Multi-behavior sequence recommendation system and method that captures user multifaceted preferences and intents

Abstract

The application belongs to the technical field of artificial intelligence, and discloses a multi-behavior sequence recommendation system and method for capturing user multi-aspect preferences and intentions, which comprises a preprocessing module, a multi-aspect preference modeling module, a multi-aspect intention modeling module, a preference and intention fusion module and a probability output module. The preprocessing module pre-processes multi-type behavior data; the multi-aspect preference modeling module models user multi-aspect preferences through a multi-aspect projection mechanism; the multi-aspect intention modeling module models user multi-aspect intentions through a behavior detail perception LSTM and a preference-guided multi-aspect attention mechanism; the preference and intention fusion module adaptively integrates user preference and intention information through a multi-aspect preference and intention fusion mechanism; and the probability output module, which is connected to the preference and intention fusion module, predicts the next item and category that the user is interested in. The system can capture user multi-aspect preferences and intentions and obtain accurate prediction results.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to a multi-behavioral sequence recommendation system and method for capturing various user preferences and intentions. Background Technology

[0002] In today's rapidly growing information society, recommender systems that attempt to predict items users might be interested in have become an important means of helping users overcome information overload. Because user preferences can change over time, a significant number of researchers have devoted themselves to sequence recommendation. Sequence recommendation can model sequence-related patterns (such as sequential patterns, co-occurrence patterns, and distance patterns) in user-item interaction data.

[0003] Due to the increasing sophistication of user data descriptions, the addition of new features around users, and the integration of multiple data sources, heterogeneous behavioral data encompassing various behavioral types is becoming increasingly common. For example, an e-commerce website might collect clicks, add-to-cart actions, and purchase actions. However, most existing methods only consider a single type of behavior, namely, target-oriented behavior, because it is the strongest signal reflecting user preferences. In reality, other types of behavior should also be considered. One of the most important reasons is that supporting behaviors can also reflect user intent. Generally speaking, user preferences and user intent are two key factors influencing whether a user is interested in a project.

[0004] In fact, user preferences and intentions have multifaceted characteristics. For example, user preferences and intentions may be reflected in terms of category and brand.

[0005] To leverage multi-type behavior, researchers have proposed several methods. These methods can be categorized into two types. The first type primarily involves extended matrix factorization algorithms, extended Bayesian personalized ranking algorithms, or the introduction of multilayer perceptrons. This type of method models multi-type behavior data from a static perspective, ignoring sequence-related patterns in the behavior data. The second type of method considers the sequence characteristics of behavior. This type can be further divided into two subclasses. The first subclass treats all behaviors as a long sequence. Then, the model is designed to process this long sequence. This subclass fails to capture the inherent patterns of each behavior type. The second subclass models at least two sequences. This subclass alleviates the limitations of the first subclass. However, this subclass fails to consider that user preferences and intentions come from multiple sources. Furthermore, most methods ignore the fact that the behaviors supporting the type can be very noisy. If noise is not handled properly, it can impair the model's performance. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a multi-behavioral sequence recommendation system and method for capturing multiple user preferences and intentions. This invention enables the learning of multiple user intentions from noisy, multi-type behaviors and the learning of multiple user preferences from target behaviors, thereby obtaining accurate prediction results.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] First, this invention provides a multi-behavioral sequence recommendation system for capturing multiple user preferences and intentions, comprising: a preprocessing module, a multi-faceted preference modeling module, a multi-faceted intention modeling module, a preference and intention fusion module, and a probability output module.

[0009] The preprocessing module processes the raw behavioral data according to the input format supported by the system to obtain the target behavioral sequence. and a multi-type behavior sequence X = {x1, x2, ..., x N},

[0010] The multi-faceted preference modeling module is based on the target behavior sequence. Generate user preference representations across multiple aspects. The multi-aspect preference modeling module includes a shared vector embedding layer, LSTM, and a multi-aspect projection mechanism. LSTM models the target behavior sequence. The hidden state representation is projected onto multiple spaces, i.e., multiple aspects, using a multi-faceted projection mechanism to generate representations with multiple preferences. The multi-faceted intent modeling module generates user intent representations across multiple aspects based on multi-type behavior sequences X. This module includes a shared vector embedding layer, a behavior detail-aware LSTM, and multiple parallel preference-guided multi-faceted attention mechanisms. The behavior detail-aware LSTM models the multi-type behavior sequences X. Under each aspect, a preference-guided multi-faceted attention mechanism is used, and these multiple parallel preference-guided multi-faceted attention mechanisms generate representations of multiple intents. The preference and intent fusion module adaptively balances the importance of user preferences and intents in each aspect, utilizing a multi-faceted preference and intent fusion mechanism to... and The data is then merged to generate the final user representation h. F The probability output module will output the user's final representation h. F As input, output the probability of the user purchasing all items at the next time point. And the probability of purchasing all item categories.

[0011] Furthermore, the original behavioral data consists of multiple types of behaviors. The preprocessing module arranges these behaviors into a long sequence based on their timestamps, takes the N most recent behaviors, calculates the time interval between adjacent behaviors, and obtains a multi-type behavior sequence X. Each behavior in the multi-type behavior sequence X consists of an item number, an item category number, a timestamp, a behavior type number, and the time interval with the next behavior. Additionally, the target behavior is extracted from the multi-type behavior sequence X of length N to form a target behavior sequence.

[0012] Furthermore, the embedding layer of the multi-faceted preference modeling module and the multi-faceted intent modeling module is shared. Using the embedding layer, corresponding embedding vectors are assigned to items, item content, behavior type and time interval respectively.

[0013] Furthermore, the present invention also provides a multi-behavioral sequence recommendation method for capturing various user preferences and intentions, comprising the following steps:

[0014] Step 1: Construct the multi-behavior sequence recommendation system as described in claim 1, specifically as follows:

[0015] a) Data preprocessing: Arrange the various types of behaviors into a long sequence based on their timestamps. Take the N most recent behaviors and calculate the time interval between any two adjacent behaviors to obtain a multi-type behavior sequence X. Each behavior consists of the item number, item category number, timestamp, behavior type number, and the time interval before the next behavior. Furthermore, extract the target behavior from the multi-type behavior sequence X of length N to form the target behavior sequence.

[0016] b) Constructing a multi-faceted preference modeling module: The input to the multi-faceted preference modeling module is the target behavior sequence. The output consists of multiple preference representations.

[0017] c) Constructing a multi-faceted intent modeling module: The input of the multi-faceted intent modeling module is a multi-type behavior sequence X, and the output is multiple intent representations.

[0018] d) Construct a preference and intent fusion module and a probability output module: The input to the preference and intent fusion module is multiple preference representations. and multiple intent representations Finally, the probability output module outputs the probability of the user purchasing all items at the next time point. And the probability of purchasing all item categories.

[0019] Step 2, Training: Train the recommendation system built in Step 1 using the processed data;

[0020] Step 3: Use the recommendation system trained in Step 2 to recommend items of interest to the user: Take the user's historical behavior data as input, the recommendation system outputs the probability of the user purchasing all items at the next time point, and recommends the items with the highest purchase probability to the user, thus generating recommendation results.

[0021] Furthermore, step b) includes the following when constructing the multi-faceted preference modeling module:

[0022] b-1) Target behavior sequence As input, after passing through the embedding layer, we obtain embedding vectors for the item, item content, behavior type, and time interval. Then, we use the formula... 1≤m≤M yields intermediate results for calculating user preferences, where [,] denotes vector concatenation. It is the hidden state of the LSTM in the m-th step. This is the hidden state at the (m-1)th step of the LSTM, where M is the target action sequence. Length, p m It is the embedding vector of the item, q m b-2) This is the embedding vector of the item content; b-2) Multiple preference representations are generated using a multi-faceted projection mechanism. Specifically, the hidden state representation of the last step of the LSTM is used. Projected onto multiple latent semantic spaces, i.e., multiple aspects: 1≤j≤J, where It is the projection function responsible for projecting onto the j-th aspect. This represents the user preference representation for the j-th aspect, where J represents the number of aspects / spaces.

[0023] Furthermore, step c) of constructing the multifaceted intent modeling module includes, in addition to constructing the embedding layer, the following:

[0024] c-1) The multi-type behavior sequence X = {x1, x2, ..., x...} N As input, the behavioral detail-aware LSTM takes behavioral details into account when modeling item sequences, specifically treating behavioral details as strong signals that control the input gate, forget gate, and output gate.

[0025] c-2) A preference-guided multi-faceted attention mechanism is constructed by using multiple parallel preference-guided multi-faceted attention mechanisms. Preferences are used as guides to refine user intent from multiple types of behavior sequences. Specifically, a preference-guided multi-faceted attention mechanism is used under each implicit aspect. The preference-guided multi-faceted attention mechanism takes the user preference projected onto that aspect as the guide, and reduces the influence of noisy items by assigning different attention scores to each item in the item sequence, thereby integrating the information of the entire sequence to generate the user intent under that aspect.

[0026] Furthermore, step d) includes the following two steps:

[0027] d-1) Adaptively fuse user preferences and intentions across different aspects using a multi-faceted preference and intention fusion mechanism. Then, concatenate the user representations from different aspects that have fused preferences and intentions, and apply projection again. This yields the final user representation h. F ;

[0028] d-2) Simultaneously predict the user's next item of interest and the next category of interest: Among them W O,I and W O,C The matrix is ​​a parametric matrix. and These represent the probability of a user purchasing all items and categories, respectively.

[0029] Compared with the prior art, the advantages of the present invention are as follows:

[0030] 1) The recommendation system proposed in this invention can extract and integrate multiple preferences and intentions for multi-behavioral sequence recommendation.

[0031] 2) This invention proposes a multi-faceted preference modeling approach, which can characterize user preferences from multiple aspects with fine granularity, thereby generating multiple preference representations. It also proposes a multi-faceted intent modeling approach, which considers behavioral details (including behavior type) and noise when modeling user intent. Simultaneously, it also characterizes user intent from multiple aspects with fine granularity, generating multiple intent representations. Finally, it proposes a fusion method that adaptively integrates multi-faceted preferences and intents, ultimately providing the final user representation.

[0032] 3) This invention is significantly superior to existing recommendation systems, and has objective recommendation accuracy. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0034] Figure 1 It is the method flowchart of the present invention;

[0035] Figure 2 It is the system structure schematic diagram of the present invention;

[0036] Figure 3 It is the schematic diagram of behavior detail perception LSTM. Specific embodiments

[0037] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0038] Embodiment 1

[0039] Combined with Figure 2 As shown, a multi-behavior sequence recommendation system that captures users' multi-faceted preferences and intentions includes: a preprocessing module, a multi-faceted preference modeling module, a multi-faceted intention modeling module, a preference and intention fusion module, and a probability output module.

[0040] The preprocessing module processes the original behavior data according to the input format supported by the system to obtain the target behavior sequence and the multi-type behavior sequence X = {x1, x2,..., x N}.

[0041] The original behavior data is multi-type behavior. The preprocessing module arranges the multi-type behavior into a long sequence according to the timestamp size, takes the N most recently occurred behaviors, calculates the time interval between adjacent two behaviors, and then discretizes the time interval by means of bucketing. In this way, each behavior component of the multi-type behavior sequence X is the item number, item category number, timestamp, behavior type number, and the time interval with the next behavior; in addition, based on the multi-type behavior sequence X with a length of N, the target behavior (purchase behavior in this embodiment) is extracted to form the target behavior sequence M represents the length of the target behavior sequence, M < N. Here, is used to represent a target behavior.

[0042] The multi-faceted preference modeling module generates the preference representation of the user in multiple aspects based on the target behavior sequence The multi-faceted preference modeling module includes a shared vector embedding layer, an LSTM, and a multi-faceted projection mechanism. The LSTM models the target behavior sequence The hidden state representation is projected onto multiple spaces, i.e., multiple aspects, using a multi-faceted projection mechanism to generate representations with multiple preferences.

[0043] The multi-faceted intent modeling module generates user intent representations across multiple aspects based on multi-type behavior sequences X. This module includes a shared vector embedding layer, a behavior detail-aware LSTM, and multiple parallel preference-guided multi-faceted attention mechanisms. The behavior detail-aware LSTM models the multi-type behavior sequences X. For each aspect, a preference-guided multi-faceted attention mechanism is used, and these multiple parallel preference-guided multi-faceted attention mechanisms generate representations of multiple intents.

[0044] The preference and intent fusion module adaptively balances the importance of user preferences and intents in each aspect, utilizing a multi-faceted preference and intent fusion mechanism to... and The data is then merged to generate the final user representation h. F .

[0045] The probability output module will output the user's final representation h. F As input, output the probability of the user purchasing all items at the next time point. And the probability of purchasing all item categories.

[0046] It should be noted here that the embedding layer of the multi-faceted preference modeling module and the multi-faceted intent modeling module is shared. Using the embedding layer, corresponding embedding vectors are assigned to items, item content, behavior type and time interval respectively.

[0047] Example 2

[0048] Combination Figure 1 As shown, this embodiment provides a multi-behavioral sequence recommendation method for capturing various user preferences and intentions, including the following steps:

[0049] Step 1: Construct a multi-behavior sequence recommendation system as described in Example 1, which captures various user preferences and intentions, as follows:

[0050] a) Data Preprocessing: The purpose of this step is to generate two sequences: a multi-type behavior sequence and a target behavior sequence. The target behavior sequence is input into the multi-faceted preference modeling module obtained in step b), and the multi-type behavior sequence is input into the multi-faceted intent modeling module obtained in step c).

[0051] The method is as follows: Arrange multi-type behaviors into a long sequence according to the timestamp size, and calculate the time interval between adjacent behaviors based on the timestamp. Then discretize the time interval by means of bucketing. In this way, each behavior x in the multi-type behavior sequence X n The constituent elements are the item number i n , the item category number c n , the timestamp t n , the behavior type number b n , and the time interval with the next behavior bucketize(Δt n ); In addition, based on the multi-type behavior sequence X with length N, extract the target behavior (in this embodiment, the purchase behavior) to form the target behavior sequence M represents the length of the target behavior sequence, M < N. Here, is used to represent a target behavior.

[0052] b) Construct a multi-faceted preference modeling module: The input of the multi-faceted preference modeling module is the target behavior sequence The output is multiple preference representations

[0053] First, use a long short-term memory network (LSTM) to model the target behavior sequence Furthermore, in order to capture the multi-faceted characteristics of preferences in a fine-grained manner, project the hidden state representation of the LSTM into multiple spaces, that is, multiple aspects, using the proposed multi-faceted projection mechanism, which will generate multiple preference representations where J represents the number of aspects / spaces.

[0054] c) Construct a multi-faceted intention modeling module: The input of the multi-faceted intention modeling module is the multi-type behavior sequence X, and the output is multiple intention representations

[0055] Information such as the behavior type and the time interval between behaviors describes the details of a user's interaction with an item. This kind of information is called behavior detail information. In order to capture the impact of behavior detail information on the item sequence modeling, use the proposed behavior detail-aware LSTM to model the multi-type behavior sequence X = {x1, x2,..., x N}. Then, in order to reduce the noise in the multi-type behavior sequence, use the proposed preference-guided multi-faceted attention mechanism on top of the detail-aware LSTM. Furthermore, in order to capture the multi-faceted characteristics of intentions in a fine-grained manner, under each aspect, use a preference-guided multi-faceted attention mechanism, which will generate multiple representations of intentions The parallel multiple preference-guided multi-faceted attention mechanisms are called preference-guided multi-faceted attention mechanisms.

[0056] d) Construct a preference and intent fusion module and a probability output module: The input to the preference and intent fusion module is multiple preference representations. and multiple intent representations Finally, the probability output module outputs the probability of the user purchasing all items at the next time point. And the probability of purchasing all item categories.

[0057] To adaptively integrate multiple user preferences and intentions, a proposed multi-faceted preference and intention fusion mechanism is used, which will... as well as The fusion ultimately yields the user's final representation h. F The probability output module then outputs the user's final representation h. F As input, output the probability of the user purchasing all items at the next time point. And the probability of purchasing all item categories. At this point, a complete recommendation model has been built, and the model can be trained using data to learn appropriate model parameters.

[0058] Step 2, Training: Train the recommendation system built in Step 1 using the processed data.

[0059] Step 3: Use the recommendation system trained in Step 2 to recommend items of interest to the user: Take the user's historical behavior data as input, the recommendation system outputs the probability of the user purchasing all items at the next time point, and recommends the items with the highest purchase probability to the user, thus generating recommendation results.

[0060] To augment the training data, this embodiment predicts each purchase action (predicting items and categories): The trained recommendation system recommends items that users are interested in: Once the recommendation system is trained, it can be fed with the user's historical behavior data, and the system will output the probability of the user purchasing all items at the next point in time. Items with a high probability of being recommended will be presented to the user.

[0061] Example 3

[0062] Steps b) and c) both involve the design of the embedding layer. In this embodiment, the embedding layer is designed to be shared by the multi-faceted preference modeling module and the multi-faceted intent modeling module. Using the embedding layer, item i... n Item content c n Behavioral type b n and time interval bucketize(Δt) n Assign the corresponding embedding vectors, i.e., the original vector representations: p n=W I i n q n =W C c n r n =W B b n s n =W Δ bucketize(Δt n ), where W I W C W B and m Δ They are all parameter matrices.

[0063] Example 4

[0064] Step b) In constructing the multi-faceted preference modeling module, in addition to building the embedding layer, it also includes: b-1) dividing the target behavior sequence As input, after passing through the embedding layer, we obtain item i. n Item content c n Behavioral type b n and time interval bucketize(Δt) n The embedding vector of ) is then obtained through the formula 1≤m≤M yields intermediate results for calculating user preferences, where [,] denotes vector concatenation. It is the hidden state of the LSTM in the m-th step. This is the hidden state at the (m-1)th step of the LSTM, where M is the target action sequence. Length, p m It is the embedding vector of the item, q m This is the embedding vector of the item content. b-2) To model user preferences in a fine-grained manner, this embodiment utilizes a multi-faceted projection mechanism to generate representations of multiple preferences. Specifically, the hidden state representation of the last step (i.e., the Mth step) of the LSTM is... Projected onto multiple latent semantic spaces, i.e., multiple aspects: 1≤j≤J, where It is the projection function responsible for projecting onto the j-th aspect. This represents the user preference representation for the j-th aspect, where J represents the number of aspects / spaces. The projection function can be linear or non-linear. For simplicity, this embodiment uses a linear projection function, i.e. 1≤j≤J, where All matrices are parametric matrices. The rationale for the multi-faceted projection mechanism is easily proven: numerous studies have demonstrated that different filters within a convolutional layer focus on features of different aspects of the image; the proposed projection function is analogous to a filter. Each projection function focuses on automatically extracting features of a specific aspect.

[0065] The other steps in this embodiment are the same as those in the multi-behavioral sequence recommendation method for capturing multiple user preferences and intentions described in Embodiment 2, and will not be repeated here.

[0066] Example 5

[0067] Step c) When constructing the multi-faceted intent modeling module, in addition to constructing the embedding layer, it also includes:

[0068] c-1) The multi-type behavior sequence X = {x1, x2, ..., x...} N As input, behavior-detail-aware LSTM considers behavioral details when modeling item sequences. Specifically, it treats behavioral details as strong signals controlling the input gate, forget gate, and output gate. This improved LSTM is called behavior-detail-aware LSTM. Figure 3 As shown. The formal definition is:

[0069]

[0070]

[0071]

[0072]

[0073]

[0074] Among them, W i W f , and W o They are all parameter matrices. and Both are bias vectors; σ(·) is the element-wise sigmoid function, and ⊙ is element-wise multiplication. f n and o n These are the input gate, the forget gate, and the output gate. and h n These are the internal and external states at step n, respectively.

[0075] (c-2) Multi-type behavioral data may contain a lot of noise. To reduce noise and generate more accurate user intent, a preference-guided multi-faceted attention mechanism is constructed by using multiple parallel preference-guided multi-faceted attention mechanisms. Preferences act as guides to refine user intent from multi-type behavioral sequences. Specifically, under each implicit aspect, a preference-guided multi-faceted attention mechanism is used. This mechanism uses the user preference projected onto that aspect as a guide (query vector), and reduces the influence of noisy items by assigning different attention scores to each item in the item sequence, thereby integrating the information of the entire sequence to generate the user intent for that aspect.

[0076] This fine-grained modeling approach enables the learning of J intent representations from J aspects. The formal definition of a preference-guided multi-faceted attention mechanism is given below:

[0077]

[0078]

[0079]

[0080] Where 1≤j≤J, and Both are parameter matrices, v j It is a weight vector. It is a bias vector. It is calculated from step b-2). It is calculated from step c-1). α j,n It is the weight assigned to the nth item under the j-th aspect. This represents the corresponding user intent in the j-th aspect. It is the scoring function in the attention mechanism.

[0081] The other steps in this embodiment are the same as those in the multi-behavioral sequence recommendation method for capturing various user preferences and intentions described in Embodiment 4, and will not be repeated here.

[0082] Example 6

[0083] Step d) includes the following two steps:

[0084] d-1) Adaptively fuse user preferences and intentions across different aspects using a multi-faceted preference and intention fusion mechanism. Then, concatenate the user representations from different aspects that have fused preferences and intentions, and apply projection again. This yields the final user representation h. F .

[0085] The details are as follows: In order to adaptively integrate user preferences and intentions under each aspect, the multi-aspect preference and intention fusion mechanism in this embodiment is implemented through the following formula: 1≤j≤J. Where... For the weight vector, For bias. σ(·) is the element-wise sigmoid function, β j It is the weight assigned to the user's intent under the j-th aspect. This represents the corresponding user preference for the j-th aspect. This represents the user intent for the j-th aspect. Let h be the user representation that integrates preferences and intentions for the j-th aspect. Then, the user representations that integrate preferences and intentions for different aspects (i.e., space) are concatenated and projected again to obtain the final user representation h. F As shown in the formula below: Among them W ρ The matrix is ​​a parametric matrix, h F and The dimensions are equal.

[0086] d-2) Simultaneously predict the user's next item of interest and the next category of interest: Among them W O,I and W O,C The matrix is ​​a parametric matrix. and These represent the probability of a user purchasing all items and categories, respectively.

[0087] The other steps in this embodiment are the same as those in the multi-behavioral sequence recommendation method for capturing various user preferences and intentions described in Embodiment 5, and will not be repeated here.

[0088] At this point, a complete recommender system model has been built, and it can be trained using data to learn appropriate model parameters. To augment the training data, this chapter predicts each purchase action (predicting items and categories):

[0089] Therefore, the loss function during training is:

[0090]

[0091] Where Θ represents all trainable parameters, γ is a hyperparameter controlling the influence of class information, and λ is a hyperparameter controlling the strength of L2 regularization used to prevent overfitting. U is the set of all users in the training set. Let be the probability of a user purchasing all items at step m+1. Let be the probability of a user purchasing all categories at step m+1.

[0092] The technical effectiveness of this invention is verified through specific experimental data. The experimental data are taken from two real e-commerce datasets, Taobao and Retailrocket, both of which contain multiple types of behavior (Taobao contains four types of behavior; Retailrocket contains three). These two datasets come from the Taobao App and a retail website, respectively. Following the leave-one-out method widely used in existing research, each user's most recently purchased items are used for testing, the second most recently purchased items for validation, and the remaining purchased items for training. For fair and efficient evaluation, each user's most recently purchased items are considered positive samples, and 100 items that the user has never interacted with are randomly selected as negative samples to form the test set. To evaluate the performance of each model, two widely used metrics are used: hit ratio (HR) and normalized discounted cumulative gain (NDCG). The experimental results are shown in Table 1. The multi-behavior sequence recommendation system and method proposed in this invention, which can capture multiple user preferences and intentions, outperforms existing state-of-the-art models and methods.

[0093] Table 1

[0094]

[0095] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should be protected by the present invention.

Claims

1. A multi-behavioral sequence recommendation system that captures various user preferences and intentions, characterized in that, include: It includes a preprocessing module, a multi-faceted preference modeling module, a multi-faceted intent modeling module, a preference and intent fusion module, and a probability output module. The preprocessing module processes the raw behavioral data according to the input format supported by the system to obtain the target behavioral sequence. and multi-type behavior sequences Specifically: the original behavioral data consists of multiple types of behaviors. The preprocessing module arranges these behaviors into a long sequence based on their timestamps, and then selects the most recently occurring behavior. For each action, the time interval between two adjacent actions is calculated to obtain a multi-type action sequence. Multi-type behavior sequences Each action consists of an item number, an item category number, a timestamp, an action type number, and a time interval before the next action; furthermore, it is based on a multi-type action sequence of length N. Extract target behaviors and form a sequence of target behaviors. ; The multi-faceted preference modeling module is based on the target behavior sequence. The system generates user preference representations across multiple aspects. The multi-aspect preference modeling module includes a shared vector embedding layer, LSTM, and a multi-aspect projection mechanism. LSTM models the target behavior sequence. The hidden state representation is projected onto multiple spaces, i.e., multiple aspects, using a multi-faceted projection mechanism to generate representations with multiple preferences. ; The multi-faceted intent modeling module is based on multi-type behavior sequences. The system generates user intent representations across multiple aspects. The multi-faceted intent modeling module includes a shared vector embedding layer, a behavior detail-aware LSTM, and multiple parallel preference-guided multi-faceted attention mechanisms. The behavior detail-aware LSTM models multiple types of behavior sequences. In each aspect, a preference-guided multi-faceted attention mechanism is used, and multiple parallel preference-guided multi-faceted attention mechanisms generate representations of multiple intentions. The embedding layer of the multi-faceted preference modeling module and the multi-faceted intent modeling module is shared. Using the embedding layer, corresponding embedding vectors are assigned to items, item content, behavior type and time interval respectively. The preference and intent fusion module adaptively balances the importance of user preferences and intents in each aspect, utilizing a multi-faceted preference and intent fusion mechanism to... and The data is then integrated to generate the final user representation. ; The probability output module will output the user's final representation. As input, output the probability of the user purchasing all items at the next time point. And the probability of purchasing all item categories. .

2. A multi-behavioral sequence recommendation method for capturing various user preferences and intentions, characterized in that, Includes the following steps: Step 1: Construct the multi-behavior sequence recommendation system as described in claim 1, specifically as follows: a) Data preprocessing: Sort the various types of behaviors into a long sequence according to their timestamps, and take the most recently occurred one. For each action, the time interval between two adjacent actions is calculated to obtain a multi-type action sequence. Thus, each action consists of an item number, an item category number, a timestamp, an action type number, and a time interval before the next action; furthermore, it is based on a multi-type action sequence of length N. Extract target behaviors and form a sequence of target behaviors. ; b) Constructing a multi-faceted preference modeling module: The input to the multi-faceted preference modeling module is the target behavior sequence. The output consists of multiple preference representations. ; c) Constructing a multi-faceted intent modeling module: The input to the multi-faceted intent modeling module is a sequence of multiple types of behaviors. The output consists of multiple intent representations. ; d) Construct a preference and intent fusion module and a probability output module: The input to the preference and intent fusion module is multiple preference representations. and multiple intent representations Finally, the probability output module outputs the probability of the user purchasing all items at the next time point. And the probability of purchasing all item categories. ; Step 2, Training: Train the recommendation system built in Step 1 using the processed data; Step 3: Use the recommendation system trained in Step 2 to recommend items of interest to the user: Take the user's historical behavior data as input, the recommendation system outputs the probability of the user purchasing all items at the next time point, and recommends the items with the highest purchase probability to the user, thus generating recommendation results.

3. The multi-behavioral sequence recommendation method for capturing multiple user preferences and intentions according to claim 2, characterized in that, Step b) includes the following when constructing the multi-faceted preference modeling module: b-1) The target behavior sequence As input, after passing through the embedding layer, we obtain embedding vectors for the item, item content, behavior type, and time interval. Then, we use the formula... The intermediate results of calculating user preferences are obtained, where This indicates the concatenation of vectors. It is the LSTM number The hidden state of the step, It is the LSTM number -1 step of hidden state, M is the target behavior sequence length, It is the embedding vector of the item. It is the embedding vector of the item's content; b-2) Utilize a multi-faceted projection mechanism to generate multiple preference representations. Specifically, this involves representing the hidden state of the last step in the LSTM process. Projected onto multiple latent semantic spaces, i.e., multiple aspects: ,in It is responsible for projecting to the first Projection functions in various aspects It is in the The corresponding user preferences are represented in each aspect. The number of aspects / spaces represented.

4. The multi-behavioral sequence recommendation method for capturing multiple user preferences and intentions according to claim 2, characterized in that, Step c) When constructing the multi-faceted intent modeling module, in addition to constructing the embedding layer, it also includes: c-1) Multi-type behavior sequences As input, behavioral detail-aware LSTM takes behavioral details into account when modeling item sequences, specifically treating behavioral details as strong signals that control the input gate, forget gate, and output gate; c-2) A preference-guided multi-faceted attention mechanism is constructed by using multiple parallel preference-guided multi-faceted attention mechanisms. Preferences are used as guides to refine user intent from multiple types of behavior sequences. Specifically, a preference-guided multi-faceted attention mechanism is used under each implicit aspect. The preference-guided multi-faceted attention mechanism takes the user preference projected onto that aspect as the guide, and reduces the influence of noisy items by assigning different attention scores to each item in the item sequence, thereby integrating the information of the entire sequence to generate the user intent under that aspect.

5. The multi-behavioral sequence recommendation method for capturing multiple user preferences and intentions according to claim 2, characterized in that, Step d) includes the following two steps: d-1) Adaptively fuse user preferences and intentions across each aspect using a multi-faceted preference and intention fusion mechanism. Then, concatenate the user representations from different aspects that have fused preferences and intentions, and apply projection again to obtain the final user representation. ; d-2) Simultaneously predict the user's next item of interest and the next category of interest: ,in and The matrix is ​​a parametric matrix. and These represent the probability of a user purchasing all items and categories, respectively.

Images