit interest mining and weight adjustment, realizing an automatic closed loop of "recommendation-feedback-mining-adjustment". This enables continuous evolution and dynamic optimization of recommendation accuracy without human intervention, and ultimately builds an intelligent recommendation system with self-evolution capabilities.

[0136] In summary, the present invention, through hierarchical clustering and knowledge graph verification, can accurately tap into users' hidden needs, significantly improve the coverage of uninteracted items, effectively break through information cocoons, and achieve accurate dynamic expansion of hidden interests. The present invention dynamically adjusts the weights of user interest levels and introduces a sliding window attenuation mechanism and semantic correlation factors, which can accurately balance short-term preferences with long-term stable interests, effectively improve user satisfaction and reduce invalid repeated recommendations. The present invention combines reinforcement learning with a multi-strategy adaptive mechanism and uses real-time user interaction feedback to dynamically adjust the recommendation strategy, which can significantly improve the recommendation accuracy.

[0137] Example 2, an embodiment of the present invention, provides an automatic artificial intelligence project recommendation system, including:

[0138] The data collection and preprocessing module is responsible for collecting multi-dimensional user data and preprocessing the data to provide a basis for subsequent modeling;

[0139] The user interest hierarchical modeling module builds a multi-level user interest model based on the collected multi-dimensional user data. It divides user interests into short-term project preference layer, long-term project preference layer, project professional field layer, and project interaction behavior pattern layer, and sets initial weight values for each layer to accurately characterize the user interest structure.

[0140] The recommendation strategy generation module builds an item recommendation strategy that matches the user's interest level based on the user's interest level model;

[0141] The implicit demand mining module combines the user's interactive behavior under the influence of the recommendation strategy and uses a hierarchical clustering method to identify the user's potential needs, extract the user's implicit interest preferences for uninteracted items, and verify them through the knowledge graph;

[0142] The hierarchical weight dynamic adjustment module dynamically updates the weights in the user's interest hierarchy in real time based on the similarity matching results between implicit interest items and each interest hierarchy;

[0143] The recommendation feedback loop mechanism module uses the user interaction data generated by the new recommendation strategy as new input and feeds it back to the implicit demand mining module and the hierarchical weight adjustment module, thus realizing a closed-loop structure of recommendation strategy-user feedback-model update, forming an intelligent recommendation system with self-evolution capabilities.

[0144] The reinforcement learning recommendation optimization module introduces a reinforcement learning model and uses user interaction feedback as a reward signal. It combines a multi-strategy adaptive reinforcement learning algorithm to dynamically optimize the recommendation strategy and form a recommendation behavior that can be continuously learned and evolved.

[0145] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A project recommendation method applied to automatic artificial intelligence, characterized in that: The following steps are involved: Collect multi-dimensional user data, build a user interest hierarchy model, and generate an initial recommendation strategy based on the user interest hierarchy model; Based on the user interest hierarchy model, a reinforcement learning model is introduced, user interaction feedback is used as a reward signal, and a multi-strategy adaptive reinforcement learning method is combined to continuously optimize the recommendation strategy; Combined with the user click behavior under the influence of the project recommendation strategy, the implicit demand of users for potential projects is explored. Based on the mining results of users' implicit demand, the hierarchical weights in the interest hierarchy model are adjusted in real time. The project recommendation strategy is adjusted through an automated feedback mechanism, and the user interaction data generated by the new recommendation strategy is used as input to cyclically perform implicit interest mining and hierarchical weight adjustment.

2. The project recommendation method for automatic artificial intelligence according to claim 1, characterized in that: The construction of the user interest hierarchy model includes dividing the user interest hierarchy based on the collected multi-dimensional user data, dividing the user interest hierarchy into short-term project preferences, long-term project preferences, professional fields to which projects belong, and project interaction behavior patterns, and assigning initial hierarchy weights to each hierarchy.

3. The project recommendation method for automatic artificial intelligence according to claim 1, characterized in that: The hierarchical clustering method is used to mine the user's implicit demand for potential projects, which specifically includes the following steps: The user's interaction frequency, stay time, and cross-category jump paths for the project are used as feature vectors; The Ward variance minimization criterion was used to calculate the inter-class distance to perform clustering; When the silhouette coefficient of the clustering result is greater than the first set threshold, the Calinski-Harabasz index improvement relative to the previous iteration is less than the second threshold, and the number of iterations reaches the upper limit of 100, clustering is stopped and the clustering result is output; Through knowledge graph verification, accurate identification of implicit interests is obtained, and the obtained implicit interests are injected into the corresponding level according to the similarity ratio.

4. The project recommendation method for automatic artificial intelligence according to claim 3, characterized in that: The knowledge graph verification includes the following steps: Build a knowledge graph containing users, projects, categories and their relationships based on users' historical interaction data, project information and external knowledge sources; For uninteracted items, calculate their relevance to the user's interests at all levels in the knowledge graph; According to the calculated correlation, uninteracted items with high correlation with user interests are screened out as potential implicit interest items; Based on the calculated correlation, the non-interacted items with high correlation with the user's interests are screened out through the screening mechanism as potential implicit interest items. The potential implicit interest items are injected into the corresponding level, and the weights of each level in the user interest hierarchy model are dynamically updated.

5. The project recommendation method for automatic artificial intelligence according to claim 4, characterized in that: The association calculation includes the following steps: Calculate the cosine similarity between users and items: Introduce the hierarchical correction coefficient to calculate the correlation between users and projects: in: Represents the user's interest characteristics; A feature vector representing an item; Represents the user's short-term interest vector, which represents the user's recent interest preferences; Represents the user's professional field interest vector, which represents the user's long-term and stable interest in a certain professional direction; Represents the cosine similarity between users' short-term interests and items; Represents the cosine similarity between the user's professional interests and the project; α represents the hierarchical correction coefficient of short-term interest, which indicates the weight of the impact of short-term interest on matching degree; β is the hierarchical correction coefficient of professional field interest, which represents the influence weight of professional field interest on matching degree; max(·) means taking the maximum value of the short-term interest matching and the professional field interest matching, so as to adjust the corresponding level when updating the user interest hierarchy model in the future; if Higher than This indicates that implicit needs are closer to the user's short-term interest layer; otherwise, they are closer to the professional field interest layer.

6. The project recommendation method for automatic artificial intelligence according to claim 4, characterized in that: The screening mechanism includes the following rules: when When , the uninteracted items are injected into the short-term interest layer and trigger the short-term weight update; when When the interaction is complete, the non-interaction items will be added to the corresponding categories of the professional field layer, and the professional field layer weight adjustment will be triggered; the layer weight adjustment will be triggered; For non-interactive items that meet both conditions, priority is given to The corresponding layer and trigger weight update.

7. The project recommendation method for automatic artificial intelligence according to claim 6, characterized in that: The weight update formula is: When the invisible interest item i is satisfied When , the short-term interest weight update formula is: Where α is the short-term correction coefficient; The correlation between the hidden interest items and the short-term interest layer; active is the number of active days of the sliding window of the invisible interest item; T decay Decay period; ΔW short is the short-term layer weight increment; is the weight of the short-term interest layer before adjustment; is the adjusted short-term interest layer weight; When implicit interest item i satisfies When , the weight update formula of professional field layer is: Where ΔW domain is the weight increment of the professional field layer; β is the professional field correction coefficient; Indicates category c i The number of hidden interest items; N total Indicates the total number of projects at the professional level; is the weight of the professional field layer before adjustment; is the adjusted professional field layer weight; The long-term interest weight update formula is: Where ΔW long is the long-term layer weight increment; k is the domain association factor, ranging from [0,1], and the k value is calculated through the semantic distance of the knowledge graph. d is the shortest path length of the knowledge graph. When d ≥ the set value, k = 0 to block the transfer of irrelevant interests. τ is the smoothing coefficient, which is used to control the smoothness of the transfer process. is the long-term interest weight before adjustment; is the adjusted long-term interest layer weight.

8. The project recommendation method for automatic artificial intelligence according to claim 7, characterized in that: After completing the weight update of each level, normalize the weights of all levels, which includes the following steps: Calculate the sum of the updated weights of each layer; The weight value of each level is divided by the sum to obtain a normalized weight value so that the sum of the weights of all levels is equal to 1.

9. The project recommendation method for automatic artificial intelligence according to claim 7, characterized in that: The values of the short-term correction coefficient α and the long-term correction coefficient β are determined by grid search, wherein the value range of α is 0.6-0.8, the value range of β is 0.8-1.0, and the constraint condition α+β≤1.5 is satisfied.

10. A project recommendation system for automatic artificial intelligence, based on the project recommendation method for automatic artificial intelligence according to any one of claims 1 to 9, characterized in that: include: The data collection and preprocessing module is responsible for collecting multi-dimensional user data and preprocessing the data to provide a basis for subsequent modeling; The user interest hierarchical modeling module builds a multi-level user interest model based on the collected multi-dimensional user data. It divides user interests into short-term project preference layer, long-term project preference layer, project professional field layer, and project interaction behavior pattern layer, and sets initial weight values for each layer to accurately characterize the user interest structure. The recommendation strategy generation module builds an item recommendation strategy that matches the user's interest level based on the user's interest level model; The implicit demand mining module combines the user's interactive behavior under the influence of the recommendation strategy and uses a hierarchical clustering method to identify the user's potential needs, extract the user's implicit interest preferences for uninteracted items, and verify them through the knowledge graph; The hierarchical weight dynamic adjustment module dynamically updates the weights of the user's interest hierarchy in real time based on the similarity matching results between implicit interest items and each interest hierarchy; The recommendation feedback loop mechanism module uses the user interaction data generated by the new recommendation strategy as new input and feeds it back to the implicit demand mining module and the hierarchical weight adjustment module, thus realizing a closed-loop structure of recommendation strategy-user feedback-model update, forming an intelligent recommendation system with self-evolution capabilities. The reinforcement learning recommendation optimization module introduces a reinforcement learning model and uses user interaction feedback as a reward signal. It combines a multi-strategy adaptive reinforcement learning algorithm to dynamically optimize the recommendation strategy and form a recommendation behavior that can be continuously learned and evolved.

Images