Multi-dimensional reverse mining and content recommendation method and device, and electronic equipment

Abstract

A multi-dimensional reverse mining and content recommendation method, apparatus, and electronic device are disclosed. The method involves a server establishing a primary interest vector based on user historical behavior data and executing a multi-dimensional mining program to obtain multiple reverse content objects, multiple bridging content objects, and multiple potential interest content objects. These objects are then collectively defined as multiple surprise candidate contents. Next, the surprise candidate contents are dynamically weighted and scored against similar candidate contents corresponding to the primary interest vector to generate a recommendation list. Finally, the primary interest vector and the dynamically weighted score are adjusted based on user feedback data regarding the recommendation list, thereby achieving the technical effect of increasing the diversity of the content recommendation list.

Description

Technical Field

[0001] This invention relates to a method, apparatus, and electronic device, particularly a method, apparatus, and electronic device for multi-dimensional reverse mining and content recommendation based on user habit models. Background Technology

[0002] In recent years, with the popularization and rapid development of mobile networks and streaming content platforms, various personalized content recommendation applications have sprung up like mushrooms after rain. Among them, content recommendation systems that improve user click-through rates and retention rates have attracted the most attention.

[0003] Generally, traditional content recommendation systems mostly rely on users' historical clicks or purchase records to calculate the similarity of content objects and push similar content. However, because this method tends to repeatedly push content that is highly consistent with users' existing preferences, it easily leads to homogenization of recommendation results and the formation of an echo chamber effect. Therefore, traditional content recommendation systems suffer from insufficient diversity in their content recommendation lists.

[0004] In light of this, some manufacturers have proposed introducing an exploration mechanism, which involves mixing a small amount of random or long-tail content into the recommendation list to enhance user engagement. However, this approach typically fails to structurally mine potential user interests and cross-domain connections, nor does it weight user hesitation behavior, thus remaining insufficient to address the lack of diversity in content recommendation lists.

[0005] In conclusion, it is clear that the existing technology has long suffered from insufficient diversity in content recommendation lists, thus necessitating the proposal of improved technical methods to address this issue. Summary of the Invention

[0006] This invention discloses a method, apparatus, and electronic device for multi-dimensional reverse mining and content recommendation based on a user habit model.

[0007] First, this invention discloses a multi-dimensional reverse mining and content recommendation method based on a user habit model, executed by a server connected to a user-end device and a database storing multiple content objects. The method comprises: the server acquiring historical behavioral data of the user from the user-end device to establish a primary interest vector for the corresponding user; the server executing a multi-dimensional mining program, which includes: calculating opposing regions in the feature space that are inversely related to the primary interest vector, and retrieving multiple reverse content objects corresponding to the opposing regions from the database; extracting multiple interest features with low correlation from the primary interest vector, and connecting the common attributes between each interest feature through a preset knowledge graph to retrieve multiple bridging content objects with cross-domain features from the database; and detecting user hesitation behavior data for specific content objects, and increasing the weight of the hesitation behavior data through an amplification algorithm to retrieve multiple potential interest content objects corresponding to the hesitation behavior data from the database. Next, the server defines the reverse content object, bridging content object, and potential interest content object as multiple surprise candidate contents, and dynamically weights and scores the surprise candidate contents with multiple similar candidate contents corresponding to the main interest vector to generate a content recommendation list and transmit it to the user's terminal device for display; and the server obtains the user's click feedback data on the content recommendation list to correct the parameters of the main interest vector and the dynamic weighted score.

[0008] Next, this invention also discloses a multi-dimensional reverse mining and content recommendation device based on a user habit model, characterized by comprising: a user profile modeling module, a multi-dimensional mining module, a candidate and scoring module, and a feedback and adaptive learning module. The user profile modeling module is used to acquire historical behavioral data of users from user-end devices and establish the corresponding user's main interest vector. The multi-dimensional mining module is connected to the user profile modeling module and is used to execute a multi-dimensional mining program, which includes: calculating the opposing regions in the feature space that are inversely related to the main interest vector, and retrieving multiple reverse content objects corresponding to the opposing regions from the database; extracting multiple interest features with low correlation from the main interest vector, and connecting the common attributes between each interest feature through a preset knowledge graph, in order to retrieve multiple bridging content objects with cross-domain features from the database; and detecting user hesitation behavior data for specific content objects, and increasing the weight of the hesitation behavior data through an amplification algorithm, in order to retrieve multiple potential interest content objects corresponding to the hesitation behavior data from the database. Next, in the candidate and scoring module, it connects to a multi-dimensional mining module to define reverse content objects, bridging content objects, and potential interest content objects as multiple surprise candidate contents. The surprise candidate contents are then dynamically weighted and scored against multiple similar candidate contents corresponding to the main interest vector to generate a content recommendation list, which is then transmitted to the user's device for display. The feedback and adaptive learning module connects to the candidate and scoring module to obtain user feedback data on the content recommendation list, which is used to correct the parameters of the main interest vector and the dynamically weighted scoring.

[0009] In addition, the present invention also discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, characterized in that, when the processor executes the computer program, it implements the multi-dimensional reverse mining and content recommendation method based on the user habit model as described above.

[0010] The methods, apparatus, and electronic devices disclosed in this invention are as described above. The difference between this invention and existing technologies lies in that this invention establishes a primary interest vector based on user historical behavior data via a server, and executes a multi-dimensional mining process to obtain multiple reverse content objects, multiple bridging content objects, and multiple potential interest content objects. These objects are then collectively defined as multiple surprise candidate contents. Next, the surprise candidate contents are dynamically weighted and scored against similar candidate contents corresponding to the primary interest vector to generate a recommendation list. Finally, the primary interest vector and the dynamically weighted score are adjusted based on user feedback data regarding their selections of the recommendation list.

[0011] Through the above-mentioned technical means, the present invention can achieve the technical effect of increasing the diversity of content recommendation lists. Attached Figure Description

[0012] Figure 1 This is a block diagram of the device for multi-dimensional reverse mining and content recommendation based on user habit models according to the present invention.

[0013] Figure 2A and Figure 2B This is a flowchart illustrating the multi-dimensional reverse mining and content recommendation method based on a user habit model, as described in this invention.

[0014] Figure 3 This is a schematic diagram illustrating the application of historical behavior data in this invention.

[0015] Figure 4 This is a schematic diagram illustrating the application of the knowledge graph of this invention.

[0016] Figure 5 This is a schematic diagram of the hardware architecture of the electronic device of the present invention.

[0017] Explanation of reference numerals in the attached figures: 110: User Profile Modeling Module 120: Multi-dimensional Mining Module 130: Candidate and Scoring Module 140: Feedback and Adaptive Learning Module 300: Structured behavioral data 400: Knowledge Graph 401a, 401b: First node 402: Second Node 500: Electronic Equipment 511: Processor 512: Memory 513: Network interface device 514: Internet 515: Display device 516: Graphical User Interface 517: Input device 518: Data storage device 519: Bus Step 210: The server obtains the user's historical behavior data from the user's terminal device to establish the corresponding user's primary interest vector. Step 220: The server executes a multi-dimensional data mining program. Step 221: Calculate the opposing regions in the feature space that are inversely related to the main interest vector, and retrieve multiple inverse content objects corresponding to the opposing regions from the database. Step 222: Extract multiple interest features with low correlation from the main interest vectors, and connect the common attributes between each interest feature through a pre-defined knowledge graph to retrieve multiple bridging content objects with cross-domain features from the database. Step 223: Detect user hesitation behavior data for specific content objects, and increase the weight of hesitation behavior data through an amplification algorithm to retrieve multiple potential content objects of interest corresponding to the hesitation behavior data from the database. Step 230: The server defines the reverse content object, bridging content object, and potential interest content object as multiple surprise candidate contents, and dynamically weights and scores the surprise candidate contents with multiple similar candidate contents corresponding to the main interest vector to generate a content recommendation list, which is then transmitted to the user's terminal device for display. Step 240: The server obtains user feedback data on their clicks on the content recommendation list to adjust the parameters of the main interest vector and the dynamic weighted score. Detailed Implementation The following will describe in detail the implementation of the present invention with reference to the accompanying drawings and embodiments, so as to fully understand how the present invention uses technical means to solve technical problems and achieve technical effects and to implement it accordingly.

[0018] Please refer to the following first. Figure 1 , Figure 1This is a block diagram of the device for multi-dimensional reverse mining and content recommendation based on a user habit model according to the present invention. The system includes: a user profile modeling module 110, a multi-dimensional mining module 120, a candidate and scoring module 130, and a feedback and adaptive learning module 140. The user profile modeling module 110 is used to acquire historical behavior data of users from user-end devices and establish the corresponding user's main interest vector. In some embodiments, the user profile modeling module 110 is used to establish the main interest vector, which specifically receives historical behavior data of users from user-end devices. The historical behavior data includes at least: content object identification code, event type, timestamp, and event weight, wherein the event type includes: click, favorite, share, or complete viewing. The user profile modeling module 110 can first preprocess the historical behavior data to remove duplicate events and map different event types to corresponding event weights, and then obtain the content feature vector of each content object. The content feature vector can be a sparse vector composed of content tags or categories, or an embedded vector formed by content text, tags, or image features. Subsequently, the user profile modeling module 110 can weight and combine the content feature vectors according to the event weights and timestamps of each event to form the main interest vector. A time decay function can be introduced to reduce the impact of earlier behaviors on the main interest vector. This time decay function can be exponential decay, piecewise linear decay, or moving average. Finally, the formed main interest vector is normalized and output as input data for the subsequent multi-dimensional mining module 120 for content retrieval and ranking. This ensures that the main interest vector is established based on specific data fields, calculable weight relationships, and a clear vector calculation process. Additionally, the user terminal device, which presents a content recommendation list and sends back click feedback data, can be a smartphone, tablet, or personal computer, etc. It can also execute content applications on the user terminal device to display specific content objects and detect user interaction behaviors such as dwell time, scrolling frequency, repeated clicks, and screen zooming on the display page to form hesitation behavior data. Furthermore, the user terminal device can also display web-based content services through a browser, and the front-end script can collect and send back interaction behaviors as the user's historical behavior data.

[0019] The multi-dimensional mining module 120 is connected to the user profile modeling module 110 to execute a multi-dimensional mining program. This program includes: calculating opposing regions in the feature space that are inversely related to the main interest vector, and retrieving multiple inverse content objects corresponding to these opposing regions from the database; extracting multiple interest features with low correlation from the main interest vector, and connecting the common attributes between each interest feature through a preset knowledge graph to retrieve multiple bridging content objects with cross-domain features from the database; and detecting user hesitation behavior data for specific content objects, and increasing the weight of the hesitation behavior data through an amplification algorithm to retrieve multiple potential interest content objects corresponding to the hesitation behavior data from the database. In some embodiments, the multi-dimensional mining module 120 can obtain multiple interest features with low weights and higher than a minimum threshold value from the main interest vector, and map these interest features to multiple first nodes in the knowledge graph. It then retrieves second nodes commonly associated with these first nodes as common attributes, and matches bridging content objects with cross-domain features from the database based on these common attributes. Furthermore, the multi-dimensional mining module 120 can receive hesitation behavior data transmitted back from user devices. This hesitation behavior data may include dwell time and scrolling frequency. Weight coefficients can be calculated based on this data, and then an amplification algorithm can be used to improve the ranking of content objects corresponding to the hesitation behavior data in the candidate set, thereby retrieving multiple potential interest content objects. The similarity can be cosine similarity, Euclidean distance, or inner product similarity. The preset threshold can be a fixed value or dynamically adjusted according to the similarity distribution, allowing the multi-dimensional mining module 120 to simultaneously generate reverse content objects, bridging content objects, and potential interest content objects through specific vector calculations, graph structure retrieval, and calculable weight adjustment processes. Additionally, the database may contain content objects and their corresponding feature vectors, as well as knowledge graph data, etc. The content objects may be articles, videos, audio, or product items; the feature vectors may be composed of content text, tags, categories, image features, or user interaction statistics, and a vector index structure can be used to support fast cosine similarity retrieval. In practice, the database can be a combination of a relational database and a vector database architecture, or a distributed file system can be used to store content objects and their corresponding feature vectors.

[0020] The candidate and scoring module 130, connected to the multi-dimensional mining module 120, defines reverse content objects, bridging content objects, and potential interest content objects as multiple surprise candidate contents. It then dynamically weights and scores these surprise candidate contents against multiple similar candidate contents corresponding to the main interest vector, generating a content recommendation list which is transmitted to the user's device for display. In some embodiments, the candidate and scoring module 130 merges the obtained reverse content objects, bridging content objects, and potential interest content objects into a candidate set, defining this set as multiple surprise candidate contents. Simultaneously, it retrieves the multiple content objects with the highest similarity from the database based on the main interest vector as multiple similar candidate contents. In practice, the candidate and scoring module 130 can calculate a comprehensive score for each content object. This comprehensive score may include a similarity score and a surprise score, and these are weighted and summed according to a mixed weight ratio to form the ranking result of the content objects. The similarity score can be calculated from the similarity between the feature vector of the content object and the main interest vector, while the surprise score can be configured based on whether the content object belongs to a reverse content object, bridging content object, or potential interest content object. In addition, the candidate and scoring module 130 can dynamically adjust the mixed weight ratio according to the type of user terminal device or network environment status, or adopt a two-stage sorting method, first sorting the surprise candidate content and similar candidate content separately, and then generating a content recommendation list in an interleaved merging manner, and transmitting the content recommendation list to the user terminal device for display. In other words, the content recommendation list is generated based on a calculable score composition, weight adjustment and sorting process.

[0021] The feedback and adaptive learning module 140 is connected to the candidate and scoring module 130 to obtain user click feedback data on the content recommendation list, which is used to correct the parameters of the main interest vector and the dynamic weighted scoring. In some embodiments, the feedback and adaptive learning module 140 is used to receive feedback data (i.e., click feedback data) of the content recommendation list returned by the user's terminal device. The click feedback data at least includes the user's click event on a content object and its corresponding content object identification code. The feedback and adaptive learning module 140 can update the weights corresponding to the content feature vectors of the clicked content objects in the main interest vector according to this feedback data, and reduce the weights corresponding to the unclicked content objects, so as to gradually adjust the weight distribution of each dimension of the main interest vector. In addition, the feedback and adaptive learning module 140 can also adjust the dynamic weighted scoring parameters in the candidate and scoring module 130 based on the feedback data. For example, when the content object corresponding to the feedback data is a surprise candidate, the weight of the surprise score in the mixed weight ratio can be increased so that the ranking ratio of surprise candidate content can be increased in the subsequent content recommendation list. The weight adjustment can be updated with a fixed step size or adaptively adjusted according to the number of feedbacks, the frequency of selection, or the historical stability, so that the main interest vector and the dynamic weighted scoring parameters are adjusted successively with the user behavior through a calculable weight update rule.

[0022] It is particularly important to note that, in practical implementation, this invention can be partially or entirely based on hardware. For example, it can be implemented using hardware processors such as integrated circuit chips, system-on-chips (SoCs), complex programmable logic devices (CPLDs), and field-programmable gate arrays (FPGAs). The hardware processor (hereinafter referred to as the processor) executes a computer program implementing this invention. This computer program is stored in a computer-readable storage medium, that is, it carries computer-readable program instructions for causing the hardware processor to implement various aspects of this invention. The computer-readable storage medium can be a tangible device capable of holding and storing instructions used by an instruction execution device. The computer-readable storage medium can be, but is not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: hard disks, random access memory, read-only memory, flash memory, optical disks, floppy disks, and any suitable combination thereof. The computer-readable storage medium used herein is not to be construed as a transient signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., optical signals through fiber optic cables), or electrical signals transmitted through wires. Furthermore, the computer-readable program instructions described herein can be downloaded from the computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to external computer devices or external storage devices. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, hubs, and / or gateways. Each computing / processing device's network interface card (NIC) or network interface receives the computer-readable program instructions from the network and forwards these instructions to the computer-readable storage medium within the respective computing / processing device. The computer program instructions that perform the operations of this invention can be assembly language instructions, instruction set architecture instructions, machine instructions, machine-dependent instructions, microinstructions, firmware instructions, or source code or object code written in any combination of one or more programming languages. The programming languages ​​include object-oriented programming languages ​​such as Common Lisp, Python, C++, Objective-C, Smalltalk, Delphi, Java, Swift, C#, Perl, Ruby, and PHP, as well as conventional procedural programming languages ​​such as C or similar programming languages.The computer program instructions may be executed entirely on the computer, partially on the computer, partially on the client computer and partially on a remote computer, or entirely on a remote computer or server.

[0023] Please see Figure 2A and Figure 2B , Figure 2A and Figure 2B This is a flowchart of the multi-dimensional reverse mining and content recommendation method based on a user habit model according to the present invention. The method is executed by a server connected to a user-end device and a database storing multiple content objects. The method comprises the following steps: the server obtains historical behavioral data of the user from the user-end device to establish the corresponding user's main interest vector (step 210); the server executes a multi-dimensional mining program (step 220), which includes: calculating the opposing regions in the feature space that are inversely related to the main interest vector, and retrieving multiple reverse content objects corresponding to the opposing regions from the database (step 221); extracting multiple interest features with low correlation from the main interest vector, and connecting the common attributes between each interest feature through a preset knowledge graph to retrieve multiple bridging content objects with cross-domain features from the database (step 222); and detecting user hesitation behavior data for specific content objects, and increasing the weight of the hesitation behavior data through an amplification algorithm to retrieve multiple potential interest content objects corresponding to the hesitation behavior data from the database (step 223). Next, the server defines the reverse content objects, bridging content objects, and potential interest content objects as multiple surprise candidate contents, and dynamically weights and scores these surprise candidate contents against multiple similar candidate contents corresponding to the main interest vector to generate a content recommendation list, which is then transmitted to the user's device for display (step 230). The server also obtains user feedback data on the content recommendation list to adjust the parameters of the main interest vector and the dynamic weighted score (step 240). Through these steps, the server can establish a main interest vector based on user historical behavior data and execute a multi-dimensional mining process to obtain multiple reverse content objects, multiple bridging content objects, and multiple potential interest content objects, defining these objects together as multiple surprise candidate contents. Then, the surprise candidate contents are dynamically weighted and scored against similar candidate contents corresponding to the main interest vector to generate a recommendation list, and the main interest vector and the dynamic weighted score are adjusted based on user feedback data on the recommendation list.

[0024] In some embodiments, the step 221 of calculating the opposing region in the feature space may include the server converting the primary interest vector into its inverse vector and calculating the cosine similarity between the feature vector and the inverse vector of each content object in the database, so as to select content objects with a cosine similarity greater than a preset threshold as the inverse content objects corresponding to the opposing region. Taking the calculation of the opposing region as an example, the value of each dimension in the primary interest vector can be taken as its opposite (e.g., multiplying vector V by -1), so that it points in completely opposite directions in the multidimensional space.

[0025] In some embodiments, the method for detecting hesitation behavior data in step 223 can be that the server detects the user's dwell time and scrolling frequency on the display page of a specific content object. When the dwell time exceeds a preset time limit and the scrolling frequency is lower than a preset frequency value, the dwell time and scrolling frequency are used as hesitation behavior data. For example, when a user browses a content object about "modern architectural design" through a user terminal device (such as a smartphone), the server can start an event timer and a displacement detector monitoring program. If the display page of the content object remains in the foreground and the cumulative dwell time recorded by the event timer exceeds 15 seconds (i.e., the preset time limit), and the displacement detector detects that the user's finger scrolls at a frequency lower than 0.2 times per second along the vertical axis of the screen (i.e., the preset frequency value, representing that the user is reading deeply rather than quickly scanning), then the server defines this "long dwell time, low scrolling" data feature as hesitation behavior data. Subsequently, the server extracts feature tags (such as minimalism, concrete material) of the content object "modern architectural design" and increases the weight of this feature tag by 1.5 times through an amplification algorithm, thereby marking this content object as a potential interest with a high conversion potential, rather than just an unintentional browsing.

[0026] In some embodiments, the multi-dimensional mining procedure further includes: the server calculating the user's interest migration path in different time periods based on timestamps in historical behavior data, and predicting the offset trend of the main interest vector at the expected time point based on the interest migration path, in order to retrieve multiple extended content objects corresponding to the offset trend from the database and include them in surprise candidate content. For example, the server divides user behavior into multiple consecutive observation periods (e.g., the first to third periods in months) based on timestamps in historical behavior data. If the analysis results show that the user's main interest tag migrates from "basic cycling" in the first period to "professional road bike equipment" in the second period, and then shows a trend of "marathon training" in the third period, the server will determine this evolution pattern as an interest migration path. Subsequently, the server uses a linear extrapolation algorithm to predict the offset trend in the fourth period (i.e., the expected time point), determining that the user is highly likely to have a high interest in the cross-domain "triathlon". In this way, the server can proactively retrieve multiple extended content objects related to "triathlon" from the database and include them in the surprise candidate content, so as to achieve a forward-looking and diversified recommendation before the user actually makes a selection.

[0027] In some embodiments, the dynamic weighted scoring step can be performed by the server dynamically adjusting the mixed weight ratio between surprise candidate content and similar candidate content based on the user's current network environment and the type of the user's terminal device, to generate a content recommendation list adapted to the user's terminal device. For example, when the server detects that the user's terminal device is a "smartphone" and its network environment status is displayed as "low bandwidth mobile network," in order to ensure content loading speed and reduce user churn rate, the server will dynamically adjust the mixed weight ratio, increasing the weight of "similar candidate content" to 80% and decreasing the weight of "surprise candidate content" to 20%, prioritizing the recommendation of content that the user is familiar with and has consistent tags. Conversely, when the server detects that the user's terminal device is a "tablet computer" with a large screen and is in a "high bandwidth network environment," it can be inferred that the user has a more ample browsing environment and higher patience for exploration. Therefore, the server can increase the weight ratio of "surprise candidate content" to 50% or higher to increase the diversity and exploratory nature of the content recommendation list, thereby generating a personalized list best suited to the current hardware conditions and usage context.

[0028] In some embodiments, the step of connecting common attributes between interest features through a preset knowledge graph can be performed by the server defining multiple interest features as multiple first nodes in the knowledge graph and retrieving second nodes in the knowledge graph that have common associations with these first nodes. The second nodes are common attributes, which are used to match bridging content objects in the database based on the attribute tags of the second nodes. The knowledge graph will be explained in detail with diagrams later.

[0029] In some embodiments, the step of increasing the weight of hesitant behavior data through an amplification algorithm can be performed by the server calculating the weighted sum of multiple user interactions with a specific content object. When the interaction includes at least one of increased dwell time, repeated clicking, and screen zooming, a weight coefficient is configured on the hesitant behavior data using the amplification algorithm. When the weight coefficient is greater than the weight of regular browsing behavior in historical behavior data, the ranking of the potential interest content object among surprise candidate content is increased. For example, assuming the user's interaction includes three actions such as "increased dwell time (e.g., more than twice the average dwell time)," "repeated clicking," and "screen zooming," the server can use a preset weight formula (e.g., ...) to increase the weight of the hesitant behavior data. ,in, The weighted total score represents the final weighted total value of a specific content object after evaluation of multi-dimensional interactive behavior. This value is used to determine the user's interest in the content and serves as the benchmark for configuring weight coefficients and ranking in the subsequent "amplification algorithm". A quantitative value representing the i-th specific interaction behavior of a user with respect to a content object, such as: dwell time, scrolling frequency, number of repeated clicks, number of screen zooms, etc. The weight coefficients for specific behaviors (representing the pre-defined technical importance weights for the i-th interaction behavior) are calculated as a weighted sum. Subsequently, the server can significantly increase the weight coefficients for specific behaviors configured in the algorithm (e.g., 2.5). When this weight coefficient is much greater than the weight of a single click in the history (e.g., 1.0), the server will determine that the user intends to explore in depth, and will then improve the ranking of the corresponding potential interest content objects in the surprise candidate content, ensuring that highly relevant objects are presented first in the recommendation list.

[0030] In some embodiments, the step of correcting the parameters of the primary interest vector and the dynamic weighted scoring may include the server determining whether the click feedback data belongs to surprise candidate content. If so, the server increases the ranking of the surprise candidate content in the content recommendation list. For example, when the obtained click feedback data shows that the user clicked on a "reverse content object" belonging to the surprise candidate content (e.g., a technology enthusiast clicked on a "classical literature" article), the server determines that the user currently has a high exploration intent. Subsequently, the server executes a parameter correction procedure, increasing the weight of the surprise candidate content in the dynamic weighted scoring formula from the original 0.2 to 0.35. This correction will directly result in a significant increase in the comprehensive score of all objects identified as surprise candidate content when generating the content recommendation list in the next recommendation cycle, thereby moving their ranking position in the list forward, thus achieving the technical effect of increasing recommendation diversity in real time and adaptively based on the user's actual click preferences. Taking the dynamic weighted scoring formula as an example, it can be... ,in, The overall ranking score represents the final order of content in the list. The higher the score, the higher the position in the recommendation list (e.g., displayed at the top of the page). This is the weighting coefficient for similar content, representing the importance weight of "similar candidate content"; The similarity score represents the degree of matching between the content object and the user's "primary interest vector," and can be obtained by calculating the cosine similarity between feature vectors. The weighting of surprise candidate content reflects the intensity of intervention in "exploring potential user preferences" or "breaking the information cocoon"; The surprise score represents the rating of an object as a surprise candidate (including: reverse, bridging, or potential interest objects). This value is generated based on the aforementioned multi-dimensional mining process (such as the calculation results of the amplification algorithm).

[0031] It should be further explained that the historical behavior data may include: content object identification code, event type, timestamp, and event weight. The event type may include selection, favorite, sharing, or viewing completion. The server can establish a primary interest vector based on the historical behavior data, for example, by forming a sparse vector using the weighted frequency of each content tag, or by weighted averaging the feature vectors of multiple content objects recently interacted with by the user to form a dense vector. Alternatively, the primary interest vector can also be established based on category fields, keyword weights, or content embedding vectors, and the influence of earlier behaviors can be reduced using a time decay function. Next, the feature space refers to the vector space composed of the feature vectors of content objects, with each content object corresponding to a feature vector located in the feature space. Furthermore, the server can first take the negative value of the primary interest vector to form an inverse vector, and calculate the cosine similarity between the feature vectors of each content object in the database and the inverse vector. Then, the set of content objects corresponding to those with a cosine similarity greater than a preset threshold is defined as the "opposite region," and these content objects are considered "reverse content objects." In practice, opposing regions can also be defined using Euclidean distance, inner product similarity, or spherical distance, and the preset threshold can be obtained through offline verification or online adaptive adjustment. In actual implementation, the calculation of opposing regions can be achieved by the server first converting the primary interest vector into an inverse vector and establishing a feature vector for each content object in the database. For example, the feature vector can be composed of hot-encoded vectors of content tags, TF-IDF vectors, or content embedding vectors. Next, the server calculates the cosine similarity between the feature vector and the inverse vector, and selects content objects with a cosine similarity greater than the preset threshold as inverse content objects. In some embodiments, the preset threshold can be set between 0.6 and 0.9, and can be set separately for different content categories. In other embodiments, the preset threshold can be set as a dynamic threshold, for example, taking the value corresponding to the top N percentiles (N is a positive integer) of the cosine similarity distribution, or adjusting it according to the intensity of recent user interactions.

[0032] Next, regarding the "interest features" part, in practice, the server can select multiple dimensions with lower weights but still above a minimum threshold from the main interest vectors as multiple interest features to represent the user's secondary or potential interests. As for the "knowledge graph" part, it can be a graph data structure containing multiple nodes and edges, where nodes correspond to entities or attribute labels, and edges represent the relationships between nodes. The server can map multiple interest features to multiple first nodes in the knowledge graph and retrieve second nodes commonly associated with these first nodes as shared attributes, used to match bridging content objects with cross-domain characteristics in the database. In practice, the knowledge graph can be implemented using an ontology hierarchy or a tag co-occurrence network, and shared attributes can be common categories, common authors, or common topic events, etc. In practice, the server can perform the following steps to complete the establishment of the knowledge graph: (1) The server obtains the meta data of all content objects (including: content tags, category tags, keywords and the domain to which the content belongs) from the database, and defines each unique content tag or domain concept as a node in the knowledge graph. Each node has a unique node identification code and a corresponding attribute tag; (2) The server establishes an edge between two nodes with semantic association based on the tag co-occurrence statistics or the conceptual relationship of the domain ontology. Each edge has an edge weight value. The edge weight value is calculated based on the co-occurrence frequency or the ontology level distance. The higher the co-occurrence frequency or the smaller the ontology level distance, the higher the corresponding edge weight value; (3) The server performs normalization processing on the established knowledge graph to remove edges with edge weight values ​​lower than the preset minimum edge weight threshold, so as to retain the node connection relationship with high semantic association; and (4) The server stores the established knowledge graph in the database in a graph index structure (such as an adjacency matrix or an adjacency string). In addition, the server can perform incremental updates to the knowledge graph according to a preset update cycle (e.g., weekly or monthly). These incremental updates are based on newly added content object post-data and the latest user interaction statistics, adding nodes, adding edges, or adjusting the edge weights of existing edges to ensure the knowledge graph reflects the latest content domain relationships and user interest distribution. In some embodiments, the knowledge graph may not rely on offline pre-construction. Instead, the server can instantly build a lightweight local knowledge subgraph based on the tag co-occurrence statistics in the current database before each multi-dimensional mining program execution. This local knowledge subgraph serves as the retrieval basis for bridging content objects, adapting to application scenarios where database content objects are rapidly updated.

[0033] Regarding the "hesitation behavior data," user devices can collect dwell time and scrolling frequency on the display page of a specific content object. When the dwell time exceeds a preset time limit and the scrolling frequency is lower than a preset frequency value, the dwell time and scrolling frequency are used as hesitation behavior data and sent back to the server. Furthermore, to convert the hesitation behavior data into a weighted calculation for ranking improvement, the server executes an "amplification algorithm" to calculate weight coefficients based on the hesitation behavior data. These weight coefficients are then used to improve the ranking of the content object corresponding to the hesitation behavior data in the candidate set, thereby retrieving multiple potential interest content objects from the database. For example, the weight coefficients calculated by the amplification algorithm can be obtained by weighted combination of the normalized results of dwell time and scrolling frequency. For instance, when a user's dwell time on a specific content object is 30 seconds, the user's historical average dwell time is 10 seconds, the scrolling frequency is 0.1 times / second, and the preset maximum scrolling frequency reference value is 2.0 times / second, the server will calculate the dwell time ratio T / T. avg =3.0 (where T is the dwell time in seconds; T avg (Historical average dwell time, in seconds) and the normalized value of the scrolling frequency (1) F / F max =0.95 (where F is the winding frequency, in seconds; F) max Multiplying the preset maximum scrolling frequency (in seconds) yields the hesitation intensity score H = 3.0 × 0.95 = 2.85 (H represents the degree of hesitation of the user's action). Substituting this into the amplification factor formula A = 1 + α... H (where α=1.0, a larger value indicates greater sensitivity to hesitation behavior, preset to 1.0 and adjustable according to desired sensitivity) yields an amplification weighting coefficient A=3.85. Finally, A is multiplied by the original surprise score to generate the enhanced weight, which serves as the ranking criterion for potential interest content objects among surprise candidate content. In addition, hesitation behavior data can also include repeated clicks, screen zooming, or cursor hovering trajectories. In practice, hesitation behavior data can be detected by starting a timer when the display page of a specific content object loads and stopping the timer when leaving the display page or switching to the next content, thus obtaining the dwell time. Furthermore, the number of scroll events triggered on the display page can be monitored and divided by the observation time window to obtain the scroll frequency. When the dwell time exceeds a preset time limit and the scroll frequency is lower than a preset frequency value, the user device sends back the dwell time and scroll frequency as hesitation behavior data. For example, the preset time limit can be set to 8 to 30 seconds, and the preset frequency value can be set to less than 0.2 scroll events per second. Alternatively, mouse movement distance or touch trajectory change rate can be used as alternative indicators of scrolling frequency, or "dwell time exceeds the preset time limit and no click event occurs" can be used as the criterion for judging hesitation behavior data.

[0034] Regarding the "surprise candidate content" portion, the server can merge reverse content objects, bridging content objects, and potential interest content objects into a candidate set, and remove duplicate content object identifiers to form multiple surprise candidate content. Additionally, for the "similar candidate content" portion, the server can retrieve the multiple content objects with the highest similarity from the database based on the main interest vector as multiple similar candidate content. Regarding the "dynamic weighted scoring" portion, in practice, the server can calculate a comprehensive score for each content object, which includes a similarity score and a surprise score, and then sum them according to a weighted average to generate a ranking. The weighted average can be adjusted based on the user's device type or network environment. In practice, the dynamic weighted scoring can be modified to a two-stage ranking: first, surprise candidate content and similar candidate content are ranked separately, and then a content recommendation list is generated using an interleaved merging method.

[0035] Regarding the "click feedback data," the server can receive click events and their corresponding content object identification codes from user devices, and update the weights of each dimension of the main interest vector accordingly. For example, it can increase the weight of the dimension corresponding to the content tag of the clicked content object and decrease the weight of the dimension corresponding to the unclicked content object. Additionally, the server can adjust the parameters of the dynamically weighted scoring. For instance, when the content object corresponding to the click feedback data belongs to the surprise candidate content, the weight of the surprise score in the mixed weight ratio is increased, so that the proportion of surprise candidate content increases in the subsequent content recommendation list. Furthermore, click feedback data can be replaced with collection feedback or viewing completion feedback, and parameter adjustments can use a fixed step size update or adaptively adjust the step size based on confidence level.

[0036] In some embodiments, the multi-dimensional mining program can calculate interest migration paths and predict shift trends. For example, the server can group user behavior into multiple time periods (e.g., morning, noon, evening, or weekdays, holidays) based on the timestamps of historical behavior data, and establish a primary interest vector for each time period to represent changes in user interests over time. The server can consider the differences in primary interest vectors between adjacent time periods as interest migration paths, and predict the shift trend at the expected time point using linear extrapolation or moving average, for example, determining that the weight of certain content tags is trending upwards. Then, the server can retrieve corresponding extended content objects from the database based on the shift trend and include them in surprise candidate content. In addition, interest migration paths can also be estimated using Markov chain transition probabilities or time series models, and the expected time point can be set as the next recommendation cycle or the next login cycle.

[0037] In some embodiments, the dynamic weighted scoring can adjust the weighting ratio based on network conditions and the type of user device. In practice, network conditions may include real-time latency, available bandwidth, or packet loss rate. Additionally, the type of user device may include smartphones, tablets, or smart displays. For example, when network conditions indicate low available bandwidth, the server can increase the weight of similar candidate content and decrease the weight of surprise candidate content to prioritize content with lower loading costs or that the user is more likely to make a quick decision about. Conversely, when network conditions are good and the user device is a large-screen device, the weight of surprise candidate content can be increased to enhance exploration. In other embodiments, the weighting ratio can switch based on the user's usage scenario (e.g., commuting mode or home mode) or adaptively adjust based on the user's historical selection feedback of surprise candidate content.

[0038] The following combinations Figures 3 to 5 The following description is provided by way of examples; please refer to the previous section first. Figure 3 , Figure 3 This is a schematic diagram illustrating the application of historical behavior data in this invention. Assume a user uses an electronic device (e.g., a device, server, etc.) that applies this invention. This electronic device can recommend articles or videos. When a user browses or operates the electronic device daily, the device continuously records the user's interactive behavior, such as: selecting an article, the duration of viewing a video, whether scrolling through content pages, and whether viewing is completed. In actual operation, the electronic device first organizes the above-mentioned historical user behavior data into structured behavior data 300, such as... Figure 3 As shown, JSON format can be used, with field names: content_id, content_tags, event_type, event_weight, timestamp, dwell_time_sec, and scroll_frequency_per_sec, each storing the corresponding data. These field names represent, in order: "Content object identifier, corresponding to a specific content object in the database," "Content tag, used to build the feature vector," "Event type, such as click," "Event weight," "Event occurrence time, used for time decay function calculations; earlier actions have lower influence," "Dwell time," and "Scroll frequency." Next, based on the content types and tags the user frequently interacts with, a primary interest vector is created to represent the user's current main preferences. For example, if a user frequently reads technology articles, the primary interest vector will have a higher weight on the "technology" related feature dimensions.

[0039] like Figure 4 The meaning is shown. Figure 4This is a schematic diagram illustrating the application of the knowledge graph of the present invention. In addition to establishing primary interest vectors, the electronic device further expands the sources of recommended content through the knowledge graph 400. For example, when multiple interest features extracted from the primary interest vectors respectively include "camping supplies" and "outdoor photography," the server can define these two as two first nodes (401a, 401b) in the knowledge graph 400. After path retrieval, it can be found that these two first nodes (401a, 401b) are connected to the same second node 402 named "High Mountain Adventure" (i.e., shared attributes) through relationships such as "mountain attributes" and "forest activities." Subsequently, the electronic device can automatically match bridging content objects from the database that do not originally belong to the user's primary interest area (such as "professional mountaineering equipment" or "mountain climbing of the Hundred Mountains"). Through this semantic association retrieval of the graph structure, the user's original content preference framework can be effectively broken, thereby retrieving bridging content objects that are technically relevant but belong to different fields.

[0040] In addition, electronic devices also search for content in the opposite direction to the primary interest vector, identifying reverse content objects that differ significantly from the user's existing preferences. Furthermore, when the electronic device detects that a user lingers on certain content pages for an extended period without immediately clicking or leaving, it interprets this behavior as hesitation and identifies content that might align with the user's potential interests (i.e., potential interest content objects). The electronic device then integrates these reverse, bridging, and potential interest content objects into a set of surprise candidate content, while simultaneously selecting content highly similar to the primary interest vector as similar candidate content. In practice, the electronic device calculates a comprehensive score for each piece of content, combining the similarity between the content and the user's interests with whether the content qualifies as surprise candidate content. This generates a content recommendation list that balances familiarity and novelty, which is then displayed.

[0041] In this way, when a user actually selects some content from the recommended list, the electronic device obtains the user's selection feedback data and adjusts the weighting parameters used in the main interest vector and content scoring based on this data. For example, if the user increases the proportion of selections for surprise candidates, the electronic device will increase the proportion of surprise candidates in subsequent recommendations, allowing the recommendation results to be dynamically adjusted according to user behavior. Thus, through this process, the electronic device can not only continuously align with the user's existing interests but also appropriately introduce cross-domain or potential interest content, thereby effectively improving the diversity and exploratory nature of the content recommendation list in practical applications.

[0042] In the comparative embodiment, it is assumed that the content recommendation mechanism calculates the similarity of content objects based solely on the content that the user has previously selected or viewed, and continuously recommends content most similar to the user's existing preferences. For example, when a user frequently reads technology articles, a traditional content recommendation mechanism will repeatedly recommend technology news or articles on the same topic. While this approach can increase content selection rates in the short term, over time, the recommended content tends to become homogenized, limiting the range of content the user receives in the long term and making it difficult to guide the user to explore new topics or content of potential interest.

[0043] In contrast, in embodiments of the present invention, in addition to establishing a primary interest vector based on the user's historical behavior, the electronic device also obtains reverse content, bridging content, and potential interest content through a multi-dimensional mining mechanism. For example, when a user's primary interests are concentrated in technology-related content, the electronic device can recommend not only content highly similar to technology but also reverse content that is opposite to or significantly different from the technology theme, or connect technology with cross-domain themes such as design and business applications through a knowledge graph to form bridging content. Furthermore, if the electronic device detects that a user spends a considerable amount of time on certain non-technology content pages without immediately clicking on them, this behavior can also be regarded as a potential interest clue and thus included in the recommendation consideration.

[0044] In the content sorting stage, traditional methods typically sort based on a single similarity score. However, embodiments of the present invention use a candidate and scoring mechanism to simultaneously calculate similarity and surprise scores for content objects and generate a content recommendation list based on a mixed weight ratio. This allows the recommendation results to maintain user familiarity while still introducing content with novelty or exploratory value.

[0045] Furthermore, when a user actually interacts with the content recommendation list, traditional methods often simply record this interaction as input data for the next similarity calculation. However, embodiments of this invention dynamically adjust the weight parameters used for the main interest vector and content scoring through feedback and adaptive learning mechanisms. For example, if a user's selection rate for surprise candidate content gradually increases, the electronic device can appropriately increase the ranking weight of surprise candidate content in subsequent recommendations, allowing the recommendation strategy to adjust according to changes in user behavior.

[0046] The above comparison shows that, compared with traditional recommendation methods based solely on similarity, the embodiments of the present invention can better balance content relevance and diversity in practical applications. Through multi-dimensional mining and adaptive learning mechanisms, it effectively reduces the problem of homogenization of recommendation results, thereby enhancing the exploratory nature and long-term usability of the content recommendation list.

[0047] like Figure 5 The meaning is shown. Figure 5This is a schematic diagram of the hardware architecture of the electronic device of the present invention. Within the electronic device 500, there are multiple computer-executable instructions (hereinafter referred to as instructions) for driving the machine to perform any or more methods discussed in the present invention. In other embodiments, the machine may be connected (e.g., network connected) to other machines in a local area network, internal network, external network, or internet. The machine may operate as a server or client in a client-server network environment, or as a peer machine in a peer-to-peer network environment. The machine may also function as a network device, server, router, switch or bridge, event generator, distributed node, centralized system, or any machine capable of executing a set of instructions (whether sequential or otherwise) specifying the actions the machine needs to take. Furthermore, although only one machine is illustrated, the term "machine" should also be considered as including a collection of any machines (e.g., tablet computers) that can individually or jointly execute a set (or more) of instructions to perform any or more methods discussed in the present invention.

[0048] The electronic device 500, such as a tablet computer, a smart display, etc., includes a processor 511, a memory 512 (such as a read-only memory, flash memory, dynamic random access memory, non-volatile resistive random access memory, embedded flash memory, or ferroelectric random access memory (FeRAM), a network interface device 513, a display device 515, an input device 517, and a data storage device 518 (which may include a fixed or removable computer-readable storage medium), and these components communicate via a bus 519.

[0049] The processor 511 may be a microprocessor, a central processing unit (CPU), or a similar device. More specifically, the processor 511 may be a Complex Instruction Set Computer (CISC) microprocessor, a Reduced Instruction Set Computer (RISC) microprocessor, a Very Long Instruction Word (VLIW) microprocessor, or a general-purpose instruction set processor implementing other instruction sets, or a general-purpose instruction set processor implementing multiple instruction sets. Furthermore, the processor 511 executes various software elements stored in memory 512 to perform various functions for the electronic device 500. In one embodiment, these software elements include an operating system, compiler elements, and a communication module (or instruction set). The operating system includes various programs, instruction sets, software elements, and / or drivers for controlling and managing general system tasks and facilitating communication between various hardware and software elements. A compiler is a computer program (or assembly of programs) that translates raw code written in a programming language into another computer language (e.g., a target language, object code). The communication module can communicate with other devices through the network interface device 513. The network interface device 513 is connected to the network 514 (such as a local area network, a wide area network, or a similar network) to communicate with other devices.

[0050] The memory 512 may store program code and / or data for use by the processor 511. The memory 512 may be implemented using random access memory (e.g., SRAM, DRAM, DDRAM), ROM, magnetic and / or optical storage devices, flash memory, or any combination thereof. The memory 512 may also include a transmission medium for carrying information-bearing signals (which may be modulated signals with or without carrier waveforms) representing instructions or data.

[0051] The display device 515, such as a liquid crystal display (LCD), LED, or cathode ray tube (CRT), is connected to the computer system via a display port and a graphics chipset, and provides a graphical user interface (GUI) 516 for user operation. In practice, the display device 515 can also be a touch screen or similar device that combines input and display functions. Additionally, the input device 517 (e.g., keyboard, mouse, touchpad, etc.) can provide user input commands and generate trigger signals.

[0052] Data storage device 518 may include a machine-readable storage medium (or more specifically, a computer-readable storage medium) storing one or more instruction sets embodying any or more methods or functions described in this invention. The disclosed data storage mechanism may be fully or at least partially implemented in memory 512. The data storage device 518 and memory 512 disclosed in electronic device 500 may be configured to implement a data storage mechanism for performing the operations and steps discussed in this invention.

[0053] In summary, the difference between this invention and existing technologies lies in the fact that the server establishes a primary interest vector based on user historical behavior data and executes a multi-dimensional mining process to obtain multiple reverse content objects, multiple bridging content objects, and multiple potential interest content objects, collectively defining these objects as multiple surprise candidate contents. Next, the surprise candidate contents are dynamically weighted and scored against similar candidate contents corresponding to the primary interest vector to generate a recommendation list. Furthermore, the primary interest vector and dynamic weighted score are adjusted based on user feedback data regarding the recommendation list. This technical approach solves the problems existing in existing technologies, thereby achieving the technical effect of improving the diversity of the content recommendation list.

[0054] Although the present invention has been disclosed above with reference to the foregoing embodiments, it is not intended to limit the present invention. Any person skilled in the art may make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of patent protection of the present invention shall be determined by the scope defined in the appended claims.

Claims

1. A multi-dimensional reverse mining and content recommendation method based on a user habit model, executed by a server, wherein the server connects to a user-end device and a database, and the database stores multiple content objects, characterized in that... The method includes: The server obtains the user's historical behavior data from the user terminal device to establish the user's main interest vector. The server executes a multi-dimensional data mining program, which includes: Calculate the opposing regions in the feature space that are inversely related to the main interest vector, and retrieve multiple inverse content objects corresponding to the opposing regions from the database; Extracting multiple interest features with low correlation from the main interest vector, and connecting the common attributes between each interest feature through a preset knowledge graph, to retrieve multiple bridging content objects with cross-domain features from the database; and The system detects user hesitation behavior data for specific content objects and increases the weight of the hesitation behavior data through an amplification algorithm in order to retrieve multiple potential interest content objects corresponding to the hesitation behavior data from the database. The server defines the reverse content object, the bridging content object, and the potential interest content object as multiple surprise candidate contents, and dynamically weights and scores these surprise candidate contents with multiple similar candidate contents corresponding to the main interest vector to generate a content recommendation list, which is then transmitted to the user's terminal device for display; and The server obtains the user's click feedback data on the content recommendation list, and uses it to correct the parameters of the main interest vector and the dynamic weighted score.

2. The multi-dimensional reverse mining and content recommendation method based on a user habit model as described in claim 1, characterized in that, The step of calculating the opposing region in the feature space includes: the server converting the main interest vector into its inverse vector, and calculating the cosine similarity between the feature vector of each content object in the database and the inverse vector, so as to select the content object with the cosine similarity greater than a preset threshold as the inverse content object corresponding to the opposing region.

3. The multi-dimensional reverse mining and content recommendation method based on a user habit model as described in claim 1, characterized in that, The steps for detecting the hesitation behavior data include: the server detecting the user's dwell time and scrolling frequency on the display page of the specific content object; when the dwell time exceeds a preset time limit and the scrolling frequency is lower than a preset frequency value, the dwell time and the scrolling frequency are used as the hesitation behavior data.

4. The multi-dimensional reverse mining and content recommendation method based on a user habit model as described in claim 1, characterized in that, The multi-dimensional mining program further includes: the server calculating the user's interest migration path in different time periods based on the timestamps in the historical behavior data, and predicting the offset trend of the main interest vector at the expected time point based on the interest migration path, so as to retrieve multiple extended content objects corresponding to the offset trend from the database and include them in the surprise candidate content.

5. The multi-dimensional reverse mining and content recommendation method based on a user habit model as described in claim 1, characterized in that, The dynamic weighted scoring step includes: the server dynamically adjusting the mixed weight ratio between the surprise candidate content and the similar candidate content based on the user's current network environment status and the type of the user's terminal device, so as to generate the content recommendation list adapted to the user's terminal device.

6. The multi-dimensional reverse mining and content recommendation method based on a user habit model as described in claim 1, characterized in that, The step of connecting the common attributes between the interest features through the preset knowledge graph includes: the server defines the multiple interest features as multiple first nodes in the knowledge graph, and retrieves a second node in the knowledge graph that has a common association with the multiple first nodes, wherein the second node is the common attribute, and is used to match the multiple bridging content objects in the database according to the attribute label of the second node.

7. The multi-dimensional reverse mining and content recommendation method based on a user habit model as described in claim 1, characterized in that, The step of increasing the weight of the hesitation behavior data through the amplification algorithm includes: the server calculating the weighted sum of multiple interactive behaviors of the user on the specific content object; when the interactive behavior includes at least one of increased dwell time, repeated clicks, and screen zooming, the server configuring a weight coefficient for the hesitation behavior data through the amplification algorithm; when the weight coefficient is greater than the weight of regular browsing behavior in the historical behavior data, the server increasing the ranking of the potential interest content object in the surprise candidate content.

8. The multi-dimensional reverse mining and content recommendation method based on a user habit model as described in claim 1, characterized in that, The step of correcting the parameters of the main interest vector and the dynamic weighted score includes: the server determining whether the click feedback data belongs to the surprise candidate content; if so, the server increasing the ranking of the surprise candidate content in the content recommendation list.

9. A multi-dimensional reverse mining and content recommendation device based on a user habit model, characterized in that, include: The user profile modeling module is used to obtain the user's historical behavior data from the user's terminal device and establish the main interest vector corresponding to the user. A multi-dimensional mining module, connected to the user profile modeling module, is used to execute a multi-dimensional mining program, which includes: Calculate the opposite region in the feature space that is inversely related to the main interest vector, and retrieve multiple reverse content objects corresponding to the opposite region from the database; Extracting multiple interest features with low correlation from the main interest vector, and connecting the common attributes between each interest feature through a preset knowledge graph, to retrieve multiple bridging content objects with cross-domain features from the database; and The system detects user hesitation behavior data for specific content objects and increases the weight of the hesitation behavior data through an amplification algorithm in order to retrieve multiple potential interest content objects corresponding to the hesitation behavior data from the database. The candidate and scoring module, connected to the multi-dimensional mining module, is used to define the reverse content object, the bridging content object, and the potential interest content object as multiple surprise candidate contents, and to dynamically weight and score the surprise candidate contents with multiple similar candidate contents corresponding to the main interest vector, so as to generate a content recommendation list and transmit it to the user terminal device for display. as well as The feedback and adaptive learning module, connected to the candidate and scoring modules, is used to obtain user feedback data on the content recommendation list, in order to correct the parameters of the main interest vector and the dynamic weighted score.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the multi-dimensional reverse mining and content recommendation method based on the user habit model as described in claim 1.

Images