Text representation alignment sequence recommendation method based on decoupling contrast learning

Abstract

The invention discloses a text representation alignment sequence recommendation method based on decoupling contrast learning, and belongs to the technical field of sequence recommendation and deep learning. The method comprises the following specific steps: constructing a text-based description data set for sequence recommendation; creating text semantic encoders of the articles and the sequences, and respectively generating corresponding text semantic vector representations; through unsupervised and supervised comparison of learning tasks, learning decoupling representation of sequences and target objects, and realizing bidirectional semantic alignment between the learning tasks and the target objects; according to the method, the overall semantics of the sequence and the features of the single object are effectively separated by decoupling the representation of the sequence and the target object, and the high-quality semantic representation of the interaction sequence and the target object is realized, so that the accuracy and effect of sequence recommendation are remarkably improved. According to the method, decoupling contrast learning is innovatively applied to sequence recommendation, the method can be widely applied to scenes such as e-commerce recommendation and content recommendation, and the method has important practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of sequence recommendation, and in particular relates to a text representation alignment sequence recommendation method based on decoupled contrastive learning. Background Art

[0002] The goal of sequential recommendation is to accurately predict the next item that users may be interested in by modeling their historical interaction sequences on online platforms, such as browsing and clicking. With the development of pre-trained language models, especially the successful application of BERT (Bidirectional Encoder Representation from Transformers, BERT) in machine translation and natural language processing tasks based on sequential text and recommendation, feasible solutions for processing sequential text data have emerged. These models have demonstrated impressive capabilities in understanding text semantics and extracting information. On this basis, the basic idea of ​​this study is to use pre-trained models such as BERT to encode the rich text information of items to extract user interest representations for sequences and improve recommendation performance. These models generate item text representations (e.g., titles) and learn item and sequence representations to directly optimize user interest representations by aligning sequence representations with textual semantic representations of target items. In this process, the textual representations of target items in the recommendation data are used as a supervisory signal. However, this approach has several limitations: (1) The unidirectional emphasis on semantic alignment with the next item causes the model to overfit to patterns in the training data related to the positive item, resulting in biased interest representations. (2) This approach ignores the benefits of bidirectional supervision in achieving comprehensive semantic representation, resulting in incomplete sequence representation and semantic matching bias. Summary of the invention

[0003] To solve the above technical problems, the purpose of the present invention is to provide a text representation alignment sequence recommendation method based on decoupled contrastive learning. The present invention proposes to decouple sequence and item representations and learn them independently to avoid information mixing. In addition, effectively aligning the sequence representation with the target item representation helps to understand more accurate user interest representations and ultimately improve the recommendation performance. Inspired by the self-supervised learning technology that uses raw data to construct self-supervisory signals and the effectiveness of bidirectional learning in sequence representation, we propose a hybrid contrastive learning strategy. This strategy combines unsupervised and supervised contrastive learning paradigms to decouple and enhance the ability of sequence representations that users are interested in. Our unsupervised contrastive learning promotes robust exploration of data features in recommendation tasks by decoupling sequence and item representations into two different subspaces. We use a bidirectional supervised contrast method to perform deep alignment between sequence representation and target item space.

[0004] To achieve the above object, the present invention proposes a text representation alignment sequence recommendation method based on decoupled contrastive learning, comprising the following steps:

[0005] Step 1: Connect multiple attribute description texts of items in sequence to construct sequence-item recommendation data based on text descriptions;

[0006] Step 2: Create a semantic encoder for the item description text based on the pre-trained language model, take the item description text as input, and output the text semantic vector representation corresponding to the description text as the item text representation;

[0007] Step 3: Create a sequence description text semantic encoder based on the pre-trained language model, take the sequence text description as input, and output the text semantic vector representation corresponding to its description text as the sequence text representation;

[0008] Step 4: Construct an unsupervised contrastive learning task based on the object text representation and sequence text representation obtained in step 2 and step 3, respectively, to learn the decoupled representation of the sequence and the target object;

[0009] Step 5: Based on the sequence text representation and item text representation obtained in steps 2 and 3, a supervised two-way contrastive learning task is constructed to achieve two-way semantic alignment between the sequence and its target item;

[0010] Step 6: Use a multi-task training strategy to jointly optimize the loss functions of the unsupervised and supervised contrastive learning training tasks in steps 4 and 5 to promote deep alignment of the sequence and the target item text semantic vector representation.

[0011] The beneficial effects of the present invention are as follows:

[0012] The text representation alignment sequence recommendation method based on decoupled contrastive learning proposed in the present invention aims to solve the problem of interest representation that the model overfits the patterns in the training data related to the positive items, resulting in semantic matching deviation and bias, by emphasizing unidirectional semantic alignment in the model training process. The method decouples the sequence and item representation spaces to independently learn their text semantic vector representations, and adopts a bidirectional supervised alignment method between the sequence text semantic vector representation and the target item text semantic vector representation, so as to achieve high-quality optimization of the sequence representation, and ultimately achieve the purpose of improving the recommendation performance, thereby solving the problems mentioned in the above background technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of the text representation alignment sequence recommendation method based on decoupled contrastive learning of the present invention. DETAILED DESCRIPTION

[0014] The following is a further clear and complete description of the technical solutions in the specific implementation of the present invention in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] The present invention provides the following technical solution: a method for recommending text representation alignment sequences based on decoupled contrastive learning. The specific steps are as follows: taking a shopping website recommendation scenario as an example, the method for recommending text representation alignment sequences based on decoupled contrastive learning is applied to a shopping recommendation scenario, such as Figure 1 As shown, the specific steps are as follows:

[0016] Step 1: Connect multiple attribute description texts of items in sequence to construct recommendation data based on text descriptions;

[0017] In the specific implementation, the sequence recommendation data based on text description is constructed according to the interactive item sequence generated by the user in the existing data set (represented by a digital identifier), the target item of the sequence (represented by a digital identifier), and the corresponding attribute description text of each item. The specific processing method is as follows:

[0018] Step 1-1. Use s = {v1, v2, ..., v n} represents the original sequence sample represented by a numeric identifier, where v i is an item that interacts with the user for the i-th time in chronological order, and n represents the total number of interactions with the user in the sequence sample. The interaction here refers to the unified description of clicks, purchases, etc. in different scenarios. Each item v i Each item v is associated with a unique identifier (represented by a number) and a descriptive text (such as the product description, title, or brand of the item). Assuming that the items in the sequence recommendation dataset contain k attribute types, this paper only focuses on the text descriptions of these items to learn the vector representation of the sequence. Specifically, for each item v in the sequence that the user interacts with, i , the present invention processes the description text into a "key-value pair" form: {(a1, d1), (a2, d2), ..., (a k , d k)}, where key a represents a property expressed in text (called property text, such as brand), and value d represents the text describing the property, i.e., property description text. For various types of property texts, [SEP] is used as a separator to concatenate various property texts and their property description texts. In this way, the item sequence expressed by the original digital identifier is processed and mapped to the corresponding text description {a1d1[SEP]a2d2[SEP]...[SEP]a k d k}.

[0019] Step 1-2. For the spliced ​​items i The attribute description text {a1d1[SEP]a2d2[SEP]...[SEP]a k d k Specifically, item v i The description text is a sequence of words, here we use T i It can be formalized as T i ={w i 1, w i 2, ..., w i j , ..., w i c}, where any word w i j It comes from a shared vocabulary V, and c is the truncated text length after the concatenation of various types of description texts. The items here refer to the clicked items and target items in the interaction sequence.

[0020] Step 1-3. For the sequence s = {v1, v2, ..., v n The attribute text and attribute description text of each item in} are spliced ​​in the order of interaction in the sequence to obtain the corresponding text-based sequence data. This constitutes a text-based sequence-target item data sample.

[0021] Step 2: Create a semantic encoder for the description text of the item (the clicked item and the target item in the interaction sequence) based on the pre-trained language model, take the text description of the item as input, and output the text semantic vector representation corresponding to the description text as the item text representation. Step 2 is specifically as follows:

[0022] Step 2-1. Build an item text semantic encoder based on the BERT model. i Add a special marker [CLS] before the text sequence corresponding to the clicked item in the sequence and the corresponding target item in the sequence.

[0023] Step 2-2. The sequentially concatenated [CLS] and multiple item attribute texts in the sequence are combined to obtain {[CLS]; T1, T2, ..., T n}. Where Ti = {w i 1, w i 2, ..., w i c} represents item v i All the words in the attribute text are input into the BERT model and the output vectors are averaged to generate the vector for the item. Represented in the real number field vector space The middle dimension is d W The parameter of BERT is θ b .

[0024] Step 2-3. For the target item v+ that the user actually clicks on corresponding to the sequence, its attribute text is T i+ ={w i 1+ , w i 2+ , ..., w i c+}, using the same parameters θ as in step 2-2 b The BERT model is used to encode the text semantic vector of the target item.

[0025] Step 3: Create a semantic encoder for the description text of the sequence based on the pre-trained language model, take the text description of the sequence as input, and output the text semantic vector representation corresponding to the description text as the sequence text representation. The specific steps of step 3 are:

[0026] Step 3-1. Based on the process of the semantic vector of the item text generated in step 2, the original representation s = {v1, v2, ..., v n}, the BERT model is used to generate a text semantic vector corresponding to each item in s, expressed as

[0027] Step 3-2. Build a Transformer-based sequence text semantic encoder to transform {x1, x2, ..., x n} as input, the output of Transformer at the last time step represents an initial sequence text semantic vector f = Transformer (θ e , [x1, x2, ..., x n ]) where θ e Represents the parameters of the Transformer-based sequence text semantic encoder;

[0028] Step 3-3. Based on step 3-2, a sequence-level residual connection block is introduced, that is, the vectors of each item in the sequence corresponding to the original representation before the input Transformer Perform average pooling on it to obtain f′=MeanPooling([x1, x2, ..., x n ]), f′ represents the sum of all item vectors [x1, x2, ..., x n ] is the vector obtained after the average pooling operation, which is a vector f′ with a dimension of d W Specifically, f′ is an average calculation of all item representations in the sequence to generate a vector that comprehensively represents the overall information of the sequence, which is used to maintain the consistency of the original input information and the retention of the overall features when the sequence-level residual connection block is introduced. Then f′ is added to the initial sequence text semantic vector f output by the Transformer to obtain h s =f+f′, The semantic vector representation of the sequence text is used as the input for the subsequent steps 4 and 5. This step is mainly used to solve the potential problems of gradient disappearance and original information loss caused by the multi-layer Transformer layer.

[0029] Step 4: Based on the target item text representation and sequence text representation obtained in step 2 and step 3, an unsupervised contrastive learning task is constructed to learn the decoupled representation of the sequence and the target item. Step 4 is specifically as follows:

[0030] Step 4-1. Sequence text semantic vector representation based on the generated sequence in step 3-3 Introduce sequence-level contrastive learning training tasks. By constructing positive sample pairs and negative sample pairs and designing sequence-level contrastive learning loss functions for model training. The process of constructing sequence-level positive sample pairs is as follows: Given a sequence sample, randomly delete items in the sequence to generate variants of the original sequence as positive samples of the sequence sample. The text semantic vector representation of the original sequence is used Indicates that the text semantic vector of the positive sample sequence is represented by The process of constructing sequence-level negative sample pairs is as follows: for a given sequence sample, other sequence samples in the batch during training are used as negative samples of the sequence sample, and their text semantic vectors are represented using Indicates. Where j represents the sample number in the batch. In the "sequence-sequence (SS)" contrastive learning training task, it is set that a batch contains B training instances during the training process. The goal of contrastive learning is to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs. The loss function is defined as:

[0031]

[0032] Among them, ex() represents an exponential function with base e; τ is an artificially defined hyperparameter used to adjust the smoothness of the similarity distribution between positive and negative sequence samples.

[0033] Step 4-2. Based on the target item text semantic vector representation generated in step 2-3, introduce the item-level "item-item" contrastive learning training task. Model training is performed by constructing positive sample pairs and negative sample pairs, and designing an item-level contrastive learning loss function. The process of constructing item-level positive sample pairs is as follows: given a target item sample, randomly delete the words in the target item sample to generate a variant of the original target item as the positive sample of the target item sample. The text semantic vector representation of the original target item is used Indicates that the text semantic vector of the positive sample target item is represented using The process of constructing item-level negative sample pairs is as follows: for a given target item sample, other target item samples in the batch during training are used as negative samples of the target item sample, and their text semantic vectors are represented using Indicates. Where j represents the sample number in the batch. In the "Item-Item (II)" contrastive learning training task, the goal of contrastive learning is to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs. The loss function is defined as:

[0034]

[0035] Among them, exp( ) represents an exponential function with base e; τ is a manually defined hyperparameter used to adjust the smoothness of the similarity distribution between positive and negative target item samples. The τ here is the same as the τ in step 4-1.

[0036] The purpose of this unsupervised contrastive learning method is to let the model learn to distinguish between similar items (target items and their positive samples) and different items (target items and negative samples). By constructing positive samples through random deletion operations, the robustness of the model to item representation can be enhanced. Using other items in the batch as negative samples can effectively utilize computing resources and provide diverse contrast samples.

[0037] Step 5: Based on the sequence text representation and the target item text representation obtained in steps 2 and 3, a supervised two-way contrastive learning task is constructed to achieve two-way semantic alignment between the sequence and its target item; the specific steps of step 5 are:

[0038] Step 5-1. When introducing the contrastive learning training tasks of "sequence-target item (SI)" and "target item-sequence (IS)", set a batch in the contrastive learning training process to contain B training instances, where each training instance contains a pair of samples: a sequence sample and its corresponding target item sample.

[0039] Step 5-2. Generate the text semantic vector representation of the training instance based on the sequence text semantic vector representation and the target item text semantic vector representation generated in step 3-3 and step 2-3: in, is the text semantic vector representation of the sequence obtained in step 3-3, It is the text semantic vector representation of the target item obtained by steps 2-3.

[0040] Step 5-3. Construct a “sequence-target item” contrastive learning training task by constructing positive and negative sample pairs. The process of constructing positive sample pairs is as follows: Given a sequence sample s j , and compare it with the corresponding target item sample v in the training instance j+ are combined together to form a positive sample pair, and its text semantic vector is expressed as The process of constructing negative sample pairs is as follows: for a given sequence sample, the other B-1 target item samples v in the training batch except the positive item samples are j′ Combined with it to form a negative sample pair, its text semantic vector is expressed as In the "Sequence-to-Item (SI)" contrastive learning training task, the goal of contrastive learning is to maximize the similarity of positive sample pairs while minimizing the similarity of negative sample pairs. The loss function is defined as:

[0041]

[0042] Among them, exp() represents an exponential function with e as the base; τ is a manually defined hyperparameter used to adjust the smoothness of the similarity distribution between positive and negative samples. The τ here is the same as the τ in steps 4-1 and 4-2.

[0043] Step 5-4. Construct the contrastive learning training task of "target item-sequence (IS)" by constructing positive and negative sample pairs. The process of constructing positive sample pairs is as follows: Given a target item sample v j+ , and compare it with the corresponding sequence sample s in the training instance j+are combined together to form a positive sample pair, and its text semantic vector is expressed as The process of constructing negative sample pairs is as follows: for a given target item sample, the other B-1 sequence samples s in the training batch except the positive sequence samples j′ Combined with it to form a negative sample pair, its text semantic vector is expressed as In the “target item-sequence (IS)” contrastive learning training task, the goal of contrastive learning is to maximize the similarity of positive sample pairs while minimizing the similarity of negative sample pairs. The loss function is defined as:

[0044]

[0045] Among them, exp( ) represents an exponential function with e as the base; τ is a manually defined hyperparameter used to adjust the smoothness of the similarity distribution between positive and negative samples. The τ here is the same as the τ in steps 4-1, 4-2, and 5-3.

[0046] Its core goal is to learn highly aligned representations of sequences and their target item positive samples in the semantic space, thus ensuring that the target item is ranked higher in the recommendation probability order.

[0047] Step 6: Use a multi-task training strategy to jointly optimize the loss functions of the unsupervised and supervised contrastive learning training tasks in steps 4 and 5 to promote the deep alignment of the sequence and the target item text semantic vector representation; finally, apply the trained text representation alignment sequence recommendation model based on decoupled contrastive learning to a specific recommendation scenario. Step 6 is specifically as follows:

[0048] Based on the loss functions of the unsupervised and supervised contrastive learning training tasks in steps 4 and 5, a weight is set for each loss function to balance their contributions, and their weighted sum is calculated as the total loss function of the entire text representation alignment sequence recommendation model based on decoupled contrastive learning, that is:

[0049]

[0050] Among them, λ i It is a manually defined hyperparameter, which represents the weight of each loss function, λ1+λ2+λ3+λ4=1.

[0051] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A text representation alignment sequence recommendation method based on decoupled contrastive learning, characterized in that: The following steps are involved: Step 1: Connect multiple attribute description texts of items in sequence to construct sequence-item recommendation data based on text descriptions; Step 2: Create a semantic encoder for the item description text based on the pre-trained language model, take the item description text as input, and output the text semantic vector representation corresponding to the description text as the item text representation; Step 3: Create a sequence description text semantic encoder based on the pre-trained language model, take the sequence text description as input, and output the text semantic vector representation corresponding to its description text as the sequence text representation; Step 4: Construct an unsupervised contrastive learning task based on the object text representation and sequence text representation obtained in step 2 and step 3, respectively, to learn the decoupled representation of the sequence and the target object; Step 5: Based on the sequence text representation and item text representation obtained in steps 2 and 3, a supervised two-way contrastive learning task is constructed to achieve two-way semantic alignment between the sequence and its target item; Step 6: Use a multi-task training strategy to jointly optimize the loss functions of the unsupervised and supervised contrastive learning training tasks in steps 4 and 5 to promote deep alignment of the sequence and the target item text semantic vector representation.

2. According to the text representation alignment sequence recommendation method based on decoupled contrastive learning in claim 1, it is characterized in that: The step 1 specifically includes the following sub-steps: Step 1-1: Use s = {v1, v2, ..., v n } represents the original sequence sample represented by a numeric identifier, where v i is an item that interacts with the user for the i-th time in chronological order, and n represents the total number of interactions with the user in the sequence sample; each item v i Each item v in the sequence that the user interacts with is associated with a unique identifier and a descriptive text; i , and process its description text into key-value pairs: {(a1, d1), (a2, d2), ..., (a k , d k )}, where key a represents a property expressed in text, and value d represents the text describing the property, i.e., the property description text; Use [SEP} as a separator to concatenate multiple attribute texts and their attribute description texts, process the item sequence expressed by the original digital identifier and map it to the corresponding text description {a1d1[SEP]a2d2[SEP]...[SEP]a k d k }; Step 1-2: The attribute description text of the spliced ​​item vi is {a1d1[SEP]a2d2[SEP]...[SEP]a k d k } is truncated; the description text Ti of item vi is a word sequence {ω i 1, ω i 2, ..., ω i c }, where any word w j From a shared vocabulary V, c is the truncated text length after the concatenation of various types of description texts; Step 1-3. For the sequence s = {v1, v2, ..., v n The attribute text and attribute description text of each item in} are spliced ​​according to the order of interaction in the sequence to obtain the corresponding text-based sequence data, which constitutes the sequence of text description-item recommendation data.

3. According to claim 2, a text representation alignment sequence recommendation method based on decoupled contrastive learning is characterized in that: The step 2 is specifically as follows: Step 2-1: Build an item text semantic encoder based on the BERT model. i Add a special marker [CLS] before the corresponding text sequence; Step 2-2: The sequentially concatenated [CLS] and multiple item attribute texts in the sequence are combined to obtain {[CLS]; T1, T2, ..., T n } Input into the pre-trained language model and take the average of the output vectors to generate the text semantic vector of the item Represents the real number field vector space The dimension is d W The vector of the pre-trained language model is θ b ; Step 2-3: For the target item v that the user actually clicks on corresponding to the sequence + , whose attribute text is T i+ ={w i 1+ , w i 2+ , ..., w i c+ }, using the same parameters θ as in step 2-2 b The pre-trained language model is used to encode the text semantic vector of the target item 4. The method for recommending text representation alignment sequences based on decoupled contrastive learning according to claim 3, characterized in that: The step 3 is specifically as follows: Step 3-1: Build a Transformer-based sequence text semantic encoder to transform the text semantic vector x i The sequence {x1, x2, ..., x n } as input, the output of the sequence text semantic encoder at the last time step represents an initial sequence text semantic vector f = Transformer (θ e , [x1, x2, ..., x n ]), where θ e Represents the parameters of the sequence text semantic encoder; Step 3-2: The vectors of each item in the sequence corresponding to the original representation before entering the Transformer Perform average pooling on it to obtain f′=MeanPooling([x1, x2, ..., x n ]), Then add f′ to the initial sequence text semantic vector f output by Transformer to get h s =f+f′, As a sequence text semantic vector.

5. The method for text representation alignment sequence recommendation based on decoupled contrastive learning according to claim 4, characterized in that: The step 4 is specifically as follows: Step 4-1: Based on the semantic vector representation of sequence text, introduce the sequence-level contrastive learning training task, construct positive sample pairs and negative sample pairs, and design the sequence-level contrastive learning loss function for model training; The process of constructing sequence-level positive sample pairs is as follows: given a sequence sample, randomly delete items in the sequence to generate variants of the original sequence as positive samples of the sequence sample. The text semantic vector representation of the original sequence is used Indicates that the text semantic vector of the positive sample sequence is represented by Indicates; where j represents the sample number in the batch; The process of constructing sequence-level negative sample pairs is as follows: for a given sequence sample, other sequence samples in the batch during training are used as negative samples of the sequence sample, and their text semantic vectors are represented using Indicates; in the sequence-level contrastive learning training task, the loss function is defined as: Among them, exp() represents an exponential function with base e; τ is a hyperparameter, and B represents the number of training instances in a batch; Step 4-2: Based on the text semantic vector representation of the target item, introduce the item-level contrastive learning training task, construct positive sample pairs and negative sample pairs, and design the item-level contrastive learning loss function for model training; The process of constructing item-level positive sample pairs is as follows: given a target item sample, randomly delete the words in the target item sample to generate a variant of the original target item as the positive sample of the target item sample. The text semantic vector representation of the original target item is used Indicates that the text semantic vector of the positive sample target item is represented using Representation; The process of constructing item-level negative sample pairs is as follows: for a given target item sample, other target item samples in the batch during training are used as negative samples of the target item sample, and their text semantic vectors are represented using express; In the item-level contrastive learning training task, the loss function is defined as:

6. The method for text representation alignment sequence recommendation based on decoupled contrastive learning according to claim 5, characterized in that: The step 5 is specifically as follows: Step 5-1: Introduce the contrastive learning training tasks of "sequence-target item" and "target item-sequence". Set a batch in the contrastive learning training process to contain B training instances, where each training instance contains a pair of samples: a sequence sample and its corresponding target item sample; Step 5-2: Generate the text semantic vector representation of the training instance based on the sequence text semantic vector representation and the target item text semantic vector representation generated in step 3-3 and step 2-3: Step 5-3: Construct positive and negative sample pairs to build a "sequence-target item" contrastive learning training task; The process of constructing positive sample pairs is as follows: given a sequence sample s j , and compare it with the corresponding target item sample v in the training instance j+ are combined together to form a positive sample pair, and its text semantic vector is expressed as The process of constructing negative sample pairs is as follows: for a given sequence sample, the other B-1 target item samples v in the training batch except the positive item samples are j′ Combined with it to form a negative sample pair, its text semantic vector is expressed as In the "sequence-target item" contrastive learning training task, the loss function is defined as: Step 5-4: Construct positive and negative sample pairs to build a "target item-sequence" contrastive learning training task; The process of constructing a positive sample pair is as follows: given a target item sample v j+ , and compare it with the corresponding sequence sample s in the training instance j+ are combined together to form a positive sample pair, and its text semantic vector is expressed as The process of constructing negative sample pairs is as follows: for a given target item sample, the other B-1 sequence samples s in the training batch except the positive sequence samples j′ Combined with it to form a negative sample pair, its text semantic vector is expressed as In the "target item-sequence" contrastive learning training task, the loss function is defined as:

7. The method for text representation alignment sequence recommendation based on decoupled contrastive learning according to claim 6, characterized in that: The total loss function during training is as follows: Among them, λ i It is a manually defined hyperparameter, representing the weight of each loss function, λ1+λ2+λ3+λ4=1.

Images