Personalized recommendation method driven by user intention recognition

Abstract

The invention discloses a personalized recommendation method driven by user intention recognition, which comprises the following steps of: collecting multi-dimensional information such as browsing records, click behaviors and comment data of a user, and constructing a user behavior data set; preprocessing the collected user behavior data to obtain a session sequence; the current demand of the user is speculated by analyzing the input text and the behavior mode of the user, and the intention of the user is recognized; constructing a user interest preference prediction model according to the identified user intention; and in combination with real-time feedback of the user, the recommendation model is dynamically updated, and the response speed and the individuation degree of the recommendation system are improved. The problem that a traditional recommendation system only depends on static data and does not comprehensively consider real-time feedback of users is solved, self-adaptive ability is injected for personalized recommendation, intention behaviors of different users in multiple scenes can be analyzed and understood, accurate recommendation service is provided, and recommendation accuracy and user satisfaction are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and information recommendation systems, specifically to a personalized recommendation method driven by user intent recognition. Background Technology

[0002] In today's era of rapid development in big data and artificial intelligence, recommender systems have become an essential component of various internet platforms, playing a crucial role in e-commerce, social networks, and content distribution. Traditional recommender systems primarily rely on collaborative filtering, content recommendation, and hybrid recommendation methods. Existing methods typically construct user profiles based on users' historical behavior, browsing history, and explicit feedback information, and then mine potential interests and preferences from these. However, these methods often have significant limitations when faced with the challenges of diverse user behavior data and the difficulty in accurately capturing implicit needs. First, traditional methods suffer from an imbalance in capturing users' short-term and long-term interests, easily influenced by popular items and ignoring individual differences. Second, recommendation methods based on behavioral statistics often struggle to deeply understand users' true needs; their recommendation results often only reflect surface-level behavioral patterns, neglecting users' multi-dimensional intentions and emotional expressions in different contexts.

[0003] In recent years, with the rapid development of natural language processing and deep learning technologies, more and more research has focused on how to infer users' true intentions using user-input text data, comments, and other unstructured information. During interactions, users not only convey their interests through browsing, clicking, and commenting, but also implicitly reveal their specific needs and expectations for products or services. Accurately extracting user intent from multi-dimensional information has become crucial for improving the personalization and accuracy of recommendation systems. Recommendation systems based on user intent recognition have emerged in this context. Their core idea is to make implicit user needs explicit through deep analysis of user language expressions and behavioral patterns, using this as a key basis for recommendation strategies. Furthermore, with the widespread adoption of social networks and mobile internet, user behavior has become more diverse and dynamic, with significant differences in interaction behaviors among different users in different scenarios. Traditional recommendation systems often treat user behavior as static data, lacking a response mechanism for dynamic changes in user behavior and context switching. This requires recommendation systems not only to accurately model users' historical behavior but also to combine real-time feedback with dynamic update mechanisms to continuously adjust user profiles and recommendation strategies to adapt to the instantaneous changes in user needs. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a personalized recommendation method driven by user intent recognition, which solves the limitation of traditional recommendation methods that rely on static data, injects intelligent adaptive capabilities into personalized recommendations, facilitates the analysis and understanding of the intent of different users in multiple scenarios, provides accurate recommendation services, and improves the accuracy of the recommendation system and user satisfaction.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] This personalized recommendation method, driven by user intent recognition, includes the following steps:

[0007] Step S1: Collect multi-dimensional information data of users and construct user datasets;

[0008] Step S2: Preprocess the collected user data to obtain the session sequence;

[0009] Step S3: By analyzing the user's input text and behavioral patterns, infer the user's current needs and identify the user's intent;

[0010] Step S4: Based on the identified user intent, construct a user interest and preference prediction model;

[0011] Step S5: Combine real-time user feedback to dynamically update the recommendation model, thereby improving the response speed and personalization of the recommendation system.

[0012] Furthermore, step S1 specifically includes: the system collects user information data through multiple channels, covering multimodal information from mobile devices, PCs, and other smart terminals; the collected data includes not only explicit behaviors, such as search queries, clicks, favorites, ratings, and comments, but also implicit behaviors, such as browsing duration, dwell location, and operation trajectory; in addition, it collects contextual information related to the user's environment, such as geographical location, device type, timestamp, and network environment.

[0013] Step S2 specifically includes:

[0014] Step S201: Preprocess the collected user data, including preprocessing operations for text data and image data:

[0015] (1) For text data, stop words and special characters are removed, and word forms are standardized. The specific process includes two steps: completing to a fixed length and adding a start marker.

[0016] The complete-to-fixed-length method completes the text by concatenating "..." <pad>The character "" makes the overall length reach the predetermined length;

[0017] The method of adding a start tag adds a tag "" before the first character of the text. <cls>The ',' is used as the start marker for the text, and a tag is added after the last character of the text. <sep>, as a marker to indicate the end of the text;

[0018] After the processed text is input into the model, the specific vector output will be adjusted according to the downstream task and context information, while the [CLS] flag added in the preprocessing will be converted into sentence vectors after the model is trained;

[0019] (2) For image data, denoising, cropping, and size normalization are performed. The normalization formula is as follows:

[0020]

[0021] Where I(x,y) represents the pixel value of the original image, μ I and σ I These are the pixel mean and standard deviation, respectively, to ensure the consistency of the image in subsequent processing;

[0022] Step S202: Perform feature extraction operations on the text data to obtain the word sequence T = {t1, t2, ..., t}. n }, where t n This represents the nth word; for image data I, which consists of image pixel values ​​I(x,y), the system uses a pre-trained convolutional neural network to extract high-level features of the image and obtain a feature vector. To capture the semantic information of the image, where This represents the image after normalization.

[0023] Step S203: Perform word vectorization processing on the text data, and use a pre-trained model to vectorize each word t. i Mapping to a low-dimensional vector space yields the vectorized word v(t). i The specific method is as follows:

[0024] v(t i )=∑ j∈C(i) α ij ·P(t j )+β·S(t i )

[0025] Where C(i) is the word t i The context window, P(t) j ) represents the pre-trained word embedding model for word t j The mapping function, α ij The attention weights are represented by Softmax(f(P(t)). i ),P(t j The function is calculated as f(·), where f(·) is the similarity measure function, and S(t) is the similarity measure function. i ) represents the embedding vector of part-of-speech tagging; β represents the weight coefficient of the part-of-speech feature;

[0026] The text is represented as a sequence of word vectors V(T) = {v(t1), v(t2), ..., v(t3)}. n The feature vector Z of the image has already been obtained in the previous step;

[0027] Construction and fusion of multimodal features:

[0028] First, for text data, the term frequency TF(t) is calculated. i ) and inverse document frequency (IDF)(t) i The TF-IDF features are obtained using the following method:

[0029]

[0030] F(t i ) = TF(t i )·IDF(t i )

[0031] in, For the word t i The number of occurrences, where N is the total number of documents in the document collection, df(t) i ) is the word containing t i The number of documents;

[0032] Then, a statistically weighted text feature matrix is ​​generated:

[0033] M=[F(t1)*v(t1),F(t2)*v(t2),...,F(t n )*v(t n )]

[0034] Finally, a feature fusion strategy is used to merge the text features V(T) and the image features Z to form a joint feature representation:

[0035]

[0036] Here, mean() performs mean pooling on the weighted text features, and U represents multimodal information from n different scenarios, where the information for each scenario is derived from v(t). i ) and the corresponding image features z i ∈Z, where R represents the set of real numbers.

[0037] Furthermore, step S3 includes:

[0038] Step S301: Based on the word embeddings of the service task request input sequence, a bidirectional LSTM network and self-attention mechanism are used to extract shared features for intent recognition and semantic slot filling; in the input layer of the model, the classification values ​​are mapped to the corresponding integer values, and each integer value is represented by a binary vector; then, positional encoding is used to record the relative or absolute position of each word in the whole sentence to learn the positional relationship between words; finally, the word vectors are added to the positional encodings to generate word embedding vectors as the input of the model;

[0039] To further capture word order information, sine and cosine positional encoding is introduced. The positional encoding is added element-wise to the word vector to generate the final embedding vector containing both semantic and positional information. The specific method is as follows:

[0040]

[0041] Where p is the position of the word in the sequence, i is the dimension index, d is the dimension of the embedding vector, and C represents a large constant;

[0042] Then, the fused embedding vector is input into a bidirectional LSTM network, and positive hidden states are generated through the dual transmission of forward and backward information. and reverse hidden state The two are then concatenated to obtain the overall context representation. The specific method is as follows:

[0043] In the bidirectional LSTM layer, for each input sentence T i =(t1,t2,…,t n The word representation E = (e1, e2, ..., e) is obtained through word embedding. n )∈R n*d As input, from e1 to e n and from e n E1 is then used for forward and reverse encoding; forward encoding generates the hidden state. Reverse encoding generates hidden states The final generated encoded representation h i Depend on and It is pieced together, and the formula is as follows:

[0044]

[0045] in, For forward LSTM units, computation is performed sequentially from 1 to n. For backward LSTM units, computation is performed sequentially from n to 1, h i This represents the final hidden state of the bidirectional LSTM, and the forward hidden state. and backward hidden state The splicing result h i ∈R n*d ;

[0046] Finally, to further improve the model's ability to capture long-range dependencies, a self-attention mechanism is introduced to perform global feature modeling on the bidirectional LSTM output representation, thereby enhancing the model's ability to model long-range correlations in sequences. The specific method is as follows:

[0047] By learning the projection matrix W of query Q, key K, and value V. Q W K and W V The hidden state matrix H = [h1, h2, ..., h T Mapping to the new space, the formula for calculating the attention weight matrix A is as follows:

[0048]

[0049] Where Q = H·W Q K = H·W K and V = H·W V d k The dimension of the key-value vector is used to obtain a self-attention weighted representation through matrix multiplication, which incorporates local temporal information and provides shared features for subsequent intent recognition and semantic slot filling tasks.

[0050] Step S302: Construct an inter-correlation interaction module, including an intent and semantic slot label attention layer, a bidirectional interaction gating mechanism layer, and an intent recognition and semantic slot filling decoding layer. The inter-correlation interaction module applies the intent and slot label attention layer to obtain explicit intent and semantic slot representations, and adds intent information and semantic slot information through the bidirectional interaction gating mechanism. The interaction and connection between these layers yield the final representation of the intent and semantic slot information. The specific method is as follows:

[0051] First, in the label attention layer, for the predefined intent label set and slot label set, each label is first mapped to a continuous vector space through the label embedding matrix. Then, for the vector at each time step in the shared feature matrix, the explicit intent and semantic slot representation is obtained using the intent and slot label attention layer, as follows:

[0052] For intent representation, the sentence encoding representation C∈R obtained from the shared encoder is used. n*2d As a query vector, W I As a key-value vector, A is obtained after passing through an attention layer. I Finally, the intention representation H is obtained. I ∈R n*2d The specific formula is as follows:

[0053] A I =Softmax(C·W I )

[0054] H I =C+W I ·A I

[0055] For semantic slot representation, the module will use C∈R n*2d Treated as a query vector, W S As a key-value vector, A is obtained through attention interaction. S Then the slot representation H is obtained. S ∈R n*2d :

[0056] A S =Softmax(C·W S )

[0057] H S =C+W S ·A S

[0058] Then, the gating mechanism layer is used to model the relationship between semantic slots and intents, incorporating intent information into the semantic slot information; H S and H I The slot gate g is obtained as input to the slot gate. S g S To determine the degree of association between semantic slot information and intent information within the same part, g S A larger value indicates a stronger correlation between the semantic slot and the intent. Intent information makes the results of the semantic slot filling task more reliable. S With H S Multiplication yields the slot representation H guided by the intent information. S The formula is as follows:

[0059] g S =tanh(H S +W·H I )

[0060] H S '=g S ·H S

[0061] Where W is the trainable weight matrix, tanh(·) is the activation function, and g S The value ranges from (-1, 1);

[0062] H S and H I The intention gate g is obtained as the input to the slot gate. I , will g I and H I The intent to obtain slot information guidance is as follows:

[0063] g I =tanh(H I +W·H S )

[0064] H I '=g I ·H I

[0065] Where W is the trainable weight matrix, tanh(·) is the activation function, and g I The value ranges from (-1, 1);

[0066] In the intent recognition and semantic slot filling decoding layer, the intent representation obtained from the cross-correlation interaction module is processed using a maxpooling operation to obtain the final sentence representation r. I This is used to achieve the final intent classification, and the specific method is as follows:

[0067] r I =Maxpooling(H I ')

[0068]

[0069] Where W is the trainable weight matrix, O I Represents an intent-based label sequence;

[0070] In semantic slot filling, the semantic slot representation serves as the input for the final semantic slot filling process, which can be formalized as follows:

[0071]

[0072] Where W is the trainable weight matrix, O S This is the final sequence of semantic slot labels.

[0073] Step S303: Perform image retrieval on the intent information and semantic slot information obtained through cross-correlation interaction: First, to improve retrieval efficiency and speed, an image index library is pre-built using an image encoder. The image index library is represented as X = {I1, I2, ..., I...} K }, where I i Let I be the i-th image, and k be the number of images in the library; process each image I using an image encoder. i To obtain visual feature representation

[0074]

[0075] Then, the user intent is fed into a text encoder as a query sentence for processing to obtain a text feature representation. A natural language sentence T for the query is constructed using "intent-key information pairs". For the i-th query T in T... i A text encoder is used to process the text in order to obtain text feature representations.

[0076]

[0077] Furthermore, calculating text feature representations and image feature representation To retrieve related images, cosine similarity is used as a similarity metric, which measures the cosine of the angle between two vectors. and The cosine similarity between them is defined as:

[0078]

[0079] Where · represents the dot product, and |·| represents the vector norm;

[0080] Finally, the images obtained from the N queries will be used as the retrieval results. These results will be further processed in the image visual concept extraction module to integrate multimodal information into the conversation recommendation module, ultimately forming a vector representation of the user's intent. This vector representation can directly reflect the user's behavioral tendencies and needs preferences at the current moment.

[0081] Furthermore, step S4 includes:

[0082] Step S401: Construct a user session from each click sequence, then build a session graph from the session, and input it into a gated graph neural network to obtain the user's preferences. The specific method is as follows:

[0083] First, each click sequence of the user is constructed into a complete user session. The click sequence of a user in a session is known to be Q = {q1, q2, ..., q...}. n }, where q i This represents the i-th click event of the user in the session; each click event contains the clicked item, timestamp, and related context information. By organizing and sorting the above information, continuous click behaviors can be integrated into a time-ordered session.

[0084] Then, for the constructed session, each click event is treated as a node in the graph, and edges in the graph are constructed according to the temporal relationship of the clicks; specifically, a session graph G = (V, E) is constructed, where the set of nodes V corresponds to the click events {q1, q2, ..., q...}. n The edge set E reflects the temporal dependency between clicks, that is, for any two consecutive clicks q... i and q i+1 Construct a directed edge e in the graph. i To capture potential nonlocal dependencies, edges are established between discontinuous nodes based on a certain weight function, let the weight function be w(q). i ,q j The edge weight matrix can be represented as:

[0085]

[0086] After constructing the conversation graph G, it is input into a gated graph neural network to capture the user's dynamic preferences; in the gated graph neural network, each node v i Having an initial state h i (0) This state is represented by the feature vector of the click event, including but not limited to the content features of the click, time information, and context information; the main role of the gated graph neural network is to achieve the fusion of node information through multi-step message passing and state updates, and its iterative update formula is expressed as:

[0087]

[0088] Where N(i) represents the set of nodes adjacent to node i. Let h be the message aggregation vector of node i in the t-th iteration, and GRU(·) be the gated recurrent unit, which controls the transmission and forgetting of information through reset and update gates; after T iterations, the hidden state h of node i is... i T It not only incorporates its own initial features, but also integrates information from other click events associated with it. This process is considered to be a deep capture of the user's behavioral patterns and interests in a single session.

[0089] Finally, all node states are pooled to obtain the global embedding representation D of the session, as shown in the formula:

[0090]

[0091] Step S402: Iteratively aggregate the session embedding representation with the initial representation in the social network through a graph attention network to obtain the user's social embedding representation. Then, connect the session embedding and the social embedding to obtain the final user interest representation. Specifically, assume the session embedding representation obtained in step S401 is D∈R. d In social networks, the initial set of each friend relationship is denoted as {F}. i ∈R d } M i＝1 Where M represents the number of a user's friends; to efficiently integrate the session embedding representation with the friend relationship representation, a graph attention network is used, which assigns different weights to each friend through an attention mechanism to achieve adaptive aggregation of information. The specific method is as follows:

[0092] First, perform a linear transformation on the features of all nodes, denoted as Y∈R. d′×d The transformed user session features are h S =Y·D, friend characteristics h F =Y·F i By constructing the attention mechanism parameter vector 'a', the attention score between the target user and their friend is calculated using the following formula:

[0093] score=LeakyReLU(a T [h S ||h F ])

[0094] Where || denotes the vector concatenation operation, and LeakyReLU(·) is the activation function, defined as:

[0095] LeakyReLU(x) = Max(α*x,x)

[0096] Where α is the slope of the negative half-axis;

[0097] Normalize the scores of all friend nodes to obtain the attention coefficient α of the i-th friend. i :

[0098]

[0099] After obtaining the attention coefficients, a graph attention network is used to weight and aggregate the friend features to obtain a preliminary social embedding representation. The specific aggregation process is as follows:

[0100]

[0101] Wherein, σ(·) is a non-linear activation function, which ensures that the aggregated representation has good expressive power;

[0102] To further enhance the depth of information fusion, the process iterates at multiple levels, with the update formula for each level expressed as:

[0103]

[0104] Where l represents the iteration level, and the initial level l = 0 is {h F } (0) =h F Through multi-level iteration, the system can capture deeper social information from users and dynamically adjust the importance of friend information at each level.

[0105] Finally, the conversation embedding representation is concatenated with the iteratively obtained social embedding representation using vector concatenation to form the final user interest representation H. U :

[0106] H U =[h S ||h F ']

[0107] Step S403: Predict the user's rating for each item in the item set, and generate a recommendation list for the user based on the rating; the final user interest representation generated by fusing conversation embedding and social embedding is denoted as H. U The feature representation of each item in the candidate item set is denoted as . Where N represents the total number of candidate items, and to establish a mapping relationship between user interests and item characteristics, a rating prediction function is constructed, defined as:

[0108]

[0109] in, To predict the user's rating for item j, the function S(·) is implemented by concatenating the user's interest representation with the item features and then inputting the result into a multi-layer fully connected network. The specific method is as follows:

[0110] First, the user interest representation is concatenated with the project features, denoted as .

[0111] z j =[H U ||I j ]

[0112] Where || denotes the vector concatenation operation, resulting in the concatenated vector z. j A rating prediction network is constructed using nonlinear transformation. Given that the network contains L layers, the calculation formula for the l-th layer is:

[0113]

[0114] in, W (l) and b (l) Let represent the weight matrix and bias vector of the l-th layer, respectively, and σ(·) be the activation function. After the final layer transformation, the output vector is obtained. Then, it is mapped to a scalar score through a linear transformation, as shown in the formula:

[0115]

[0116] Among them, w r Let b be the weight vector for the rating mapping. r For rating mapping bias;

[0117] Predicted score This reflects the user's preference for item j; a higher value indicates greater user interest in the item. During network training, the rating prediction network performs supervised learning of predicted ratings based on a predefined loss function, defined as:

[0118]

[0119] Where, r j The system assigns the user's true rating to item j, and N represents the number of training samples. Through the backpropagation algorithm, the system continuously adjusts the parameters of each layer in the network to minimize the loss function, thereby improving the accuracy of rating prediction.

[0120] Furthermore, step S5 includes:

[0121] Step S501: First, collect user interaction feedback through multiple data sources, including explicit feedback (such as user ratings and reviews) and implicit feedback (such as click-through rate, dwell time, and conversion rate), to form a feedback dataset: T = {(x i ,s i ,y i ,τ i ,c i )} i＝1 N , where x i Let s represent the comprehensive input features of the i-th recommendation. i y is the recommended score generated by the system. i τ represents the actual feedback signal. i For timestamps, c i This indicates contextual information, where N represents the number of user feedback responses;

[0122] Step S502: After collecting feedback data, a hybrid error evaluation strategy is used to quantify the deviation between the prediction value of the recommendation model and the actual feedback. Simultaneously, a feedback loss function is constructed, defined as:

[0123]

[0124] Where y represents the actual user feedback value, which typically refers to user behavioral feedback, such as clicks, ratings, or other forms of feedback. The system predicts the recommended value, which represents the predicted value of the recommendation result given by the recommendation algorithm, and δ is the preset threshold.

[0125] The weighted loss function is obtained by weighting the feedback data:

[0126]

[0127] Among them, w i The coefficients weighted for each feedback data point are used to adjust its influence in the loss function according to the importance of different feedbacks, r i The recommendation score generated by the system is the model's prediction of how well the recommendation matches the user, and N represents the number of user responses.

[0128] Step S503: Based on the error assessment, a parameter update strategy is adopted to achieve online adaptive adjustment of the model. The parameter update process is as follows:

[0129]

[0130]

[0131] Where λ is the model update parameter, η is the learning rate, β1 and β2 are the decay rates of the first and second moments, respectively, ∈ is a small constant to prevent the denominator from being zero, and m t This is a first-moment estimate, i.e., an exponentially weighted moving average of the gradient, used to capture the overall direction of the gradient, v t This is a second-moment estimate, specifically an exponentially weighted moving average of the squared gradient, used to measure the variance of the gradient. This is the bias correction value for the first-order moment estimate. This is the bias correction value for the second-order moment estimation. The gradient of the loss function L” with respect to λ represents the directional derivative of the model parameters in the current iteration, which is used to guide parameter updates;

[0132] Step S504: To further improve the recommendation system's responsiveness to feedback signals, the system dynamically updates the probability distribution of each candidate recommendation option based on the current feedback reward distribution in each recommendation decision. Specifically, let the reward of candidate option j follow a Beta distribution Beta(α) j ,β j In each recommendation, the system samples the expected reward and selects the optimal recommendation accordingly, thereby achieving real-time online learning and adaptive adjustment.

[0133] The beneficial effects of this invention include:

[0134] (1) This invention accurately infers the user’s current needs and implicit intentions by comprehensively analyzing the user’s browsing history, click behavior, comment data and other multi-dimensional information, and combining the user’s input text and behavior patterns. Traditional recommendation systems mainly rely on the user’s historical behavior and are difficult to capture the user’s immediate needs. However, the system of this invention makes the recommendation results more accurate and personalized by deeply understanding the behavioral intentions behind the user’s behavior.

[0135] (2) The present invention designs a user intent recognition module that embeds user, item and interaction behavior information into the same space and effectively integrates multiple behavioral information through low-dimensional representation to form a more comprehensive user profile. This method not only enhances the understanding of user behavior intent, but also improves the flexibility and accuracy of the recommendation system. Through this integration of multi-dimensional information, the recommendation system can better cope with changes in user preferences and the needs of different scenarios, and has better adaptability and stability.

[0136] (3) Unlike traditional static recommendation methods, this invention introduces a real-time feedback mechanism, which can dynamically update user profiles and recommendation models. This adaptive mechanism ensures that the recommendation system can respond to changes in user needs in a timely manner and provide more efficient and personalized recommendation services. The user's immediate feedback will affect the recommendation results, further improving the system's flexibility and real-time performance, and enhancing the user experience.

[0137] (4) By designing an embedded propagation layer for behavior perception, the system can distinguish the importance of different behaviors in the recommendation process and capture high-order collaborative signals between users and products. This enables the recommendation system to not only make recommendations based on users' historical behavior, but also to deeply understand the complex interaction relationship between users and products, effectively improving the accuracy and robustness of recommendations and maintaining high recommendation performance in various user behavior scenarios.

[0138] (5) By integrating user social relationship data through graph attention network, the system can effectively solve the cold start problem of new users or low-activity users. Traditional recommendation systems perform poorly when new user behavior data is sparse, while this invention uses the friend behavior patterns in social networks as auxiliary information to construct an initial user profile. When user behavior data is insufficient, social embedding representation can improve recommendation accuracy and significantly reduce the proportion of random recommendations in the cold start stage.

[0139] (6) The dynamic update module adopts a hybrid error evaluation strategy. It distinguishes between normal feedback and outliers by setting a threshold and uses a piecewise loss function to process noisy data. When a user accidentally clicks or the temporary behavior deviates from the normal, the system can automatically reduce the gradient influence of the sample to avoid the problem of model overfitting.

[0140] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained from the following description and the foregoing claims. Attached Figure Description

[0141] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0142] Figure 1 This is a flowchart of a personalized recommendation method driven by user intent recognition according to the present invention;

[0143] Figure 2 This is a schematic diagram of One-hot encoding according to an embodiment of the present invention;

[0144] Figure 3 This is a schematic diagram of session graph construction according to an embodiment of the present invention;

[0145] Figure 4 This is a flowchart illustrating the specific workflow of the gated graph neural network according to an embodiment of the present invention.

[0146] Figure 5 This is a schematic diagram of the user social network layer in an embodiment of the present invention;

[0147] Figure 6 This is a diagram of the personalized recommendation interface of the system in an embodiment of the present invention;

[0148] Figure 7 This is a schematic diagram of a two-layer feedback mechanism triggered by user interaction in an embodiment of the present invention. Detailed Implementation

[0149] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0150] This invention provides a user intent recognition-driven personalized recommendation method, such as... Figure 1 As shown, the specific steps of the above method will be further explained below through a specific embodiment.

[0151] Step S1: Collect multi-dimensional information data of users to construct a user dataset; Step S1 specifically includes: the system collects user data through multiple channels, covering multimodal information from mobile devices, PCs and other smart terminals; the collected data includes not only explicit behaviors, such as search queries, clicks, favorites, ratings and comments, but also implicit behaviors, such as browsing time, dwell location, and operation trajectory; in addition, it collects contextual information related to the user's environment, such as geographical location, device type, timestamp and network environment.

[0152] Step S2 specifically includes the following sub-steps:

[0153] Step S201: Preprocess the collected user data, removing stop words and special characters from text data and standardizing word forms;

[0154] Taking the text input "Tomorrow I plan to visit a client in Beijing!" and a corresponding image as an example, this paper discusses data preprocessing and feature extraction in detail. First, in the text data preprocessing, conventional methods such as removing stop words, special characters, and standardizing word forms are used. During preprocessing, start and end markers are added to clearly define the text boundaries. Specifically, a "..." is added at the beginning of the text. <cls>"Tag, add " at the end <sep>The tag will then be appended to the end of the text if the text length does not meet the preset length requirement (e.g., the preset length is 8). <pad>The characters are then processed to form a standardized input sequence. In this example, the text sequence after the above processing is [ <cls>Tomorrow, I plan to visit a client in Beijing. <sep> ， <pad>This preprocessing not only improves the uniformity of the data, but also provides clear starting information for training sentence vectors in the subsequent model;

[0155] For image data, denoising, cropping, and size normalization are performed. The normalization formula is as follows:

[0156]

[0157] in, μ represents the pixel value of the original image. I and σ I These are the pixel mean and standard deviation, respectively, to ensure the consistency of the image in subsequent processing;

[0158] In stage S202, the system extracts features from the word segmentation sequence output from S201. During feature extraction, the text data is first segmented into word sequences T = {t1, t2, ..., t...} by the word segmentation tool. n }; Subsequently, the text data is processed into word vectors, and a pre-trained model is used to vectorize each word t i Mapping to a low-dimensional vector space yields the vectorized word v(t). i The specific method is as follows:

[0159] v(t i )=∑ j∈C(i) α ij ·P(t j )+β·S(t i )

[0160] For example, "tomorrow" is mapped to [0.1,0.2,0.3,0.4], "plan" is mapped to [0.2,0.1,0.4,0.3], "visit" is mapped to [0.3,0.4,0.1,0.2], "Beijing" is mapped to [0.4,0.3,0.2,0.1], and "customer" is mapped to [0.5,0.6,0.7,0.8];

[0161] Meanwhile, for image data, high-level features are extracted through a pre-trained convolutional neural network to obtain the feature vector Z = [0.2, 0.4, 0.6, 0.8], thereby capturing the semantic information in the image.

[0162] Furthermore, for text data, TF-IDF features are constructed by calculating term frequency and inverse document frequency; for example, if the text contains only the five semantic words mentioned above, and each word appears only once, then the term frequency TF(t) is calculated. i =1 / 5 = 0.2; In a corpus containing 100 documents, if each word appears in 50 documents, then IDF(t) = 1 / 5 = 0.2. i )=log2≈0.693, and then F(t) can be calculated. i = 0.2 × 0.693 ≈ 0.1386.

[0163] In stage S203, a feature concatenation strategy is adopted to integrate text features with image features. Specifically, the vector of each text word is directly concatenated with the image feature vector to form a joint feature vector. For example, for the word "tomorrow," its word vector is [0.1, 0.2, 0.3, 0.4], and the image feature vector [0.2, 0.4, 0.6, 0.8] is concatenated to generate the joint feature vector [0.1, 0.2, 0.3, 0.4, 0.2, 0.4, 0.6, 0.8]. The system then fuses the text features extracted in stage S202 with other potential contextual features to construct a joint representation. in, Z represents the word vector sequence obtained through a pre-trained embedding model, and Z represents the corresponding image feature. This joint representation not only preserves the semantic and contextual information of the original text, but also enhances the saliency of key information through feature weighting, thereby providing comprehensive and detailed data information for the accurate identification of user intent and system response.

[0164] Step S3: By analyzing the user's input text and behavioral patterns, infer the user's current needs and identify the user's intent.

[0165] In step S301, assuming the user submits a service task request of "planning to visit a certain client in Beijing tomorrow," the system first preprocesses the original text input, discretizing the words "tomorrow," "plan," "visit," "go," "Beijing," and "certain client" into their respective integer indices. Then, it uses One-hot encoding to convert them into a fixed-dimensional vector representation, such as... Figure 2 As shown; finally, through the pre-trained word embedding matrix, the discrete vectors are mapped to a continuous low-dimensional vector space to form preliminary word vectors;

[0166] To further capture word order information, sine and cosine positional encoding is introduced. The positional encoding is added element-wise to the word vector to generate the final embedding vector containing both semantic and positional information. The specific method is as follows:

[0167]

[0168] Where p is the position of the word in the sequence, i is the dimension index, d is the dimension of the embedding vector, and C represents a large constant; the positional encoding is added element by element to the word vector to obtain the final input embedding vector sequence that contains both semantic and positional information.

[0169] Then, the embedded sequence is fed into a bidirectional LSTM network. In the forward LSTM, the sequence is passed from front to back to generate the forward hidden state. In the inverse LSTM, the sequence is passed from back to front to generate the inverse hidden state. By concatenating the forward and backward hidden states at each time step, a hidden representation incorporating contextual information is obtained. The specific method is as follows:

[0170] In the bidirectional LSTM layer, for each input sentence T i =(t1,t2,…,t n The word representation E = (e1, e2, ..., e) is obtained through word embedding. n )∈R n*d As input, from e1 to e n and from e n E1 is then used for forward and reverse encoding; forward encoding generates the hidden state. Reverse encoding generates hidden states The final generated encoded representation h i Depend on and It is pieced together, and the formula is as follows:

[0171]

[0172] in, For forward LSTM units, computation is performed sequentially from 1 to n. For backward LSTM units, computation is performed sequentially from n to 1, h i This represents the final hidden state of the bidirectional LSTM, and the forward hidden state. and backward hidden state The splicing result h i ∈R n*d ;

[0173] To further capture long-range dependencies, the system utilizes a self-attention mechanism to perform global feature modeling on the matrix H composed of all hidden states; and learns the projection matrix W of the query, key, and value. Q W K and W V The hidden state matrix H = [h1, h2, ..., h T Mapping to the new space, the formula for calculating the attention weight matrix A is as follows:

[0174]

[0175] Where Q = H·W Q K = H·W K and V = H·W V d k The dimension of the key-value vector is used to obtain a self-attention weighted representation through matrix multiplication, which incorporates local temporal information and provides shared features for subsequent intent recognition and semantic slot filling tasks.

[0176] Step S302: Construct an inter-correlation interaction module, including an intent and semantic slot label attention layer, a bidirectional interaction gating mechanism layer, and an intent recognition and semantic slot filling decoding layer. The inter-correlation interaction module applies the intent and slot label attention layer to obtain explicit intent and semantic slot representations, and adds intent information and semantic slot information through the bidirectional interaction gating mechanism. The interaction and connection between these layers yield the final representation of the intent and semantic slot information. The specific method is as follows:

[0177] First, in the label attention layer, for the predefined intent label set and slot label set, each label is first mapped to a continuous vector space through the label embedding matrix. Then, for the vector at each time step in the shared feature matrix, the explicit intent and semantic slot representation is obtained using the intent and slot label attention layer, as follows:

[0178] For intent representation, the sentence encoding representation C∈R obtained from the shared encoder is used. n*2d As a query vector, W I As a key-value vector, A is obtained after passing through an attention layer. I Finally, the intention representation H is obtained. I ∈R n*2d The specific formula is as follows:

[0179] A I =Softmax(C·W I )

[0180] H I =C+W I ·A I

[0181] For semantic slot representation, the module will use C∈R n*2d Treated as a query vector, W S As a key-value vector, A is obtained through attention interaction. S Then the slot representation H is obtained. S ∈R n*2d :

[0182] A S =Softmax(C·W S )

[0183] H S =C+W S ·A S

[0184] Then, the gating mechanism layer is used to model the relationship between semantic slots and intents, incorporating intent information into the semantic slot information; H S and H I The slot gate g is obtained as input to the slot gate. S g S To determine the degree of association between semantic slot information and intent information within the same part, g S A larger value indicates a stronger correlation between the semantic slot and the intent. Intent information makes the results of the semantic slot filling task more reliable. S With H S Multiplication yields the slot representation H guided by the intent information. S The formula is as follows:

[0185] g S =tanh(H S +W·H I )

[0186] H S '=g S ·H S

[0187] Where W is the trainable weight matrix, tanh(·) is the activation function, and g S The value ranges from (-1, 1);

[0188] H S and H I The intention gate g is obtained as the input to the slot gate. I , will g I and H I The intent to obtain slot information guidance is as follows:

[0189] g I =tanh(H I +W·H S )

[0190] H I '=g I ·H I

[0191] Where W is the trainable weight matrix, tanh(·) is the activation function, and g I The value ranges from (-1, 1);

[0192] In the intent recognition and semantic slot filling decoding layer, the intent representation obtained from the cross-correlation interaction module is processed using a maxpooling operation to obtain the final sentence representation r. I This is used to achieve the final intent classification, and the specific method is as follows:

[0193] r I =Maxpooling(H I ')

[0194]

[0195] Where W is the trainable weight matrix, O I Represents an intent-based label sequence;

[0196] In semantic slot filling, the semantic slot representation serves as the input for the final semantic slot filling process, which can be formalized as follows:

[0197]

[0198] Where W is the trainable weight matrix, O S This is the final sequence of semantic slot labels.

[0199] Step S303: Perform image retrieval on the intent information and semantic slot information obtained through cross-correlation interaction: First, to improve retrieval efficiency and speed, an image index library is pre-built using an image encoder. The image index library is represented as X = {I1, I2, ..., I...} K }, where I i Let I be the i-th image, and k be the number of images in the library; process each image I using an image encoder. i To obtain visual feature representation

[0200]

[0201] Then, the user intent is fed into a text encoder as a query sentence for processing to obtain a text feature representation. A natural language sentence T for the query is constructed using "intent-key information pairs". For the i-th query T in T... i A text encoder is used to process the text in order to obtain text feature representations.

[0202]

[0203] Furthermore, calculating text feature representations and image feature representation To retrieve related images, cosine similarity is used as a similarity metric, which measures the cosine of the angle between two vectors. and The cosine similarity between them is defined as:

[0204]

[0205] Where · represents the dot product, and |·| represents the vector norm;

[0206] Finally, the images obtained from the N queries will be used as the retrieval results. These results will be further processed in the image visual concept extraction module to integrate multimodal information into the conversation recommendation module, ultimately forming a vector representation of the user's intent. This vector representation can directly reflect the user's behavioral tendencies and needs preferences at the current moment.

[0207] In this embodiment, step S4 includes:

[0208] Step S401: Construct a user session from each click sequence, then build a session graph from the session, and input it into a gated graph neural network to obtain the user's preferences. The specific method is as follows:

[0209] First, each click sequence of the user is constructed into a complete user session. The click sequence of a user in a session is known to be Q = {q1, q2, ..., q...}. n }, where q i This represents the i-th click event of the user in the session; each click event contains the clicked item, timestamp, and related context information. By organizing and sorting the above information, continuous click behaviors can be integrated into a time-ordered session.

[0210] Then, for the constructed session, each click event is treated as a node in the graph, and edges in the graph are constructed according to the temporal relationship of the clicks; specifically, a session graph G = (V, E) is constructed, where the set of nodes V corresponds to the click events {q1, q2, ..., q...}. n The edge set E reflects the temporal dependency between clicks, that is, for any two consecutive clicks q... i and q i+1 Construct a directed edge e in the graph. i To capture potential nonlocal dependencies, edges are established between discontinuous nodes based on a certain weight function, let the weight function be w(q). i ,q j The edge weight matrix can be represented as:

[0211]

[0212] After constructing the conversation graph G, it is input into a gated graph neural network to capture the user's dynamic preferences; in the gated graph neural network, each node v i Having an initial state h i (0) This state is represented by the feature vector of the click event, including but not limited to the content features of the click, time information, and context information; the main role of the gated graph neural network is to achieve the fusion of node information through multi-step message passing and state updates, and its iterative update formula is expressed as:

[0213]

[0214] Where N(i) represents the set of nodes adjacent to node i, m i t Let h be the message aggregation vector of node i in the t-th iteration, and GRU(·) be the gated recurrent unit, which controls the transmission and forgetting of information through reset and update gates; after T iterations, the hidden state h of node i is... i T It not only incorporates its own initial features, but also integrates information from other click events associated with it. This process is considered to be a deep capture of the user's behavioral patterns and interests in a single session.

[0215] Finally, all node states are pooled to obtain the global embedding representation D of the session, as shown in the formula:

[0216]

[0217] Step S402: Iteratively aggregate the session embedding representation with the initial representation in the social network through a graph attention network to obtain the user's social embedding representation. Then, connect the session embedding and the social embedding to obtain the final user interest representation. Specifically, assume the session embedding representation obtained in step S401 is D∈R. d In social networks, the initial set of each friend relationship is denoted as {F}. i ∈R d } M i＝1 Where M represents the number of a user's friends; to efficiently integrate the session embedding representation with the friend relationship representation, a graph attention network is used, which assigns different weights to each friend through an attention mechanism to achieve adaptive aggregation of information. The specific method is as follows:

[0218] First, perform a linear transformation on the features of all nodes, denoted as Y∈R. d′×d The transformed user session features are h S =Y·D, friend characteristics h F =Y·F i By constructing the attention mechanism parameter vector 'a', the attention score between the target user and their friend is calculated using the following formula:

[0219] score=LeakyReLU(a T [h S ||h F ])

[0220] Where || denotes the vector concatenation operation, and LeakyReLU(·) is the activation function, defined as:

[0221] LeakyReLU(x) = Max(α*x,x)

[0222] Where α is the slope of the negative half-axis;

[0223] Normalize the scores of all friend nodes to obtain the attention coefficient α of the i-th friend. i :

[0224]

[0225] After obtaining the attention coefficients, a graph attention network is used to weight and aggregate the friend features to obtain a preliminary social embedding representation. The specific aggregation process is as follows:

[0226]

[0227] Wherein, σ(·) is a non-linear activation function, which ensures that the aggregated representation has good expressive power;

[0228] To further enhance the depth of information fusion, the process iterates at multiple levels, with the update formula for each level expressed as:

[0229]

[0230] Where l represents the iteration level, and the initial level l = 0 is {h F } (0) =h F Through multi-level iteration, the system can capture deeper social information from users and dynamically adjust the importance of friend information at each level.

[0231] Finally, the conversation embedding representation is concatenated with the iteratively obtained social embedding representation using vector concatenation to form the final user interest representation H. U :

[0232] H U =[h S ||h F ']

[0233] Step S403: Predict the user's rating for each item in the item set, and generate a recommendation list for the user based on the rating; the final user interest representation generated by fusing conversation embedding and social embedding is denoted as H. U The feature representation of each item in the candidate item set is denoted as . Where N represents the total number of candidate items, and to establish a mapping relationship between user interests and item characteristics, a rating prediction function is constructed, defined as:

[0234]

[0235] in, To predict the user's rating for item j, the function S(·) is implemented by concatenating the user's interest representation with the item features and then inputting the result into a multi-layer fully connected network. The specific method is as follows:

[0236] First, the user interest representation is concatenated with the project features, denoted as .

[0237] z j =[H U ||I j ]

[0238] Where || denotes the vector concatenation operation, resulting in the concatenated vector z. j A rating prediction network is constructed using nonlinear transformation. Given that the network contains L layers, the calculation formula for the l-th layer is:

[0239]

[0240] Among them, s j (l-1) =z j W (l) and b (l) Let represent the weight matrix and bias vector of the l-th layer, respectively, and σ(·) be the activation function. After the final layer transformation, the output vector is obtained. Then, it is mapped to a scalar score through a linear transformation, as shown in the formula:

[0241]

[0242] Among them, w r Let b be the weight vector for the rating mapping. r For rating mapping bias;

[0243] Predicted score This reflects the user's preference for item j; a higher value indicates greater user interest in the item. During network training, the rating prediction network performs supervised learning of predicted ratings based on a predefined loss function, defined as:

[0244]

[0245] Where, r j The system uses the user's actual rating for item j, where N represents the number of training samples. Through backpropagation, the system continuously adjusts the parameters of each layer in the network to minimize the loss function, thereby improving the accuracy of rating prediction. Figure 6 As shown, when a user submits a service task request of "planning to visit a client in Beijing tomorrow", the system performs a comprehensive and detailed analysis based on the user's intent and provides a series of targeted recommendations to help the user successfully complete the visit task.

[0246] Step S5 specifically includes the following steps:

[0247] In step S501, user interaction feedback is obtained by integrating multiple data sources, such as... Figure 7 As shown, the dataset includes explicit feedback (explicit feedback is the behavior of users directly expressing their preferences for items or content, such as user ratings and reviews; its advantage is that the data directly reflects user preferences, but its disadvantage is that the data volume may be small and it is affected by user subjectivity and social factors) and implicit feedback (implicit feedback is the behavior of users when using the system, indirectly reflecting user preferences, such as click-through rate, dwell time, and conversion rate; its advantage is that the data volume is large and can reflect user behavior in real time, but its disadvantage is that the data noise is high and it is difficult to judge the user's true preferences, so it needs to be used in conjunction with explicit feedback for comprehensive interpretation), forming a feedback dataset: T = {(x i ,s i ,y i ,τ i ,c i )} i＝1 N , where x i Let s represent the comprehensive input features of the i-th recommendation. i y is the recommended score generated by the system. i τ represents the actual feedback signal. i For timestamps, c i This indicates contextual information, where N represents the number of user feedback responses;

[0248] Step S502: After collecting feedback data, a hybrid error evaluation strategy is used to quantify the deviation between the prediction value of the recommendation model and the actual feedback. Simultaneously, a feedback loss function is constructed, defined as:

[0249]

[0250] Where y represents the actual user feedback value, which typically refers to user behavioral feedback, such as clicks, ratings, or other forms of feedback. The system predicts the recommended value, which represents the predicted value of the recommendation result given by the recommendation algorithm, and δ is the preset threshold.

[0251] The weighted loss function is obtained by weighting the feedback data:

[0252]

[0253] Among them, w i The coefficients weighted for each feedback data point are used to adjust its influence in the loss function according to the importance of different feedbacks, r i The system-generated recommendation score is the model's predicted value regarding the match between the recommendation item and the user, where N represents the number of user feedback responses. Assuming a recommendation item has an actual rating of y = 4 points, the system predicts the recommendation score... If a threshold δ = 0.7 is set and the weight w of the data point is... i =1.2, then the error |4.0-3.5| = 0.5≤0.7. Using a double penalty, the loss value is calculated as 1 / 2(0.5). 2 =0.125, and the weighted loss of this feedback data point is 1.2 × 0.125 = 0.15.

[0254] In step S503, the model parameters are adaptively adjusted online using a parameter update strategy based on the Adam optimization algorithm. The parameter update process is as follows:

[0255]

[0256] Where λ is the model update parameter, η is the learning rate, β1 and β2 are the decay rates of the first and second moments, respectively, ∈ is a small constant to prevent the denominator from being zero, and m t This is a first-moment estimate, i.e., an exponentially weighted moving average of the gradient, used to capture the overall direction of the gradient, v t This is a second-moment estimate, specifically an exponentially weighted moving average of the squared gradient, used to measure the variance of the gradient. This is the bias correction value for the first-order moment estimate. This is the bias correction value for the second-order moment estimation. The gradient of the loss function L” with respect to λ represents the directional derivative of the model parameters in the current iteration, which is used to guide parameter updates;

[0257] Suppose that in a certain iteration, the gradient The first moment m at the previous moment is 0.2. t-1 =0.1, second moment v t-1 =0.04, known: β1=0.9, β2=0.999, eta=0.01, ∈=10 -8 Then v can be calculated sequentially. t = 0.999 × 0.04 + 0.001 × (0.2) 2 =0.04 + 0.00004 = 0.04004, m t =0.9×0.1+0.1×0.2=0.11, and the model parameters are updated after correction.

[0258] In step S504, to further improve the system's responsiveness to feedback signals, the present invention dynamically updates the probability distribution of each recommendation candidate based on the current feedback reward distribution in each recommendation decision; specifically, it is assumed that the reward of a candidate j follows a Beta distribution Beta(α) j ,β j In each recommendation, the expected reward is obtained by sampling from the Beta distribution, thereby selecting the optimal recommendation candidate to achieve real-time online learning and adaptive adjustment; if the parameter of a candidate is α... j =2 and β j If the expected reward is 2 / (2+3) = 0.4, then the expected reward is 2 / (2+3) = 0.4. Among multiple candidate options, the system dynamically updates the probability distribution of each candidate option by comparing the sampled expected reward values, and selects the item with the highest matching degree accordingly. Through the above steps, this invention achieves comprehensive collection of user feedback data, error assessment, adaptive parameter updating, and dynamic optimization of recommended candidate options, thereby significantly improving the response speed and accuracy of the recommendation system under various interactive behaviors.

[0259] Traditional recommendation systems often treat user behavior as static data, lacking a response mechanism to dynamic changes in user behavior and context switching. The user intent recognition-driven personalized recommendation method of this invention addresses this problem through technological innovation. In the preprocessing stage, it integrates multi-dimensional information such as user browsing history, click behavior, and comment data to construct a more comprehensive dataset. Through semantic parsing and behavioral pattern analysis, it identifies the user's current true needs and intentions. Building upon this, this invention further designs a user intent recognition module, mapping users, items, and corresponding interactive behaviors into low-dimensional continuous representations, constituting the initial input features of the recommendation model. This approach not only captures the basic attributes of users and products but also unifies and integrates various behavioral information, providing rich semantic information for subsequent recommendation algorithms. Simultaneously, the model combines graph structure information and high-order collaborative signals from multi-behavioral interactions, effectively distinguishing the importance of different behaviors through embedded propagation layers and behavioral perception mechanisms, achieving deep feature interaction between users, items, and interaction relationships. This design makes the model more accurate in capturing implicit behavioral semantics and the user's true intent, thereby significantly improving the response speed and personalization of the recommendation system.

[0260] This invention, through improvements to traditional recommendation techniques and the introduction of deep learning methods, not only overcomes the shortcomings of traditional methods in capturing users' true needs but also achieves multi-dimensional data fusion and real-time dynamic updates. This technology demonstrates significant advantages in improving recommendation accuracy, enhancing system robustness, and increasing user satisfaction. Simultaneously, it provides a solid theoretical foundation and technical support for future intelligent recommendation and precision marketing, possessing broad application prospects and profound industrial value.

[0261] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0262] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0263] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0264] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0265] Finally, 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.< / pad> < / sep> < / cls> < / pad> < / sep> < / cls> < / sep> < / cls> < / pad>

Claims

1. A user intention recognition driven personalized recommendation method, characterized in that: The method comprises the following steps: Step S1: collecting multi-dimensional information data of the user, and constructing a user behavior data set; Step S2: preprocessing the collected user data to obtain a session sequence; Step S3: inferring the current demand of the user by analyzing the input text and behavior mode of the user, and identifying the intention of the user; Step S4: constructing a user interest preference prediction model according to the identified user intention; Step S5: dynamically updating the recommendation model in combination with the real-time feedback of the user, and improving the response speed and individualization degree of the recommendation system. 2.The user intention recognition driven personalized recommendation method of claim 1, wherein: The step S1 specifically comprises: The system collects user data through multiple channels, covering multi-modal information from mobile devices, PC terminals and other smart terminals; the collected multi-dimensional information data includes explicit behavior and implicit behavior; in addition, context information related to the environment of the user is also collected. 3.The user intention recognition driven personalized recommendation method of claim 1, wherein: The step S2 comprises: Step S201: preprocessing the collected user data, including preprocessing operations of text data and image data: (1) for text data, stop words, special characters are removed, and word forms are unified, the specific process includes two steps of completion to a fixed length and addition of a start mark; After the processed text is input into the model, the specific vector output is adjusted according to the downstream task and the context information; (2) for image data, denoising, cropping and size normalization processing are performed, and the normalization formula is: where I(x, y) represents the pixel value of the original image, μ I and σ I are the pixel mean and standard deviation, respectively, to ensure consistency of the image in subsequent processing; Step S202: respectively perform feature extraction operation on the text data, to obtain a word sequence T={t1, t2, …, t n}, wherein t n represents the nth word; for the image data I, which is composed of image pixel values I(x, y), the system extracts high-level features of the image using a pre-trained convolutional neural network, to obtain a feature vector to capture the semantic information of the image, wherein represents the image after normalization processing; Step S203, word vectorization processing is performed on the text data, and each word t i is mapped to a low-dimensional vector space by using a pre-trained model to obtain a vectorized word v(t i ). The specific method is as follows: v(t i ) = ∑ j∈C(i) α ij ·P(t j )+β·S(t i ) where C(i) is the context window of word t i , P(t j ) represents the mapping function of the pre-trained word embedding model for word t j , a ij represents the attention weight, which is calculated by the Softmax(f(P(t i ), P(t j ))) function, f(·) is a similarity measure function, and S(t i ) represents the part-of-speech tagged embedding vector; β represents the weight coefficient of the part-of-speech feature; The text is represented as a sequence of word vectors V(T) = {v(t1), v(t2),..., v(t n )} and the feature vector Z of the image has been obtained in the previous step. The multi-modal features are constructed and fused: First, for text data, TF-IDF features are obtained by computing term frequency TF(t i ) and inverse document frequency IDF(t i ). The specific method is as follows: F(t i ) = TF(t i ) · IDF(t i ) wherein, is the word t i occurs, N is the total number of documents in the document collection, df(t i ) is the number of documents containing the word t i . Then, a statistically weighted text feature matrix is generated: M = [F(t1)*v(t1), F(t2)*v(t2),..., F(t n )*v(t n )] Finally, the text feature V(T) and the image feature Z are combined to form a joint feature representation by using a feature fusion strategy: wherein mean() performs mean-pooling on the weighted text features, U denotes the multi-modal information in n different scenarios, wherein the information in each scenario is composed of v(t i ) and the corresponding image feature z i ∈Z, and R represents a set of real numbers. 4.The user intention recognition driven personalized recommendation method of claim 1, wherein: The step S3 comprises: Step S301: on the basis of obtaining the word embedding of the service task request input sequence, first, the shared features of intent recognition and semantic slot filling are extracted by using a bidirectional LSTM network and a self-attention mechanism; in the input layer of the model, the classification value is mapped to the corresponding integer value, and each integer value is represented by a binary vector; then, the relative or absolute position of each word in the entire sentence is recorded by position encoding to learn the positional relationship between words; finally, the word vector and the position encoding are added to generate a word embedding vector as the input of the model; Step S302: constructing a mutual correlation interaction module, which includes a label attention layer and a mutual correlation interaction layer, the label attention layer further extracts intent information features and semantic slot information, and the mutual correlation interaction layer realizes the interaction of the two kinds of information by using a gating mechanism; Step S303: image retrieval of the intention information and semantic slot information interacted by cross-correlation respectively: first, in order to improve the retrieval efficiency and speed, an image index library is constructed in advance using an image encoder, and the image index library is represented as X={I1, I2, …, In}, wherein I K is the i-th image, and k is the number of images in the library; each image I i is processed by using an image encoder to obtain a visual feature representation i . Then, the user intent is sent into the text encoder as a query sentence for processing to obtain a text feature representation, and a natural language sentence T of the query is constructed using the "intent-key information pair", for the i-th query T in T i is processed using the text encoder to obtain a text feature representation Furthermore, the similarity between the text feature representation and the image feature representation is calculated to retrieve relevant images, using cosine similarity as the similarity measure method, measuring the cosine value of the angle between two vectors, and the cosine similarity between them is defined as: Wherein, · represents dot product, and |·| represents vector norm; Finally, the images of the N queries are obtained as retrieval results, which will be further processed in the image visual concept extraction module to integrate the multi-modal information into the session recommendation module, and finally form a vector representation of the user intention, which directly reflects the behavior tendency and demand preference of the user at the current moment.

5. The user intent recognition driven personalized recommendation method of claim 4, wherein: The specific steps in the step S301 are: Firstly, the original text is obtained from the service task request input sequence, and each word in the text is mapped to a continuous vector space using word embedding technology; the classification values in the text are discretized, each category is mapped to a unique integer index, and then binary encoding or One-Hot encoding is used to convert it into a fixed-dimensional vector representation, which is embedded into a low-dimensional space using a word embedding matrix to form an initial word vector; In order to further capture the word order information, the sine-cosine position encoding is introduced, and the position encoding is added to the word vector element by element to generate the final embedding vector containing semantic and position information, the specific method is: Where p is the position of the word in the sequence, i is the dimension index, d is the dimension of the embedding vector, and C represents a large constant; Then, the fused embedding vectors are input into a bidirectional LSTM network, through the double transmission of forward and reverse information, respectively generating forward hidden states and reverse hidden states and splicing the two to obtain the overall context representation The specific method is: At the bi-directional LSTM layer for each input sentence T i = (t1, t2,..., t n ) get word representation E = (e1, e2,..., e n )∈R n*d As input, forward encoding and reverse encoding are performed from e1 to e n and from e n to e1 respectively; the forward encoding generates hidden state The reverse encoding generates hidden state The final generated encoding representation h i is spliced by and , the formula is as follows: wherein, is a forward LSTM unit computed in order from 1 to n, is a backward LSTM unit computed in order from n to 1, h i is the final hidden state of the bidirectional LSTM, h is the concatenation of the forward hidden state h and the backward hidden state h i ∈ R n*d ; Finally, in order to further improve the model's ability to capture long-distance dependencies, a self-attention mechanism is introduced to model the global features of the bidirectional LSTM output to enhance the model's ability to model long-distance relationships in the sequence, the specific method is: By learning the projection matrices W of the query Q, the key K, and the value V Q , W K , and W V , the hidden state matrix H = [h1, h2, …, h T ] is mapped into a new space, and the attention weight matrix A is calculated as follows: where Q = H·W Q , K = H·W K and V = H·W V , d k is the dimension of the key-value vector, and finally the self-attention weighted representation is obtained through matrix multiplication, which integrates local temporal information and provides shared features for subsequent intent recognition and semantic slot filling tasks. 6.The user intention recognition driven personalized recommendation method of claim 4, wherein: The step S302 of constructing the mutual correlation interaction module includes an intent and semantic slot label attention layer, a bidirectional interaction gate mechanism layer, and an intent recognition and semantic slot filling decoding layer. The mutual correlation interaction module applies the intent and slot label attention layer to obtain explicit intent and semantic slot representation, increases the intent information and semantic slot information through the bidirectional interaction gate mechanism, and interacts and connects to obtain the final representation of the intent and semantic slot information. The specific method is: Firstly, in the label attention layer, for the pre-defined intent label set and slot label set, first map each label to a continuous vector space through a label embedding matrix; then, for each time step vector in the shared feature matrix, use the intent and slot label attention layer to obtain explicit intent and semantic slot representation, the specific method is as follows: For intent representation, the sentence encoding representation C ∈ R n*2d As query vector, W I As key and value vector, A I , finally get the intent representation H I ∈ R n*2d The specific formula is as follows: A I = Softmax(C · W I ) H I = C + W I ·A I For semantic slot representation, the module will C ∈ R n*2d W S As key and value vectors, after attention interaction, A S Then get semantic slot representation H S ∈ R n*2d : A S = Softmax(C · W S ) H S = C + W S ·A S Then, the gate mechanism layer is used to model the semantic slot and intent relationship, and the intent information is introduced into the semantic slot information. H S and H I The slot gate g is obtained as input to the slot gate. S g S To determine the degree of association between semantic slot information and intent information within the same part, g S A larger value indicates a stronger correlation between the semantic slot and the intent. Intent information makes the results of the semantic slot filling task more reliable. S With H S Multiplication yields the slot representation H guided by the intent information. S The formula is as follows: g S = tanh(H S + W · H I ) H S '=g S ·H S where W is a trainable weight matrix, tanh(·) is an activation function, g S ranges between (-1, 1); H S and H I get intent gate g as input to slot gate I , g I and H I get slot information guided intent representation: g I = tanh(H I + W · H S ) H I '=g I ·H I where W is a trainable weight matrix, tanh(·) is an activation function, g I ranges between (-1, 1); In the intent recognition and semantic slot filling decoding layer, the intent representation obtained by the interrelated interaction module is subjected to a Maxpooling operation to obtain a final sentence representation r I to implement the final intent classification, and the specific method is as follows: r I = Maxpooling(H I ') where W is a trainable weight matrix, O I representative intent label sequence; For semantic slot filling, the semantic slot representation is used as the input of the final semantic slot filling, and the semantic slot filling process is formalized as follows: where W is a trainable weight matrix, O S is the final semantic slot label sequence.

7. The user intent recognition driven personalized recommendation method of claim 1, wherein: The step S4 includes: Step S401: Construct each click sequence of the user into a user session, then construct the session into a session graph, and input it into the gated graph neural network to obtain the user's preference, the specific method is as follows: First, each click sequence of a user is constructed into a complete user session, and the click sequence of a user in a session is known as Q = {q1, q2,..., q n}, where q i represents the i-th click event of the user in the session; each click event contains the clicked item, timestamp and related context information, and the continuous click behavior is integrated into a time-ordered session by sorting and arranging the above information; Then, for the constructed session, each click event is treated as a node in the graph, and edges in the graph are constructed according to the temporal relationship of the clicks; specifically, a session graph G = (V, E) is constructed, where the set of nodes V corresponds to the click events {q1, q2, ..., q...}. n The edge set E reflects the temporal dependency between clicks, that is, for any two consecutive clicks q... i and q i+1 Construct a directed edge e in the graph. i To capture potential nonlocal dependencies, edges are established between discontinuous nodes based on a certain weight function, let the weight function be w(q). i ,q j The edge weight matrix is ​​represented as follows: After constructing the session graph G, it is input into the gated graph neural network to capture the dynamic preferences of the user; in the gated graph neural network, each node v i has an initial state h i (0) , which is represented by a feature vector of the click event, including but not limited to the content features of the click, time information and context information; the main role in the gated graph neural network is to realize the fusion of node information through multi-step message passing and state updating, and its iterative update formula is expressed as: where N(i) denotes the set of nodes adjacent to node i, is the message aggregation vector for node i in the t-th iteration, and GRU(·) is the Gated Recurrent Unit that controls the transmission and forgetting of information through the reset gate and update gate. After T iterations, the hidden state h i T Not only the initial features of itself are integrated, but also the information of other click events associated with it, which is considered as a deep capture of the user's behavior patterns and interest preferences in a single session. Finally, the node states are pooled to obtain the global embedding representation D of the session, the formula is: Step S402: The session embedding representation is iteratively aggregated with the friend relationship representation in the social network through the graph attention network to obtain the social embedding representation of the user, and the final user interest representation is obtained by connecting the session embedding and the social embedding; Step S403: Predict the rating of each item in the item set by the user, and generate a recommendation list for the user according to the rating. 8.The user intention recognition driven personalized recommendation method of claim 7, wherein: The step S402 includes the following specific steps: assuming that the session embedding obtained in the step S401 is represented as D∈R d , and in the social network, each initial friend relationship set is denoted as {F i ∈R d} M i＝1 , where M represents the number of user friends; in order to efficiently fuse the session embedding representation and the friend relationship representation, a graph attention network is adopted, different weights are given to each friend through an attention mechanism, adaptive aggregation of information is realized, and the specific method is as follows: First, a linear transformation is performed on the features of all nodes, denoted as a transformation matrix Y ∈ R d′×d , and the transformed user session features are h S = Y · D, and the friend features are h F = Y · F i . The attention score between the target user and the friends is calculated by constructing an attention mechanism parameter vector a, and the formula is as follows: score = LeakyReLU(a T [h S ||h F ]) Where || represents the vector concatenation operation, and LeakyReLU(·) is an activation function defined as: LeakyReLU(x)=Max(α*x,x) Where alpha is the negative half-axis slope. The scores of all friend nodes are normalized to obtain the attention coefficient a of the ith friend i : After obtaining the attention coefficient, the friend features are weighted and aggregated by the graph attention network to obtain the preliminary social embedding representation, and the specific aggregation process is: wherein σ(·) is a nonlinear activation function, and the aggregated representation has good expression ability; In order to further improve the depth of information fusion, the process is iterated at multiple levels, and the update formula of each layer is represented as: Wherein, l represents the iteration layer number, the initial layer l = 0 when {h F} (0) = h F ; through multi-layer iteration, the system captures the deeper social information of the user, and dynamically adjusts the importance of the friend information at each layer; Finally, the session embedding representation is concatenated with the iteratively obtained social embedding representation, here in the form of vector concatenation, to form the final user interest representation H U : H U = [h S ||h F ']. 9.The user intention recognition driven personalized recommendation method of claim 7, wherein: The specific steps of capturing the implicit behavior semantics in step S403 are: The final user interest representation generated by the fusion of session embedding and social embedding is denoted as H U And the feature representation of each item in the candidate item set is denoted as Where N represents the total number of candidate items, in order to establish the mapping relationship between user interest and item features, a score prediction function is constructed, defined as: wherein, is the predicted rating of user i for item j, the function S(·) is implemented by concatenating the user interest representation and the item feature and inputting them into a multi-layer fully connected network, and the specific method is as follows: First, the user interest representation is spliced with the item features, denoted as z j = [H U ||I j ] Wherein, || represents the vector splicing operation, and a spliced vector z is obtained j , a score prediction network is constructed by using a nonlinear transformation. Given that the network contains L layers, for the lth layer, the calculation formula is: s j (l) = σ(W (l) · s j (l-1) + b (l) ) where s j (l-1) = z j , W (l) and b (l) are the weight matrix and bias vector of the l-th layer, respectively, and σ(·) is the activation function. After the transformation through the last layer, the output vector s j (L) is obtained, which is then mapped to a scalar score through a linear transformation, given by where w r is a score mapping weight vector, b r is a score mapping bias; predicted score reflects the degree of user's preference for item j, the higher the value, the more interested the user is in the item, in the training process of the network, the score prediction network will supervise the learning of the predicted score according to the pre-defined loss function, and the loss function is defined as: where r j is the true rating of item j by the user, N represents the number of training samples, and through the back propagation algorithm, the system continuously adjusts the parameters of each layer in the network to minimize the loss function, thereby improving the accuracy of rating prediction. 10.The user intention recognition driven personalized recommendation method of claim 1, wherein: The step S5 includes: Step S501: First, collect user interaction feedback through multiple data sources, including explicit feedback and implicit feedback, to form a feedback dataset: T = {(x i ,s i ,y i ,τ i ,c i )} i＝1 N , where x i represents the comprehensive input features of the i-th recommendation, s i is the recommendation score generated by the system, y i represents the actual feedback signal, τ i is the timestamp, c i represents the context information, and N represents the number of user feedbacks; Step S502: After collecting the feedback data, a hybrid error evaluation strategy is used to quantify the deviation between the recommended model prediction value and the actual feedback, and a feedback loss function is constructed, which is defined as: Wherein, y is the real user feedback value, usually refers to the user's behavior feedback, such as: click, rating or other forms of feedback, The recommended value predicted by the system, representing the predicted value of the recommended result given by the recommendation algorithm, and δ is a preset threshold value. The weighted loss function is obtained by weighting the feedback data: where w i is a coefficient for weighting each feedback data point to adjust its influence in the loss function according to the importance of different feedbacks, r i is the recommendation score generated by the system, i.e., the predicted value of the model for the matching degree of the recommendation item to the user, and N represents the number of user feedbacks. Step S503: On the basis of error evaluation, a parameter update strategy is used to realize online model adaptive adjustment, and the parameter update process is as follows: where λ is the model update parameter, η is the learning rate, β1 and β2 are the decay rates of the first and second moments, respectively, ∈ is a small constant to prevent the denominator from being zero, m t is the first moment estimate, i.e., the exponentially weighted moving average of the gradient, which is used to capture the overall direction of the gradient, v t is the second moment estimate, i.e., the exponentially weighted moving average of the squared gradient, which is used to measure the variance of the gradient, is the bias correction value for the first moment estimate, is the bias correction value for the second moment estimate, is the gradient of the loss function L with respect to λ, which represents the directional derivative of the model parameters in the current iteration, and is used to guide the parameter update; Step S504: In order to further improve the response ability of the recommendation system to the feedback signal, the system dynamically updates the probability distribution of each recommendation candidate according to the current feedback reward distribution in each recommendation decision, that is, in each recommendation, the system samples the expected reward and selects the optimal recommendation accordingly, so as to realize real-time online learning and adaptive adjustment.

Images