A heterogeneous multi-view enhanced subsequence unit learning conversation recommendation method

Abstract

The application discloses a kind of heterogeneous multi-view enhancement subsequence unit learning session recommendation method, specifically related to the technical field based on session recommendation, solve the interaction between single item in the prior art, lack rich global context information, it is difficult to understand the intention of user from higher dimensional angle technical problem;Its technical scheme is: multiple continuous items are regarded as a subsequence unit, and learning is carried out on the subsequence level;It explores the intention of user within a certain range by subsequence, but not just focus on the direct relationship between items;The number of items in subsequence can be dynamically adjusted, so as to explore the influence of subsequence of different lengths on recommendation performance;The application can apply subsequence unit learning user intention in multiple sessions, better capture the context information in global session.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of session-based recommendation, and in particular to a session recommendation method based on heterogeneous multi-view enhanced sub-sequence unit learning. BACKGROUND

[0002] With the rapid development of the Internet and the explosive growth of digital information, a large amount of data has flooded into people's lives. In order to help users find and obtain the content they want from these huge amounts of information, recommendation systems have emerged. These systems build user models by analyzing users' historical behavior and provide personalized recommendation services. Session-based recommendation (SBR) is a new recommendation paradigm that differs from sequence recommendation. Its goal is to capture users' short-term dynamic preferences by analyzing the dependency relationships between items in a session and accurately predict the next most likely item to interact with.

[0003] Graph neural network (GNN) based methods model session sequences as graphs, showing the interaction between items, and learn user behavior through multi-layer GNNs. Graph neural networks can be applied to local sessions to consider the relationships between items within a session; they can also be applied to multiple global sessions to study the relationships between items in multiple sessions, thereby expanding the scope of item influence. However, most existing work typically adopts a single perspective approach when constructing graph structures, i.e., treating each item as an independent individual. These methods can introduce information bias because they only rely on the relationship between two items and cannot integrate users' intentions from a higher-dimensional perspective. The information of sessions of different lengths can reveal different user intentions, and if the user's interest can be grasped from a higher-dimensional global context, better recommendations can be provided. Although some methods also consider learning from sessions of different granularities, they only operate in local sessions and cannot obtain rich information from global context.

[0004] Therefore, how to apply sub-sequence unit learning to multiple sessions to better capture the context information in global sessions is a technical problem to be solved by the present application. SUMMARY

[0005] Therefore, the present application solves the technical problem in the prior art that the interaction between single items lacks rich global context information and it is difficult to understand the user's intention from a higher-dimensional perspective. The session recommendation method based on heterogeneous multi-view enhanced sub-sequence unit learning provided by the present application regards multiple consecutive items as a sub-sequence unit and learns at the sub-sequence level. The number of items in the sub-sequence can be dynamically adjusted to explore the impact of sub-sequences of different lengths on recommendation performance. Sub-sequence unit learning can be applied to multiple sessions to better capture the context information in global sessions.

[0006] The application provides a heterogeneous multi-view enhanced subsequence unit learning session recommendation method, which comprises four main stages. The first stage is the view construction stage, in which the session sequence is constructed into single view and heterogeneous multi-view. The second stage is the embedding learning stage, in which a hybrid reading function is used to learn the embedding between different scale views. Next, the fusion layer aggregates the information from the two views to generate a session-level embedding vector. Finally, in the result prediction stage, all items are scored by a sample adaptive loss function, and the item with the highest score is recommended.

[0007] Further, the global session is first modeled as single view and heterogeneous multi-view structure. The single view is constructed based on a single session s and is defined as G s =(X s ,E s ), where X s is the set of items interacted by the session sequence s, is the set of directed edges; each edge records the interaction between items, indicating that the anonymous user clicks the items x i and x j in order; each item has a self-loop, represented by a self-connection edge r self ; in addition, in order to represent the influence from other items, two kinds of relationships are defined: the forward information propagation r in and the reverse information propagation r out , r in represents the transmission of information from x i to x j , and r out represents the transmission of information from x j to x j under the same conditions. The connection strength in the single view is stored by the adjacency matrix A s , which records the interaction relationship between |X s | non-repeated items, and the interaction strength between two items is defined as:

[0008]

[0009] k1 represents the result after row normalization of the adjacency matrix A s .

[0010] The heterogeneous multi-view structure is used to represent the relationship between items at multiple granularities, and each node in the heterogeneous multi-view is composed of multiple consecutive items from the same batch of sessions. The heterogeneous multi-view where is the set of sub-sequences ; any one sub-sequence where l* denotes the number of consecutive items in the current sub-sequence; when l * = 1, the multi-view degenerates into a single view; is a set of edges, denotes the number of connections of a sub-sequence; denotes the directed edge from a sub-sequence to , both of which have length l * ; the adjacency matrix stores the interaction frequency between sub-sequences and :

[0011]

[0012] k2is the result of row normalization in the adjacency matrix .

[0013] Further, after the conversation sequence is constructed into a single view and a heterogeneous multi-view, a single view attention network (SV-GAT) and a heterogeneous multi-view neural network (HMV-GNN) are used to obtain the embedding of items and sub-sequences. The SV-GAT is based on the transition relationship between items in a single view, and aims to learn the correlation between individual items. To achieve this goal, any item x i in a single view is mapped to an embedding space, thereby obtaining the corresponding embedding vector. A graph attention network is used to capture the correlation e ij between items and other items:

[0014] e ij = LeakyRelu(W ij (x i ⊙x j ))

[0015] The influence degree between items is calculated by using a LeakyRelu activation function, and the element-level product of the vector sum is a learnable parameter.

[0016] Normalization is performed by a Softmax function:

[0017]

[0018] N xi is the 1st-order neighbor set of item x i , and a ij denotes the transition relationship between the item and its neighbors; the final feature representation of the item is obtained by weighted sum of the neighbor information:

[0019]

[0020] The set of all item-level embedding vectors in a single view is represented as X SV-GAT ; the item embedding vector is updated by a multi-layer attention network, as follows:

[0021]

[0022] l1 represents the number of attention layers of SV-GAT, A s represents the normalized value of the interaction relationship between items.

[0023] Further, unlike a single view, each node in a heterogeneous multi-view is composed of multiple consecutive items; for a sub-sequence of length l * , in order to learn the complete intent of the sub-sequence, a hybrid reading function R H is designed to generate the embedding vector m i of the heterogeneous sub-sequence, as follows:

[0024]

[0025] Specifically, both set-based reading functions (such as Mean and Max) and sequence-based reading functions (such as Gated Recurrent Unit, GRU) are considered. The former learns order-invariant intent, while the latter is better at capturing order-sensitive intent. For a sub-sequence of length l * , the final intent of the sub-sequence is calculated as follows:

[0026]

[0027] and respectively preserve the order-invariant intent and order-sensitive intent of the sub-sequence; the set of all sub-sequences is represented as M.

[0028] Further, HMV-GNN captures high-dimensional user intent information based on the transition relationship in the heterogeneous multi-view. To achieve this effect, a multi-head attention mechanism is used, which not only considers the influence of the sub-sequence, but also enhances the generalization ability of the present application. For a single head attention, the calculation formula is as follows:

[0029]

[0030] head i = Attention(MW i Q , MW i K , MW i V );

[0031] M represents the input of the sub-sequence, W​i Q , W i K and W i V are the learnable parameters of the i-th head, and the operation of the multi-head attention mechanism is to concatenate the results of n single-head attention networks and normalize them through the softmax function; form the latent vector representation of the session sequence:

[0032] M HMV-GNN = Multi-Attention (M; W att ) = [head1, …, head n ]W att

[0033] W att denotes the learnable parameters of the multi-head attention network;

[0034] After passing through the information to the l2 layer, the final embedding vector is represented as:

[0035]

[0036] M and M HMV-GNN represent the initial vectors and learned vectors of all sub-sequences in the heterogeneous multi-view, respectively.

[0037] Further, in order to avoid overfitting of the trained model, a Dropout operation is introduced when learning two levels of information:

[0038]

[0039] The item-level and sub-sequence-level embedding representations corresponding to the single-view attention network and the heterogeneous multi-view neural network are extracted from and respectively; for each session, these two types of embedding representations are denoted as X SV-GAT and M HMV-GAT ; in order to better utilize the information in the sub-sequence, a separate session embedding representation is generated for the sub-sequence level; given a session s and its corresponding sub-sequence unit where l * denotes the number of consecutive items in the sub-sequence, and the last sub-sequence is selected as the local representation; in order to capture rich comprehensive information, the influence weights of the local representation and the context are calculated, and the attention score is calculated using a soft attention mechanism, thereby generating the global representation of the sub-sequence

[0040]

[0041] and is a learnable weight matrix, is a bias vector, the embedding representation of the heterogeneous multi-view content generated by information concatenation The formula is as follows:

[0042]

[0043] ||represents a concatenation operation, is a learnable parameter; the sub-sequence level embedding contains the information of the global view and the local view of the sub-sequence;

[0044] The last item clicked by the user is often highly consistent with his current preference, and is significantly helpful to the recommendation task. Therefore, by the item x l ∈X SV-GAT the final session-level embedding representation s f is generated.

[0045] μis a tunable parameter used to control the weight distribution between session information and last-click item information.

[0046] Further, in order to calculate the probability of all candidate items, the session embedding s f is interacted with the initial item vector x i , and the next item most likely to interact with the current session is selected by the softmax function

[0047] In addition, since the prediction difficulty of different sessions varies with samples, a sample adaptive loss function is proposed. Through sample adaptive loss function scoring all items, the loss function introduces a modulation factor SA i , and adds it to the cross-entropy loss function; according to the prediction deviation of the sample, the weight is distributed:

[0048]

[0049] L SA =-∑(2-2·SA i ) τ log(SA i )

[0050] y i represents the true label of the next click item, τis a temperature coefficient; the SA loss function distributes the weight according to the prediction deviation of the sample; when τdecreases, the modulation factor reduces the loss contribution of simple samples; on the contrary, when τincreases, more attention is paid to high-confidence samples and their contributions.

[0051] In the above technical solutions, the present application provides technical effects and advantages:

[0052] 1. The present application regards multiple continuous items as a subsequence unit and learns at the subsequence level. It explores the user's intention within a certain range through the subsequence, rather than just focusing on the direct relationship between items. In addition, the number of items in the subsequence can be dynamically adjusted to explore the impact of subsequence of different lengths on recommendation performance.

[0053] 2. The present application uses a hybrid reading function to fuse the intention of each item in the corresponding subsequence unit, considering both order-independent and order-sensitive intentions. In addition, the sample adaptive loss function distinguishes the weight of items in the session, enhancing the fitting ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0055] Figure 1 The overall framework diagram of the present application is a heterogeneous multi-view enhanced subsequence unit learning session recommendation method.

[0056] Figure 2 The flowchart of the present application is a heterogeneous multi-view enhanced subsequence unit learning session recommendation method.

[0057] Figure 3 The construction example diagram of the present application is a single view and a heterogeneous multi-view.

[0058] Figure 4 The recommendation performance diagram of the present application is a recommendation method in subsequence of different lengths. DETAILED DESCRIPTION

[0059] In order to make those skilled in the art better understand the technical solutions of the present application, the present application will be further described in detail below with reference to the drawings.

[0060] Example 1:

[0061] Please refer to Figure 1A heterogeneous multi-view enhanced sub-sequence unit learning session recommendation method is shown in the overall framework diagram. The technical scheme provided by the application comprises four main stages: the first stage is a view construction stage, and multiple session sequences are constructed into single views and heterogeneous multi-views. The second stage is an embedding learning stage, and a hybrid reading function is used to learn the embedding between different scale views. Next, the fusion layer aggregates the information from the two graph neural networks to generate a session-level embedding vector. Finally, in the result prediction stage, all items are scored by sample adaptive loss, and the item with the highest score is recommended.

[0062] A flowchart of a heterogeneous multi-view enhanced sub-sequence unit learning session recommendation method is shown in Figure 2 .

[0063] First, the global session is modeled as a single view and a heterogeneous multi-view structure, please refer to Figure 3 An example diagram of the single view and the heterogeneous multi-view of the application. The single view is constructed based on a single session s and is defined as G s =(X s ,E s ), wherein X s is a set of items interacted by the session sequence s, is a set of directed edges; each edge records the interaction between items, indicating that the anonymous user clicks the items x i and x j in order; in order to preserve the information of the items themselves, each item has a self-loop, which is represented by a self-connection edge r self ; in addition, in order to represent the influence from other items, two kinds of relationships are defined: the forward information propagation r in and the reverse information propagation r out , r in represents the transmission of information from x i to x j , and r out represents the transmission of information from x j to x j under the same conditions. The connection strength in the single view is stored by an adjacency matrix A s , which records the interaction relationship between |X s | non-repeated items, and the interaction strength between two items is defined as:

[0064]

[0065] k1 represents the result after row normalization of the adjacency matrix A s .

[0066] The heterogeneous multi-view structure is used to represent the relationships between items at multiple granularities. Unlike single views, each node in the heterogeneous multi-view is composed of multiple consecutive items from a sub-sequence in the same batch session; the heterogeneous multi-view wherein is a set of sub-sequences ; any one sub-sequence wherein l * represents the number of consecutive items in the current sub-sequence; when l * = 1, the multi-view degenerates into a single view; is a set of edges, represents the number of connections of the sub-sequence; represents a directed edge from the sub-sequence to , both of which have a length of l * ; the adjacency matrix stores the interaction frequency between sub-sequences and :

[0067]

[0068] k2 is the result of row normalization in the adjacency matrix .

[0069] After the session sequence is constructed into a single view and a heterogeneous multi-view, a single view attention network (SV-GAT) and a heterogeneous multi-view neural network (HMV-GNN) are used to obtain the embedding of items and sub-sequences. The SV-GAT is based on the transition relationship between items in the single view, and aims to learn the correlation between individual items. To achieve this goal, any item x i in the single view is mapped to an embedding space, thereby obtaining the corresponding embedding vector. A graph attention network is used to capture the degree of association between items and other items e ij :

[0070] e ij = LeakyRelu (W ij (x i ⊙x j ))

[0071] The degree of influence between items is calculated using the LeakyRelu activation function, and the element-level product of the vector sum is a learnable parameter.

[0072] Normalization is performed by the Softmax function:

[0073]

[0074] N xi is the item xi The first-order neighbor set, α ij This represents the transformation relationship between an item and its neighbors; a weighted sum of the neighbor information is performed to obtain the final feature representation of the item.

[0075]

[0076] The set of all item-level embedding vectors in a single view is represented as X. SV-GAT The item embedding vector is updated through a multi-layer attention network, using the following formula:

[0077]

[0078] l1 represents the number of attention layers in SV-GAT, A s Normalized value representing the interaction relationship between stored items.

[0079] Unlike single-view architecture, each node in heterogeneous multi-view architecture consists of multiple consecutive items; for a length of l * To learn the complete intent of a subsequence, a hybrid readout function R was designed. H To generate embedding vectors m of heterogeneous subsequences i The formula is as follows:

[0080]

[0081] Specifically, both set-based read functions (such as Mean and Max) and sequence-based read functions (such as Gated Recurrent Units, GRUs) are considered. The former learns order-invariant intentions, while the latter is better at capturing order-sensitive intentions. For a length of l... * subsequence of The final intention of this subsequence is calculated using the following formula:

[0082]

[0083] and The order-invariant and order-sensitive intentions of the subsequences are preserved respectively; the set of all subsequences is represented by M.

[0084] HMV-GNN captures high-dimensional user intent information based on transition relationships in heterogeneous multi-view architecture. To achieve this, a multi-head attention mechanism is employed, which not only considers the influence of sub-sequences but also enhances the generalization ability of this invention. The calculation formula for the attention of a single head is as follows:

[0085]

[0086] head i= Attention(MW i Q ,MW i K ,MW i V );

[0087] M represents the input of the subsequence, W i Q , W i K and W i V is the learnable parameter of the i-th head, and the operation of the multi-head attention mechanism is to concatenate the results of n single-head attention networks and normalize them through the softmax function; the latent vector representation of the session sequence is formed as follows:

[0088] M HMV-GNN = Multi-Attention(M; W att ) = [head1,..., head n ] W att

[0089] W att represents the learnable parameter of the multi-head attention network;

[0090] After passing through the information to the l2 layer, the final embedding vector is represented as:

[0091]

[0092] M and M HMV-GNN represent the initial vector and the learned vector of all sub-sequences in the heterogeneous multi-view, respectively.

[0093] In order to avoid overfitting of the trained model, a Dropout operation is introduced when learning two levels of information:

[0094]

[0095] The single-view attention network and the heterogeneous multi-view neural network learn based on the same batch of sessions, and the corresponding item-level and sub-sequence-level embedding representations are extracted from and respectively; for each session, the two types of embedding representations are denoted as X SV-GAT and M HMV-GAT ; in order to better utilize the information in the sub-sequence, a separate session embedding representation is generated for the sub-sequence level; given a session s and its corresponding sub-sequence unit where l * represents the number of consecutive items in the sub-sequence, and the last sub-sequence As a local representation; to capture rich comprehensive information, compute local representation and context The influence weight, and the attention score is calculated by using soft attention mechanism Thus generating the global representation of subsequence

[0096]

[0097]

[0098] And The learnable weight matrix, Is the bias vector, and the embedding representation of heterogeneous multi-view content is generated by information splicing The formula is as follows:

[0099]

[0100] || Denotes the splicing operation, The learnable parameter; subsequence level embedding Contains the information of global view and local view of subsequence;

[0101] The last item clicked by the user is often highly consistent with his current preference, and is significantly helpful to the recommendation task. Therefore, by the last clicked item x l ∈X SV-GAT Generate the final session-level embedding representation s f :

[0102]

[0103] μ is a adjustable parameter, used to control the weight distribution between session information and last clicked item information.

[0104] In order to calculate the probability of all candidate items, the session embedding s f Interact with the initial item vector x i And through the softmax function, the next item most likely to interact with the current session is selected

[0105] In addition, since the prediction difficulty of different sessions varies with samples, a sample adaptive loss function is proposed. Through sample adaptive loss function to score all items, the loss function introduces modulation factor SA i And added to the cross entropy loss function; According to the prediction deviation of the sample, the weight is allocated:

[0106]

[0107] L SA = -∑(2 - 2 · SA i ) τ log(SA i )

[0108] y i denotes the true label of the next clicked item, and τ is the temperature coefficient; the SA loss function assigns weights according to the prediction bias of the sample; when τ decreases, the modulation factor reduces the loss contribution of simple samples; on the contrary, when τ increases, more attention is paid to high-confidence samples and their contributions.

[0109] Embodiment 2:

[0110] This embodiment selects three basic data sets for experiments: Tmall, Gowalla and Diginetica. These data sets are used to evaluate the recommendation performance of the present application in different fields and scenarios. The Tmall data set was originally published in the IJCAI-15 competition, recording the user's interaction information on the Tmall e-commerce platform. Gowalla is a real-world data set that includes user check-in behavior in different geographic locations, reflecting the spatial distribution of user activities and interests. The Diginetica data set is derived from CIKM Cup 2016, providing transaction records and purchase behavior of users on an online shopping platform.

[0111] In order to evaluate the performance of the present application, comparative experiments are carried out between the present application and the baseline model, and P@20 and MRR@20 are used as evaluation indicators. P@20 is used to measure the proportion of the user's actual interaction goods in the top 20 recommended goods. MRR@20 focuses on the ranking of the first relevant goods in the top 20 list, and calculates the average value of the reciprocal ranking. The larger the values of P@20 and MRR@20, the higher the recommendation quality.

[0112] This embodiment constructs a single view based on a single item and a heterogeneous multi-view based on multiple consecutive items. In the heterogeneous multi-view, the number of items for each node can be dynamically adjusted, so as to explore the influence of different sub-session sequence lengths on the recommendation performance. In all three data sets, the present application performs well, and the specific experimental results are shown in Table 1. Table 1 verifies that the overall performance of the present application is better than that of the baseline model, indicating that the single view and heterogeneous multi-view modeling method proposed by the present application can better capture user behavior, thereby generating accurate predictions that better meet user intentions.

[0113] Table 1: Comparison of overall performance of recommendation methods and baselines model

[0114]

[0115] Embodiment 3:

[0116] To understand the impact of sub-sequence length on the recommendation performance in heterogeneous multi-view structure, the embodiment experiments the sub-sequence length, and the value range is 1 to 7. Figure 4 The recommendation effects of different sub-sequence lengths on three data sets are shown. In the Tmall data set, the recommendation performance fluctuates with the change of the sub-sequence length, which shows that the sub-sequence length is closely related to the actual intention of the user. When the user performs the click behavior, the dynamic change of interest will affect the click result. It is observed that when the sub-sequence is 6, the P@20 and MRR@20 of the Tmall data set reach the peak value. For the Gowalla data set, the change of P@20 is not significant, but the recommendation performance of MRR@20 is the worst when the sub-sequence length is 5. In the Diginetica data set, when the sub-sequence length is 3, both P@20 and MRR@20 reach the best effect.

[0117] The above only describes some exemplary embodiments of the present application in a descriptive manner, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above drawings and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. A conversation recommendation method for heterogeneous multi-view enhanced subsequence unit learning, characterized in that, Includes the following steps, S1: In the view construction phase, the session sequence is constructed into a single view and heterogeneous multiple views; the single view is constructed through a single session sequence, and the heterogeneous multiple views are constructed through multiple session sequences. S2: Embedding learning phase, using a hybrid read function to learn the embeddings between views at different scales; S3: The fusion layer aggregates information from the two views to generate a session-level embedding vector; S4: In the result prediction stage, all items are scored using the sample adaptive loss function, and the item with the highest score is recommended; The embedding learning stage in step S2, after the session sequence is constructed into a single view and a heterogeneous multi-view, includes the single-view attention network and the heterogeneous multi-view neural network to obtain the embeddings of items and sub-sequences; The single-view attention network is based on the transition relationships between items in a single view and aims to learn the correlations between individual items; any item x in the single view i The data is mapped to the embedding space to obtain the corresponding embedding vector; a graph attention network is used to capture the correlation e between items and other items. ij : e ij =LeakyRelu(W ij (x i ⊙x j )); The influence between items is calculated using the LeakyReLU activation function, x. i ⊙x j Represents vector x i and x j element-wise product, W ij These are learnable parameters; Normalization is performed using the Softmax function: It is item x i The first-order neighbor set, α ij This represents the transformation relationship between an item and its neighbors; a weighted sum of the neighbor information is performed to obtain the final feature representation of the item. The set of all item-level embedding vectors in a single view is represented as X. SV-GAT The item embedding vector is updated through a multi-layer attention network, using the following formula: l1 represents the number of attention layers in SV-GAT, A s Normalized values ​​representing the interaction relationships between stored items; Each node in the heterogeneous multi-view neural network consists of multiple consecutive items; for a length of l * To learn the complete intent of a subsequence, a hybrid readout function R was designed. H To generate embedding vectors m of heterogeneous subsequences i The formula is as follows: For length l * subsequence of The final intention of this subsequence is calculated using the following formula: and The order-invariant and order-sensitive intentions of the subsequences are preserved respectively; the set of all subsequences is represented by M.

2. The conversation recommendation method based on heterogeneous multi-view enhanced subsequence unit learning according to claim 1, characterized in that, In step S1, a single view is defined as G. s =(X s E s ), where X s It is the set of items that have been interacted with in the session sequence s. It is a set of directed edges; each edge The interactions between the items were recorded, indicating that the anonymous user clicked on item x in sequence. i and x j Each project has a self-loop, connected by self-connecting edges r. self Representation; Definition of forward information propagation r in and reverse information propagation r out r in Indicates information from x i To x j The transmission of r, and r out This means that under the same conditions, from x j To x j Information transmission; connection strength in a single view is achieved through the adjacency matrix A. s The matrix stores |X s The interaction relationships between two unique items, and the interaction strength between the two items. Defined as: k1 indicates that in the adjacency matrix A s The result after row normalization.

3. The conversation recommendation method based on heterogeneous multi-view enhanced subsequence unit learning according to claim 2, characterized in that, In step S1, the heterogeneous multi-view structure is used to represent the relationships between items at multiple granularities. Each node in the heterogeneous multi-view consists of multiple consecutive items, which come from subsequences within the same batch of sessions. in It is a subsequence The set; any subsequence Among them l * Indicates the number of consecutive items in the current subsequence; when l * When = 1, multiple views degenerate into a single view; It is a set of edges. Indicates the number of connections in a subsequence; Indicates from subsequence arrive A directed edge, where both subsequences have length l. * Adjacency matrix Subsequences are stored and Frequency of interaction between them: k2 is in the adjacency matrix The result after row normalization.

4. The conversation recommendation method based on heterogeneous multi-view enhanced subsequence unit learning according to claim 1, characterized in that, The heterogeneous multi-view neural network captures high-dimensional user intent information based on the transition relationships in heterogeneous multi-views, and adopts a multi-head attention mechanism. The calculation formula for the attention of a single head is as follows: head i =Attention(MW i Q ,MW i K ,MW i V ); W i Q W i K and W i V These are the learnable parameters of the i-th head. The operation of the multi-head attention mechanism is to concatenate the results of n single-head attention networks and normalize them using the softmax function. The potential vector representation of the conversation sequence: M HMV-GNN =Multi-Attention(M;W att )=[head1,…,head n ]W att W att Represents the learnable parameters of a multi-head attention network; After the information is passed to the L2 layer, the final embedding vector is represented as: M and M HMV-GNN Let represent the initial vector and the learned vector of all subsequences in the heterogeneous multi-view, respectively.

5. The conversation recommendation method based on heterogeneous multi-view enhanced subsequence unit learning according to claim 4, characterized in that, In step S3, before information aggregation, a Dropout operation is introduced when learning the information at both the single-view and multi-view levels: The item-level and subsequence-level embedding representations corresponding to single-view attention networks and heterogeneous multi-view neural networks are respectively derived from... and Extract from; for each session, these two types of embedding representations are denoted as X. SV-GAT and M HMV-GAT Generate separate session embedding representations at the subsequence level; given a session s and its corresponding subsequence units. Among them l * This indicates the number of consecutive items in a subsequence, selecting the last subsequence. As a local representation; compute the local representation and context The influence weights are determined, and a soft attention mechanism is used to calculate the attention score. This generates a global representation of the subsequence. : and The weight matrix is ​​a learnable matrix. It is a bias vector that generates an embedded representation of heterogeneous multi-view content through information concatenation. The formula is as follows: || indicates a concatenation operation. Learnable parameters; subsequence-level embedding It contains information about the global and local views of the subsequences; The item x that the user last clicked in the single view. l ∈X SV-GAT Generate the final session-level embedding representation s f : μ is an adjustable parameter used to control the weight distribution between session information and last clicked item information.

6. The conversation recommendation method based on heterogeneous multi-view enhanced subsequence unit learning according to claim 5, characterized in that, In step S4, the session is embedded. f With the initial item vector x i The interaction is performed, and the softmax function is used to select the next item that is most likely to be interacted with in the current session. All items are scored using a sample adaptive loss function, which incorporates a modulation factor SA. i And add it to the cross-entropy loss function; assign weights based on the prediction bias of the samples: L SA =-∑(2-2·SA i ) τ log(IN i ) y i The true label of the next clicked item is represented by τ, which is the temperature coefficient. The SA loss function assigns weights based on the prediction bias of the samples. When τ decreases, the modulation factor reduces the loss contribution of simple samples. Conversely, when τ increases, more attention is paid to high-confidence samples and their contributions.

Images