Behavior-enhanced semantic recommendation method and device for zero-shot cross-domain sequential recommendation

Abstract

The application discloses a behavior-enhanced semantic recommendation method and device for zero-sample cross-domain sequence recommendation, and belongs to the technical field of sequence recommendation. The method comprises the following steps: obtaining a target query in a target domain; determining a semantic candidate score corresponding to each candidate item by a semantic scoring model based on text semantic embeddings corresponding to historical interaction sequences, historical items and candidate items; determining a time feature corresponding to a historical item based on time information in the historical interaction sequence; determining a behavior candidate score corresponding to a candidate item by a behavior scoring model based on the time feature, interaction attributes corresponding to the historical item and the candidate item, and the text semantic embeddings corresponding to the candidate item; fusing the semantic candidate score and the behavior candidate score to obtain a fusion candidate score corresponding to the candidate item; and determining a recommendation result based on the fusion candidate score. The method can improve the accuracy of the recommendation result in a new field with insufficient in-domain interaction data.

Description

Technical Field

[0001] This invention belongs to the field of sequence recommendation technology, specifically relating to a behavior-enhanced semantic recommendation method and apparatus for zero-sample cross-domain sequence recommendation. Background Technology

[0002] Sequence recommendation models model users’ historical interactions in chronological order to infer their preferences over time and predict which items they might interact with next. They have been widely used in e-commerce, content distribution and information services.

[0003] Existing sequence recommendation methods typically rely on sufficient in-domain interaction data. They require learning from historical user-item interaction data to determine user behavior sequences and item relationships, thereby acquiring user behavioral preferences and item behavioral characteristics. Based on this learned information, candidate items are then rated and ranked. Therefore, the quantity and completeness of in-domain interaction data directly impact the sequence recommendation model's ability to learn user behavior patterns and item features.

[0004] However, when a new platform or product category is first launched, sufficient user interaction data is usually not yet accumulated. Furthermore, migrating recommendation models between different business domains may be limited by factors such as data privacy, data isolation, and operational strategies, making it difficult to obtain historical interaction data in the new domain or to use it for training and adjusting the recommendation model. Without domain-specific interaction data, existing sequence recommendation models struggle to directly learn the interaction patterns between users and items in the new domain, as well as the item characteristics reflected by these interactions, resulting in lower recommendation accuracy in the new domain. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that the methods commonly used in the prior art usually rely on sufficient intra-domain interaction data, which leads to low recommendation accuracy of the recommendation model in new domains. In order to solve the above problem, the present invention provides a behavior-enhanced semantic recommendation method and apparatus for zero-sample cross-domain sequence recommendation.

[0006] The content of this invention includes: In a first aspect, embodiments of the present invention provide a behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation, comprising: Obtain a target query in the target domain. The target query includes a historical interaction sequence and multiple candidate items. The historical interaction sequence includes multiple historical items and interaction times that the target user interacted with before the current predicted position. The target domain includes multiple target domain items, multiple target domain users, and target domain interaction data between the target domain items and the target domain users. The historical items and candidate items both belong to the target domain items, and the target user belongs to the target domain users. Based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item, a semantic candidate score is determined for each candidate item through a semantic scoring model. The semantic scoring model is pre-trained based on data from the source domain, which includes multiple source domain items, multiple source domain users, and source domain interaction data between the source domain items and the source domain users. Based on the time information in the historical interaction sequence, the time characteristics corresponding to the historical item are determined; Based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embeddings corresponding to the candidate items, a behavior candidate score is determined for each candidate item through a behavior scoring model. The interaction attributes include popularity-related proxies and co-occurrence information proxies. The interaction attribute mapper includes a popularity mapper and a co-occurrence mapper. The popularity mapper is used to determine the popularity-related proxies based on text semantic embeddings, and the co-occurrence mapper is used to determine the co-occurrence information proxies based on text semantic embeddings. The popularity mapper and the co-occurrence mapper are pre-trained based on data from the source domain. The popularity-related proxies are used to represent the proxy results of the relative popularity of items learned from the source domain in the target domain, and the co-occurrence information proxies are used to represent the proxy results of the collaborative association features between items learned from the source domain in the target domain. The semantic candidate score and the behavioral candidate score are fused to obtain a fused candidate score for each candidate item. The recommended result corresponding to the target query is determined based on the fusion candidate score.

[0007] Optionally, the time feature includes a time interval feature and a relative time position feature; Based on the time information in the historical interaction sequence, the time characteristics corresponding to the historical item are determined, including: Based on the time information in the historical interaction sequence, the time interval feature and relative time position feature corresponding to each historical item are determined respectively, wherein the time interval feature corresponding to the j-th historical item is... for: ; ; ; The relative time location feature of the j-th historical item for: ; in, Let j be the time interval corresponding to the j-th historical item. Let j be the interaction time corresponding to the j-th historical item. Let j-1 be the interaction time corresponding to the historical item. The quantity of the historical items, For a fixed time scale, and These are all preset constants used to prevent division by zero. Used to determine the maximum value among multiple values ​​within parentheses.

[0008] Optionally, the step of determining the semantic candidate score for each candidate item using a semantic scoring model based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item includes: The candidate item representation is obtained by mapping the text semantic embedding corresponding to the candidate item through the candidate item mapping function; The text semantic embedding corresponding to the historical item is mapped by the historical item mapping function to obtain the historical item representation corresponding to the historical item. Encode and read out multiple historical item representations arranged according to the historical interaction sequence to obtain the sequence representation corresponding to the historical interaction sequence; The semantic candidate score for each candidate item is determined based on the sequence representation and the representation of each candidate item.

[0009] Optionally, before determining the behavioral candidate score for each candidate item using a behavioral scoring model based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embedding corresponding to the candidate items, the method further includes: Based on the source domain interaction data, the popularity attribute and co-occurrence attribute of the source domain item are determined. The popularity attribute is used to characterize the relative popularity of the source domain item among the multiple source domain items. The co-occurrence attribute is used to characterize the association features between the source domain item and other source domain items based on user interaction behavior. The text semantic embedding corresponding to the source domain item is input into the popularity mapper to obtain the popularity-related proxy corresponding to the source domain item. The text semantic embedding corresponding to the source domain item is input into the co-occurrence mapper to obtain the co-occurrence information proxy corresponding to the source domain item. The parameters of the popularity mapper are adjusted based on the difference between the popularity-related agent and the popularity attribute to obtain a trained popularity mapper. Similarly, the parameters of the co-occurrence mapper are adjusted based on the difference between the co-occurrence information agent and the co-occurrence attribute to obtain a trained co-occurrence mapper.

[0010] Optionally, determining the popularity attribute corresponding to the source domain item based on the source domain interaction data includes: The source domain interaction data is divided into multiple time periods according to a preset time period, and the number of interactions of the source domain items in each time period is counted. Based on the number of interactions and the preset attenuation coefficient, the attenuation status of the source domain item in each time period is determined; For source domain items with non-zero decay state, the average parallel rank is assigned, and then normalized after being divided by the target value to obtain the normalized rank value of the source domain item in each time period. The target value is the maximum average rank plus one. The popularity attribute corresponding to the source domain item is determined based on the normalized rank value; Among them, source domain items Corresponding popularity attribute for: ; ; in, This indicates the last time period mentioned. Indicates the number of time periods at the end of the aggregation. Source Domain Items In the The normalized rank value corresponding to each time period Source Domain Items In the The decay state over a time period, Source Domain Items In the Number of interactions per time period The preset attenuation coefficient is denoted as .

[0011] Optionally, determining the co-occurrence attribute of the source domain item based on the source domain interaction data includes: Construct a source domain user-item association matrix and an initial co-occurrence matrix based on the source domain interaction data; After removing self-connections from the initial co-occurrence matrix, the co-occurrence values ​​between different source domain items are subjected to logarithmic transformation and symmetric normalization to obtain a symmetric normalized matrix. The symmetric normalized matrix elements in Used to characterize source domain items and source domain items Co-occurrence weights between them; For a target source domain item, calculate the reweighting coefficient between the target source domain item and its co-occurring neighbors. The co-occurring neighbors are source domain items whose co-occurrence weight with the target source domain item meets a preset requirement. The target source domain item is any one of the source domain items. Based on the reweighting coefficient, the text semantic embeddings corresponding to all the co-occurring neighbors are weighted and aggregated to obtain the neighborhood embedding of the target source domain item; The co-occurrence attribute of each source domain item is determined based on the neighborhood embedding; Among them, source domain items Corresponding co-occurrence attributes for: ; ; in, Represents a set of source domain items with non-zero neighborhood embeddings. for The mean of the middle neighborhood embedding, In order to be in The principal component transformation matrix fitted on top, Source Domain Items Corresponding co-occurring neighbors The reweighting coefficients, Indicates source domain items Corresponding text semantic embedding, Indicates co-occurring neighbors Corresponding text semantic embedding, This indicates the preset weight parameters. Indicates source domain items The mean of the cosine similarity with all source domain items. Used to convert input values ​​into modulation coefficients between 0 and 1.

[0012] Optionally, the step of determining the behavioral candidate score for each candidate item based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embedding corresponding to the candidate items through a behavioral scoring model includes: Based on the time features and the popularity-related proxies corresponding to the historical items, construct the historical features corresponding to the historical items; Based on the historical features and the co-occurrence information agent corresponding to the historical items, determine the behavioral location input corresponding to the historical items; According to the interaction order of the historical items in the historical interaction sequence, the behavior position input corresponding to the historical item is sequence encoded, and the historical behavior representation corresponding to the historical interaction sequence is determined based on the encoding state corresponding to the last valid historical position. Based on the text semantic embedding corresponding to the candidate item and the co-occurrence information proxy corresponding to the candidate item, the candidate behavior representation corresponding to the candidate item is determined. Based on the degree of matching between the historical behavior representation and the candidate behavior representation, a behavior candidate score is determined for each candidate item.

[0013] Secondly, embodiments of the present invention provide a behavior-enhanced semantic recommendation apparatus for zero-sample cross-domain sequence recommendation, comprising: The acquisition module is used to acquire a target query in the target domain. The target query includes a historical interaction sequence and multiple candidate items. The historical interaction sequence includes multiple historical items and interaction times that the target user interacted with before the current predicted position. The target domain includes multiple target domain items, multiple target domain users, and target domain interaction data between the target domain items and the target domain users. The historical items and the candidate items all belong to the target domain items, and the target user belongs to the target domain users. The semantic scoring module is used to determine the semantic candidate score of each candidate item based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item, through a semantic scoring model. The semantic scoring model is pre-trained based on data from the source domain, and the source domain includes multiple source domain items, multiple source domain users, and source domain interaction data between the source domain items and the source domain users. The first determining module is used to determine the time characteristics corresponding to the historical item based on the time information in the historical interaction sequence; The behavior scoring module is used to determine the behavior candidate score for each candidate item based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embeddings corresponding to the candidate items, through a behavior scoring model. The interaction attributes include popularity-related proxies and co-occurrence information proxies. The interaction attribute mapper includes a popularity mapper and a co-occurrence mapper. The popularity mapper is used to determine the popularity-related proxies based on the text semantic embeddings, and the co-occurrence mapper is used to determine the co-occurrence information proxies based on the text semantic embeddings. The popularity mapper and the co-occurrence mapper are pre-trained based on data from the source domain. The popularity-related proxies are used to represent the proxy results of the relative popularity of items learned from the source domain in the target domain, and the co-occurrence information proxies are used to represent the proxy results of the collaborative association features between items learned from the source domain in the target domain. The fusion module is used to fuse the semantic candidate score and the behavior candidate score to obtain a fused candidate score for each candidate item; The second determining module is used to determine the recommendation result corresponding to the target query based on the fusion candidate score.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor is configured to read the program in the memory to implement the steps in the behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation as described in the first aspect.

[0015] Fourthly, embodiments of the present invention provide a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation as described in the first aspect.

[0016] The beneficial effects of this invention are as follows: In the embodiments of this invention, firstly, a semantic scoring model pre-trained on the source domain is used to determine the semantic candidate score of the candidate item based on the historical interaction sequence corresponding to the target query and the textual semantic embeddings of historical items and candidate items, thereby utilizing the textual semantic information of the items to achieve cross-domain recommendation. Secondly, the temporal features corresponding to historical items are determined based on the temporal information in the historical interaction sequence, and the corresponding interaction attributes are determined based on the textual semantic embeddings of the target domain items using an interaction attribute mapper pre-trained on the source domain. This transfers the item interaction attribute mapping relationship learned in the source domain to the target domain without needing to retrain the interaction attribute mapper or reconstruct the item interaction attribute statistics at the target domain level using target domain interaction data. Thirdly, a behavior candidate score is determined by combining the behavior scoring model pre-trained on the source domain with temporal features, interaction attributes, and the textual semantic embeddings of candidate items. By fusing the behavior candidate score with the semantic candidate score, the recommendation result can simultaneously consider the semantic information corresponding to the target query and the behavior-related information transferred from the source domain, reducing the dependence of zero-sample cross-domain recommendation on target domain interaction data training and target domain behavior statistics, and improving the accuracy of candidate item recommendation in target domains with insufficient interaction data. Attached Figure Description

[0017] Figure 1 A flowchart of a behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the implementation framework of the behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation provided in an embodiment of the present invention. Figure 3 A schematic diagram of a behavior-enhanced semantic recommendation device for zero-sample cross-domain sequence recommendation provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] In the embodiments of this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first," "second," etc., in the specification of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such terms can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are usually of the same class, without limiting the number of objects. For example, the first object can be one or multiple.

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] This application provides a behavior-enhanced semantic recommendation method and apparatus for zero-shot cross-domain sequence recommendation, which aims to directly transfer models trained on a source domain with rich interactions to an unseen target domain, thereby improving the accuracy of recommendation results in the new domain.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating a behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation provided in an embodiment of the present invention. The method specifically includes the following steps: Step 101: Obtain the target query in the target domain. The target query includes a historical interaction sequence and multiple candidate items. The historical interaction sequence includes multiple historical items and interaction times that the target user interacted with before the current predicted position. The target domain includes multiple target domain items, multiple target domain users, and target domain interaction data between the target domain items and the target domain users. The historical items and the candidate items all belong to the target domain items, and the target user belongs to the target domain users.

[0023] Step 102: Based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item, determine the semantic candidate score corresponding to each candidate item through a semantic scoring model. The semantic scoring model is pre-trained based on data from the source domain. The source domain includes multiple source domain items, multiple source domain users, and source domain interaction data between the source domain items and the source domain users.

[0024] Step 103: Determine the time characteristics corresponding to the historical item based on the time information in the historical interaction sequence.

[0025] Step 104: Based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embeddings corresponding to the candidate items, a behavior candidate score is determined for each candidate item using a behavior scoring model. The interaction attributes include popularity-related proxies and co-occurrence information proxies. The interaction attribute mapper includes a popularity mapper and a co-occurrence mapper. The popularity mapper is used to determine the popularity-related proxies based on the text semantic embeddings, and the co-occurrence mapper is used to determine the co-occurrence information proxies based on the text semantic embeddings. The popularity mapper and the co-occurrence mapper are pre-trained based on data from the source domain. The popularity-related proxies are used to represent the proxy results of the relative popularity of items learned from the source domain in the target domain, and the co-occurrence information proxies are used to represent the proxy results of the collaborative association features between items learned from the source domain in the target domain.

[0026] Step 105: Fuse the semantic candidate score and the behavioral candidate score to obtain the fused candidate score for each candidate item.

[0027] Step 106: Determine the recommendation result corresponding to the target query based on the fusion candidate score.

[0028] In practical implementation, the source domain is an existing domain rich in interactive data, and the target domain is a new domain. The source and target domains can be two completely non-overlapping or only partially overlapping domains. For ease of description and distinction, items and users in the source domain are referred to as source domain items and source domain users, respectively, and items and users in the target domain are referred to as target domain items and target domain users, respectively.

[0029] In practical implementation, an item refers to an actual recommended object that can be provided or displayed to the user, such as goods or videos on an e-commerce platform, video or audio content on an audio content distribution platform, news content on a news information platform, or applications on an application distribution platform. Interaction data refers to the data generated and recorded by the information service system during operation based on user interactions with items. Interactions include at least one of browsing, clicking, favorites, purchasing, playing, and downloading. Interaction data may include the user identifier, item identifier, and corresponding interaction time.

[0030] In this embodiment, the source domain and target domain correspond to different recommendation business domains, which can correspond to different product categories, content categories, service categories, application scenarios, or different information service platforms. The source domain is the recommendation business domain with historical interaction data used to train the semantic scoring model, behavioral scoring model, and interaction attribute mapper, while the target domain is the recommendation business domain for which item recommendations are to be performed. In the zero-sample cross-domain recommendation scenario, the semantic scoring model, behavioral scoring model, and interaction attribute mapper are not retrained using the interaction data of the target domain. Instead, the semantic scoring model, behavioral scoring model, and interaction attribute mapper pre-trained based on the source domain are used to process the target user's current recommendation request in the target domain to determine the fusion candidate scores corresponding to multiple candidate items. Based on the fusion candidate scores, the item to be recommended is determined from the multiple candidate items. Finally, based on the recommendation results, the recommendation interface of the user terminal is controlled to display the item to be recommended, or the item information corresponding to the item to be recommended is sent to the user terminal to provide the target user with the corresponding item recommendation service. Represented as , ,in, Used to characterize the source domain. Used to characterize the target domain and Each is a field The user set and the item set in the middle. Represents a collection of metadata. This represents a set of interaction sequences. As a specific example, the source and target domains do not overlap at the user and item levels, i.e. ,and .

[0031] Specifically, text metadata is the text used to describe an item. The corresponding text metadata is In some embodiments, the text metadata includes at least one of a title, description, and category. The interaction sequence is used to describe the interaction between the user and the item and the timing of the interaction. Corresponding interaction sequence for: ; in, Indicates user In position Interactive items For the timestamp corresponding to the interaction, Interactive sequence The sequence length, and .

[0032] For target domain users and predicted location , Only the target domain prefix preceding this position can be observed: ; In practice, each query is processed independently. Items and timestamps from the current prefix can be used during a query, but cross-user or cross-query aggregation, storage, or reuse of target domain interactive statistics is prohibited. Represents a historical interaction sequence. This indicates the subsequence of historical items contained within. This represents the nth historical item. , This represents the set of candidate items.

[0033] In some embodiments, the set of candidate items is predetermined before performing semantic scoring and behavioral scoring; the specific method of determination is not limited here. For example, multiple candidate items can be determined from the target domain item set according to a preset candidate generation rule, and the determined multiple candidate items, together with the current historical interaction sequence, constitute the target query. The specific candidate generation rule is not limited here and can be implemented by the candidate recall module in the recommendation system, preset business rules, or a pre-configured set of candidate items.

[0034] In some embodiments, prior to step 102, the method further includes: The text metadata corresponding to the target domain item is input into a pre-trained text semantic encoder for encoding processing to obtain the text semantic embedding corresponding to the target domain item. The text metadata includes at least one of title, description and category.

[0035] For the field Items in Its corresponding text semantic embedding for: ; in, Used to characterize a pre-trained and frozen text semantic encoder. Used to characterize the domain Items in The specific structure of the text metadata is not limited here.

[0036] As an optional implementation, before the query phase, the text semantic embeddings corresponding to all items in the target domain are pre-acquired. In the subsequent query phase, the text semantic embeddings corresponding to historical items and candidate items can be directly retrieved. Alternatively, after determining the historical items and candidate items, they are respectively input into a pre-trained text semantic encoder for encoding processing to obtain the text semantic embeddings corresponding to the historical items and the candidate items.

[0037] In this embodiment, a semantic scoring model pre-trained and frozen in the source domain is used to determine the semantic candidate score corresponding to each candidate item in the target domain. In some embodiments, the semantic scoring model is called through a general interface that is compatible with semantic scoring models using masked positional sequence encoding or causal end-positional sequence encoding, and the numerical scale of the original semantic candidate scores output by different semantic scoring models does not need to be consistent.

[0038] Optionally, in some embodiments, step 102 includes: The candidate item representation is obtained by mapping the text semantic embedding corresponding to the candidate item through the candidate item mapping function; The text semantic embedding corresponding to the historical item is mapped by the historical item mapping function to obtain the historical item representation corresponding to the historical item. Encode and read out multiple historical item representations arranged according to the historical interaction sequence to obtain the sequence representation corresponding to the historical interaction sequence; The semantic candidate score for each candidate item is determined based on the sequence representation and the representation of each candidate item.

[0039] In some embodiments, the semantic scoring model includes a historical item mapping function, a candidate item mapping function, a sequence encoder, and readout rules. The historical item mapping function maps the text semantic embeddings corresponding to historical items to obtain historical item representations. The candidate item mapping function maps the text semantic embeddings corresponding to candidate items to obtain candidate item representations. The sequence encoder and readout rules encode and read out each historical item representation arranged according to the historical interaction sequence to obtain a sequence representation corresponding to the historical interaction sequence, and determine the semantic candidate score corresponding to the candidate item based on the sequence representation and the candidate item representation.

[0040] For example, in some embodiments, the semantic scoring model can be built based on BERT4Rec-Llama, SAGERec, or UniSRec. Specifically, when the semantic scoring model is built based on BERT4Rec-Llama, the learned item projection is combined with its native bidirectional encoder to sequence-encode historical items in the historical interaction sequence and obtain corresponding semantic candidate scores; when the semantic scoring model is built based on SAGERec, independent recommended projection and generalized projection are used on the historical side, while recommended projection is retained on the candidate side; when the semantic scoring model is built based on UniSRec, parametric whitening experts are used on both the historical and candidate sides, and deterministic routing is used during the inference phase.

[0041] For the field Historical interaction sequence in Candidate items Corresponding semantic candidate score for: ; in, Used to characterize the historical item mapping function Used to characterize the candidate-side item mapping function. Used to characterize sequence encoders and readout rules Candidate items Corresponding text semantic embedding, The nth historical item in the historical interaction sequence The corresponding text semantic embedding.

[0042] The behavioral scoring model provides a scoring channel decoupled from the frozen semantic scoring model parameters. Given a historical interaction sequence... and candidate items The behavior scoring model outputs the behavior candidate scores corresponding to the candidate items. Specifically, the behavioral scoring model integrates two types of evidence: the time channel extracts local descriptions of queries from the time information of historical interaction sequences, while the interaction attribute channel transfers popularity proxies and co-occurrences defined by the source domain through an interaction attribute mapper.

[0043] In this embodiment, instead of directly statistically analyzing the behavioral attributes of items in the target domain based on the target domain interaction data, the popularity and co-occurrence attributes corresponding to the source domain items are first constructed based on the source domain interaction data. These popularity and co-occurrence attributes are then used as supervised targets to train an interaction attribute mapper that establishes a mapping relationship between the textual semantic embeddings of items and their corresponding behavioral attributes. After training the interaction attribute mapper, its model parameters are kept fixed, and the textual semantic embeddings of the target domain items are input into the interaction attribute mapper to obtain the interaction attributes defined by the behavioral patterns of the source domain and transferred to the target domain. Therefore, the interaction attributes corresponding to the target domain items are not obtained by re-statistically analyzing the target domain interaction data, but rather predicted through the mapping relationship between the textual semantic information learned from the source domain and the behavioral attributes of the items.

[0044] Optionally, in some embodiments, the time feature includes a time interval feature and a relative time position feature, and step 103 includes: Based on the time information in the historical interaction sequence, the time interval feature and relative time position feature corresponding to each historical item are determined respectively, wherein the time interval feature corresponding to the j-th historical item is... for: ; ; ; The relative time location feature of the j-th historical item for: ; in, Let j be the time interval corresponding to the j-th historical item. Let j be the interaction time corresponding to the j-th historical item. Let j-1 be the interaction time corresponding to the historical item. The quantity of the historical items, For a fixed time scale, and These are all preset constants used to prevent division by zero. Used to determine the maximum value among multiple values ​​within parentheses.

[0045] Specifically, for historical interaction sequences with non-decreasing timestamps : ; in, Used to represent the j-th historical item in the historical interaction sequence. The interaction times are used to represent the interaction time corresponding to the j-th historical item. The interaction times in the above historical interaction sequence are arranged in non-decreasing order.

[0046] The time interval for each historical item is determined based on the time difference between the interaction times of adjacent historical items. For the first historical item, the corresponding time interval is set to... , and when From time to time Furthermore, logarithmic transformation and normalization are performed on the time intervals corresponding to each historical item to obtain the time interval characteristics for each historical item. The normalized logarithmic time interval characteristics are defined as follows: ; in, The time interval feature corresponding to the j-th historical item is... For a fixed time scale, This is a preset constant used to prevent division by zero. It is used to measure the size of the time interval before the current position relative to the maximum logarithmic interval of the historical interaction sequence. This is a maximum value function used to determine the maximum value among multiple values ​​within parentheses. In this embodiment, the maximum value determination operation sets a preset lower limit for the denominator to avoid the denominator being zero or too small.

[0047] Furthermore, based on the interaction time corresponding to each historical item, the interaction time corresponding to the first historical item in the historical interaction sequence, and the interaction time corresponding to the last historical item, the relative time position feature corresponding to each historical item is determined according to the following formula, whereby the relative time position feature is defined as: ; in, Let j be the relative time location feature corresponding to the j-th historical item. A pre-set constant, This decreases gradually from earliest to latest position over the time span of the historical interaction sequence. It should be understood that when all timestamps are identical, each valid position has... .

[0048] For each target query, the time interval feature and the relative time position feature are determined based on the time information in the historical interaction sequence corresponding to that target query, without needing to determine the target domain-level time statistics based on the historical interaction data of different target users in the target domain. Therefore, the time features are recalculated independently for each target query, without relying on any target domain-level aggregate statistics.

[0049] In this embodiment, the interaction attribute includes a popularity-related proxy and a co-occurrence information proxy, and the interaction attribute mapper includes a popularity mapper and a co-occurrence mapper; The popularity-related agent is determined by embedding the text semantics corresponding to the target domain item into the trained popularity mapper, which is pre-trained based on the data of the source domain. The co-occurrence information agent is determined by embedding the text semantics corresponding to the target domain item into the pre-trained co-occurrence mapper, which is pre-trained based on the data of the source domain.

[0050] Both the popularity mapper and the co-occurrence mapper are pre-trained and frozen on the source domain. Before step 104, the text semantic embeddings corresponding to historical items are input into the popularity mapper and the co-occurrence mapper respectively to obtain the popularity-related agent and the co-occurrence information agent corresponding to the historical items. The text semantic embeddings corresponding to candidate items are input into the popularity mapper and the co-occurrence mapper respectively to obtain the popularity-related agent and the co-occurrence information agent corresponding to the candidate items.

[0051] It should be understood that popularity and co-occurrence attributes describe the interactive behavior of source domain items from different dimensions. By learning the mapping relationship between text semantic embeddings and the above two types of behavioral attributes, even if the target domain items lack intra-domain interaction data for reconstructing the corresponding behavioral statistics, they can still obtain proxy representations that correspond to the behavioral attributes of the source domain.

[0052] Optionally, in some embodiments, before step 104, the method further includes: Based on the source domain interaction data, the popularity attribute corresponding to the source domain item is determined, and the popularity attribute is used to characterize the relative popularity of the source domain item among the multiple source domain items; The text semantics corresponding to the source domain item are embedded and input into the popularity mapper to obtain the popularity-related proxy corresponding to the source domain item. The parameters of the popularity mapper are adjusted based on the difference between the popularity-related agent and the popularity attribute to obtain the trained popularity mapper.

[0053] Optionally, in some embodiments, determining the popularity attribute corresponding to the source domain item based on the source domain interaction data includes: The source domain interaction data is divided into multiple time periods according to a preset time period, and the number of interactions of the source domain items in each time period is counted. Based on the number of interactions and the preset attenuation coefficient, the attenuation status of the source domain item in each time period is determined; For source domain items with non-zero decay state, the average parallel rank is assigned, and then normalized after being divided by the target value to obtain the normalized rank value of the source domain item in each time period. The target value is the maximum average rank plus one. The popularity attribute corresponding to the source domain item is determined based on the normalized rank value; Among them, source domain items Corresponding popularity attribute for: ; ; in, This indicates the last time period mentioned. Indicates the number of time periods at the end of the aggregation. Source Domain Items In the The normalized rank value corresponding to each time period Source Domain Items In the The decay state over a time period, Source Domain Items In the Number of interactions per time period The preset attenuation coefficient is denoted as .

[0054] In this embodiment, source domain interaction data is divided according to a preset time period. The decaying interaction state is determined based on the number of interactions of each source domain item within different time periods. A normalized rank value is determined based on the relative order of the decaying interaction states of each source domain item within the corresponding time period. Furthermore, the normalized rank values ​​corresponding to multiple time periods within a preset aggregation window are aggregated to obtain the popularity attribute corresponding to the source domain item. Therefore, the popularity attribute does not directly use the original number of interactions of the source domain item, but rather reflects the popularity of the source domain item relative to other source domain items in recent multiple time periods.

[0055] As a concrete example, using the week as the preset time period, the week number corresponding to the source domain is denoted as... Source domain items In the The number of interactions per week is recorded as For source domain items Its initial decay state is After the weekly count is added, update the number of weeks according to the following formula. The corresponding decay state of the week : ; in, Used to characterize source domain items In the The corresponding decay state of the week, The preset attenuation coefficient is used, and its value is not limited here. According to the above formula, in this embodiment, the attenuation is performed by adding the number of interactions in the current time period to the attenuation state of the previous time period, so that the number of interactions in the current time period and the accumulated interaction information in the previous time period are updated according to the preset attenuation coefficient.

[0056] For each time period, the normalized rank value of each source domain item is determined based on its decay state. Specifically, the normalized rank value of source domain items with a decay state of zero is set to zero. For source domain items with a non-zero decay state, the rank is assigned according to the average rank, and then normalized by dividing by "the maximum average rank plus one" to obtain the source domain item. In the The normalized rank value corresponding to Zhou is .

[0057] Based on the normalized rank values ​​corresponding to multiple time periods within the aggregation window at the end of the source domain, the popularity attribute of the source domain item is determined. Source Domain Items Corresponding popularity attribute for: ; in, This indicates the last time period corresponding to the source domain. This indicates the number of time periods at the end of the aggregation, i.e., how many recent time periods are taken backwards from the last time period for averaging when calculating the popularity attribute. In this embodiment, the popularity attribute is determined based on the decay state and relative ranking of source domain items over multiple time periods to characterize the relative popularity of source domain items over recent time periods. This popularity attribute summarizes the recent percentile trajectory of source domain items, rather than the original number of interactions.

[0058] For any field Items in The popularity mapper determines the corresponding popularity-related agent according to the following formula. : ; in, For the field Items in Corresponding text semantic embedding, Used to characterize popularity mappers Parameters used to characterize the popularity mapper Used to characterize numerical truncation operations.

[0059] After determining the popularity attributes of each source domain item, a popularity mapper is trained based on the text semantic embeddings and corresponding popularity attributes of the source domain items to establish a mapping relationship between the text semantic embeddings and popularity attributes of the items. During the training phase, the text semantic embeddings of the source domain items are input into the popularity mapper, and the prediction results output by the popularity mapper are used as the basis for the training. Popularity attribute obtained by calculation The model parameters are adjusted based on the differences between them. In some embodiments, mean squared error is used as the training loss for the popularity mapper.

[0060] Before inference in the target domain, the text semantic embeddings corresponding to the target domain items are input into the frozen popularity mapper after training. This yields a popularity-related agent for the target domain items. The popularity-related agent is determined based on the mapping relationship between the text semantic embeddings and popularity attributes learned from the source domain data, without needing to statistically analyze the popularity of target domain items based on historical interaction data. Since the popularity mapper is trained based on the source domain, applying the frozen popular mapper to the text semantic embeddings of target domain items results in a popularity agent defined by the source domain, rather than the true popularity estimated based on target domain interactions.

[0061] Optionally, in some embodiments, before step 104, the method further includes: Based on the source domain interaction data, the co-occurrence attribute corresponding to the source domain item is determined. The co-occurrence attribute is used to characterize the association features between the source domain item and other source domain items based on user interaction behavior. The text semantic embedding corresponding to the source domain item is input into the co-occurrence mapper to obtain the co-occurrence information proxy corresponding to the source domain item; The parameters of the co-occurrence mapper are adjusted based on the difference between the co-occurrence information agent and the co-occurrence attribute to obtain the trained co-occurrence mapper.

[0062] Optionally, in some embodiments, determining the co-occurrence attribute of the source domain item based on the source domain interaction data includes: Construct a source domain user-item association matrix and an initial co-occurrence matrix based on the source domain interaction data; After removing self-connections from the initial co-occurrence matrix, the co-occurrence values ​​between different source domain items are subjected to logarithmic transformation and symmetric normalization to obtain a symmetric normalized matrix. The symmetric normalized matrix elements in Used to characterize source domain items and source domain items Co-occurrence weights between them; For a target source domain item, calculate the reweighting coefficient between the target source domain item and its co-occurring neighbors. The co-occurring neighbors are source domain items whose co-occurrence weight with the target source domain item meets a preset requirement. The target source domain item is any one of the source domain items. Based on the reweighting coefficient, the text semantic embeddings corresponding to all the co-occurring neighbors are weighted and aggregated to obtain the neighborhood embedding of the target source domain item; The co-occurrence attribute of each source domain item is determined based on the neighborhood embedding; Among them, source domain items Corresponding co-occurrence attributes for: ; ; in, Represents a set of source domain items with non-zero neighborhood embeddings. for The mean of the middle neighborhood embedding, In order to be in The principal component transformation matrix fitted on top, Source Domain Items Corresponding co-occurring neighbors The reweighting coefficients, Indicates source domain items Corresponding text semantic embedding, Indicates co-occurring neighbors Corresponding text semantic embedding, This indicates the preset weight parameters. Indicates source domain items The mean of the cosine similarity with all source domain items. Used to convert input values ​​into modulation coefficients between 0 and 1.

[0063] To leverage the collaborative signals of similar items, this embodiment constructs a co-occurrence neighborhood that depends solely on the source domain. To determine the co-occurrence attributes corresponding to source domain items, a binary source domain user-item association matrix is ​​constructed based on historical interaction data between source domain users and source domain items. And further determine the initial co-occurrence matrix. The elements in the initial co-occurrence matrix are the co-occurrence values ​​between items from different source domains.

[0064] Remove the initial co-occurrence matrix After self-joining, the co-occurrence values ​​of items from different source domains are subjected to the following logarithmic transformation to obtain the co-occurrence matrix. : , ; ; ; in, Source Domain Items and source domain items Co-occurrence values ​​between them Source Domain Items and source domain items The unnormalized co-occurrence weights between them.

[0065] According to the degree matrix For co-occurrence matrix Perform symmetric normalization to obtain the symmetric normalized matrix. for: .

[0066] For an item with a degree of zero, the corresponding inverse square root degree is set to zero.

[0067] For source domain items From the symmetric normalized co-occurrence matrix Items from the Source Domain Select from the non-zero elements of the corresponding matrix rows, in descending order of element value. Select a non-zero element, and determine the source domain item corresponding to the selected non-zero element as the source domain item. Co-occurring neighbors, obtain source domain items The corresponding set of co-occurring neighbors After determining the set of co-occurring neighbors, the set of co-occurring neighbors is fixed, and the co-occurrence weight between the source domain item and each co-occurring neighbor in the set is modulated based on the textual semantic similarity between the source domain item and each co-occurring neighbor in the set.

[0068] make Indicates source domain items The mean of the cosine similarity between the source domain item set and all source domain items in the source domain item set. For Preserve the reweighting coefficients of the edges for: ; in, Indicates source domain items Corresponding text semantic embedding, Indicates source domain items Corresponding text semantic embedding, This indicates the preset weight parameters. Used to convert input values ​​into modulation coefficients between 0 and 1, in this embodiment, the Sigmoid maps input values ​​to between 0 and 1 to modulate the co-occurrence weights between the source domain item and its corresponding co-occurring neighbors based on the textual semantic similarity between the source domain item and its corresponding co-occurring neighbors. It should be understood that the Sigmoid factor only modulates already selected neighbors and does not introduce new neighbors.

[0069] The reweighted coefficients corresponding to each co-occurring neighbor are normalized, and the text semantic embeddings corresponding to each co-occurring neighbor are weighted and aggregated based on the normalized reweighted coefficients to obtain the source domain items. Corresponding neighborhood embedding There are no items that effectively preserve their neighborhoods, and their neighborhood embeddings are set to zero vectors.

[0070] make Let represent the set of source domain items with non-zero neighborhood embeddings. Dimensionality reduction is performed on the neighborhood embeddings corresponding to the source domain items with non-zero neighborhood embeddings to obtain the co-occurrence attributes corresponding to each source domain item. In some embodiments, a centralized principal component analysis method without whitening is used for dimensionality reduction to obtain the source domain items. Corresponding co-occurrence attributes for: ; in, yes The mean of the above, The principal component transformation matrix fitted on this set, For the dimension of text semantic embedding, This represents the dimension of the co-occurrence attribute obtained after principal component transformation.

[0071] For any field Items in The co-occurrence mapper determines the co-occurrence information agent corresponding to the item according to the following formula. : ; in, For the field Chinese items Corresponding text semantic embedding, Used to characterize co-occurrence mappers Parameters used to characterize the co-occurrence mapper.

[0072] Based on the textual semantic embeddings and co-occurrence attributes of source domain items, a co-occurrence mapper is trained to establish a mapping relationship between the textual semantic embeddings of items and co-occurrence information proxies. In some embodiments, the co-occurrence mapper is fitted to a specified subset of source domain items, stratified by degree, using mean squared error.

[0073] After training, the text semantic embeddings corresponding to the target domain items are input into the trained and frozen co-occurrence mapper to obtain the co-occurrence information proxy for the target domain items. The co-occurrence information proxy is determined based on the mapping relationship between the text semantic embeddings and co-occurrence attributes learned from the source domain data, without needing to reconstruct the co-occurrence relationship between target domain items based on historical interaction data of the target domain. In this embodiment, the co-occurrence mapper is pre-trained and frozen based on the source domain; therefore, the target domain obtains a proxy representation defined and transferred from the source domain, rather than statistics reconstructed based on target interactions.

[0074] In some embodiments, the behavior scoring model encodes the historical interaction sequence and each candidate item based on time features, interaction attributes corresponding to historical items, and interaction attributes corresponding to candidate items, and determines the behavior candidate score corresponding to each candidate item based on the encoded historical behavior representation and candidate behavior representation.

[0075] Optionally, in some embodiments, step 104 includes: Based on the time features and the popularity-related proxies corresponding to the historical items, construct the historical features corresponding to the historical items; Based on the historical features and the co-occurrence information agent corresponding to the historical items, determine the behavioral location input corresponding to the historical items; According to the interaction order of the historical items in the historical interaction sequence, the behavior position input corresponding to the historical item is sequence encoded, and the historical behavior representation corresponding to the historical interaction sequence is determined based on the encoding state corresponding to the last valid historical position. Based on the text semantic embedding corresponding to the candidate item and the co-occurrence information proxy corresponding to the candidate item, the candidate behavior representation corresponding to the candidate item is determined. Based on the degree of matching between the historical behavior representation and the candidate behavior representation, a behavior candidate score is determined for each candidate item.

[0076] As a specific implementation, the interaction attributes include popularity-related proxies and co-occurrence information proxies, and the temporal features include time interval features and relative temporal position features. Specifically, the popularity-related proxies are used to construct three-dimensional historical features, while the co-occurrence information proxies are injected at the position level before causal sequence encoding. Thus, popularity-related information, co-occurrence coordination information, and temporal information participate in the determination of historical behavior representation and candidate behavior representation according to their corresponding information meanings.

[0077] like Figure 2 As shown, for the j-th historical item in the historical interaction sequence Based on the time interval characteristics corresponding to the historical item Relative time and location characteristics And popularity-related agents Constructing three-dimensional historical features for: ; in, Used to characterize splicing operations.

[0078] Based on the historical characteristics of the j-th historical item And co-occurrence information proxy for historical items The input of the behavior position corresponding to the j-th historical item is determined according to the following formula. : ; in, This indicates the historical feature mapping module. This indicates the historical co-occurrence information projection module. Representing historical items Corresponding co-occurrence information agent, This represents the historical scaling factor.

[0079] In this embodiment, the popularity-related proxies corresponding to historical items are concatenated with time interval features and relative time location features to form historical features. The co-occurrence information proxies corresponding to historical items are then fused with the mapping results corresponding to historical features in a location-level manner after projection processing.

[0080] According to the interaction order of each historical item in the historical interaction sequence, the behavior position corresponding to each historical item is input into the causal history encoding module for sequence encoding, and the historical behavior representation corresponding to the historical interaction sequence is determined based on the encoding state corresponding to the last valid historical position. : ; in, Indicates the causal history encoding module, subscript " "" indicates that the encoding state output by the causal history encoding module at the last valid historical position in the historical interaction sequence has been obtained. This indicates the historical normalization module, used to perform normalization operations. This refers to the behavioral representation dimension in the behavioral scoring model.

[0081] As a specific implementation, the causal history encoding module takes the user's historical interaction sequence (item identifiers and their interaction timestamps) as input. First, it constructs a behavioral feature vector for each position in the historical interaction sequence, including: a log-normalized feature of the time interval between the previous interaction and the previous interaction, a recentity feature of the interaction time, a popularity quantile feature obtained by a popularity predictor trained in the source domain based on the textual semantic embedding of the item, and behavioral features output by a behavior mapper based on item co-occurrence statistical learning. Each feature is normalized by layers and encoded by a two-layer feedforward network (linear layer—GELU activation—Dropout—linear layer), then superimposed with a sinusoidal positional encoding and fed into a multi-head Transformer encoder. This encoder uses an upper triangular causal attention mask, ensuring that each position can only focus on itself and earlier historical positions, guaranteeing that the modeling of historical states relies solely on "past" interaction information. Finally, the hidden states of the effective positions at the end of the sequence are extracted, normalized by layers, and output as a causal encoded representation of the user's historical interests.

[0082] For any candidate item Based on the text semantic embedding and co-occurrence information proxy of the candidate items, the candidate items are determined according to the following formula. Corresponding candidate behavior input : ; in, Indicates candidate items Corresponding text semantic embedding, This indicates the candidate-side text semantic projection module. This indicates the candidate co-occurrence information projection module. Indicates candidate items Corresponding co-occurrence information agent, This represents the candidate-side scaling factor.

[0083] Furthermore, regarding candidate items The candidate behavior input is mapped and normalized to obtain candidate items. Corresponding candidate behavior representation : ; in, Indicates candidate encoding modules. This indicates the candidate-side normalization module, which is used to perform normalization operations.

[0084] As a specific implementation, the candidate encoding module employs a dual-branch structure to encode each candidate item: the semantic branch performs layer normalization and linear projection on the frozen large language model semantic embedding (item text vector generated by the pre-trained language model) of the candidate item to obtain a semantic latent vector; the behavior branch obtains the behavior features of the candidate item through a behavior lookup table (generated by a behavior mapper constructed from the co-occurrence statistics of items in the source domain), performs layer normalization and linear projection, multiplies it by a learnable scalar gating coefficient, and adds it element-wise with the output of the semantic branch to achieve adaptive weighted fusion of semantic and behavioral information. The fused vector is then encoded and normalized by a shared two-layer feedforward network to output the final encoded representation of the candidate item; this representation is then multiplied by the sequence state vector output by the causal history encoding module to obtain the user's preference score for the candidate item.

[0085] Based on historical behavioral representation With candidate items Corresponding candidate behavior representation The degree of matching between them determines the candidate items. The corresponding behavioral candidate score. In some embodiments, candidate items are determined according to the following formula. behavioral candidate scores : .

[0086] Through the above steps, the behavioral candidate score and semantic candidate score of each candidate item can be obtained. In some embodiments, the fusion candidate score of the candidate item is obtained by weighted summing of the behavioral candidate score and the semantic candidate score.

[0087] Optionally, in some embodiments, step 105 includes: determining the mean and overall standard deviation of the semantic candidate scores based on the semantic candidate scores of all the candidate items, and determining the mean and overall standard deviation of the behavioral candidate scores based on the behavioral candidate scores of all the candidate items; The semantic candidate scores of each candidate item are standardized based on the mean and overall standard deviation of the semantic candidate scores to obtain standardized semantic candidate scores. The behavioral candidate scores of each candidate item are also standardized based on the mean and overall standard deviation of the behavioral candidate scores to obtain standardized behavioral candidate scores. The standardized semantic candidate scores and standardized behavioral candidate scores are fused based on the fusion coefficient to obtain the fused candidate score corresponding to the candidate item, wherein the fusion coefficient is predetermined based on the data of the source domain.

[0088] Specifically, for each of the multiple candidate items, its semantic candidate score and behavioral candidate score are standardized at the query level. For each candidate item... The corresponding standardized candidate scores are determined according to the following formula. : ; in, , and Represent the candidate score set respectively The mean and population standard deviation, For example, a fixed lower limit for the value. .

[0089] It should be understood that, in practice, semantic candidate scores and behavioral candidate scores are standardized separately. The mean of the semantic candidate scores is calculated. and the overall standard deviation of semantic candidate scores Based on this, the semantic candidate score of each candidate item is standardized. Similarly, the mean of the behavioral candidate scores is calculated. and the overall standard deviation of candidate scores for behavior Based on this, the behavioral candidate scores for each candidate item are standardized.

[0090] Through the above standardization process, even if the original candidate scores output by the semantic scoring model and the behavioral scoring model have different numerical scales, standardized candidate scores within a fusionable scale can still be obtained. Furthermore, the mean and overall standard deviation used in the above standardization are determined based on the candidate scores of the candidate items corresponding to the current target query, without needing to determine statistical parameters based on other queries in the target domain or interaction data from other users.

[0091] It should be noted that this embodiment does not simply apply the model trained on source domain interaction data directly to the target domain. For sequence recommendation models that primarily rely on item text semantics for cross-domain transfer, although text semantic embedding can establish a shared semantic representation space among non-overlapping items in different domains, text semantic information cannot explicitly reflect the time rhythm, recentity, or other local query behavior features corresponding to the current historical interaction sequence, nor can it directly characterize the popularity of items and co-occurrence relationships between items generated by user interaction behavior. Therefore, when performing zero-shot transfer using only a semantic scoring model trained on the source domain, behavioral information in the target domain is difficult to utilize effectively.

[0092] In this embodiment, the semantic scoring model and the behavioral scoring model can adopt different model structures and perform parameter optimizations separately. The candidate scores output by the two models may have different numerical positions and numerical scales. If the original candidate scores are directly weighted or summed, the fusion coefficient is difficult to stably represent the relative contribution of the two types of scoring information. Furthermore, training a dynamic gating or routing model based on target domain interaction data introduces target domain supervision, which conflicts with the information constraints of zero-shot cross-domain recommendation. To address this, this embodiment utilizes an interaction attribute mapper trained based on source domain interaction data to transfer the mapping relationship between the popularity attribute and co-occurrence attribute defined in the source domain to the target domain, and generates behavioral candidate scores by combining the time information in the current target query. On the other hand, the semantic candidate scores and behavioral candidate scores are query-level standardized separately, and then fused using a fusion coefficient predetermined only based on the source domain data. Thus, semantic evidence is combined with transferable behavioral evidence without using target domain interaction data for model training or fusion calibration.

[0093] As an optional implementation, the standardized semantic candidate scores and standardized behavioral candidate scores are fused to obtain a fused candidate score for each candidate item. For example, the standardized semantic candidate scores and standardized behavioral candidate scores are summed to obtain the corresponding fused candidate score.

[0094] In some embodiments, fusion coefficients are pre-determined based on data from the source domain, and standardized semantic candidate scores and standardized behavioral candidate scores are fused to obtain fused candidate scores. : ; in, The fusion coefficients are pre-determined in the source domain.

[0095] Specifically, a set of fusion coefficients, including multiple candidate fusion coefficients, is pre-defined. For each candidate fusion coefficient in the set... Based on the candidate fusion coefficient, the standardized semantic candidate scores and standardized behavioral candidate scores in the source domain are fused, and the recommendation performance evaluation index corresponding to the candidate fusion coefficient is determined according to the corresponding recommendation results. For example, in some embodiments, NDCG@10 is used as the recommended performance evaluation metric.

[0096] Final fusion coefficient for: ; ; in, (NDCG numerical unit) This rule includes the exact parallel case, where when the performance of multiple source domain settings is approximately equivalent, the coefficient with the smaller behavioral contribution is preferred, thus forming a more conservative and predictable fusion strategy.

[0097] In some embodiments, after obtaining the fusion candidate score for each candidate item, the candidate items are ranked from highest to lowest according to the fusion candidate score using a fixed parallel processing strategy, thereby determining the recommendation result. In other embodiments, the candidate item with the highest fusion candidate score is determined as the recommendation result.

[0098] As a specific example, please refer to Figure 2 ,like Figure 2 As shown, the semantic scoring model, behavioral scoring model, popularity mapper, co-occurrence mapper, and fusion coefficients are all predetermined and frozen in the source domain.

[0099] make , and This represents a non-overlapping source domain partition for three users, used for parameter optimization, checkpoint and non-fusion hyperparameter selection, and fusion coefficient selection, respectively. In the specific implementation, all trainable components are selected only from... Receive interactive supervision. Non-overlapping source domain verification partitions for users are used for checkpoint and non-fusion hyperparameter selection. Specifically designed for source domain fusion calibration. During the target phase, neither target popularity counts nor target co-occurrence maps are constructed, and no development decisions are made using target labels or target ranking metrics. During inference, all parameters remain frozen, and training-era stochastic mechanisms such as dropout are disabled; deterministic routing within the frozen semantic scoring model still functions correctly. Therefore, the information available on the target domain side is strictly limited to current historical items and their timestamps, the current set of candidate items, and item metadata allowed by the evaluation protocol.

[0100] In the interactive attribute channel, first according to Construct an item-level training objective, and then fit it to the semantic embedding from the text. Two mapping functions to these targets yield a popularity mapper and a co-occurrence mapper. In some embodiments, for each trained semantic scoring model and running instance, Only The selection is made above and remains fixed across all target queries corresponding to that instance. During the source domain development phase, the semantic backbone, attribute mapper, and behavior expert are trained separately, and the source domain data is used to select... .

[0101] The semantic scoring model and the behavior scoring model do not share trainable parameters. After the source domain checkpoint selection is completed, all participants... All computational components are frozen, so behavioral expert training does not update the semantic scorer. The popularity mapper and co-occurrence mapper are fitted independently to the semantic scoring model and remain frozen during behavioral scoring model training and target domain inference. During specific training, after the source domain fitting is complete, the two mappers are shared between different semantic scoring models and random seed instances and remain frozen throughout all subsequent stages.

[0102] In this embodiment, a parameter-decoupled behavioral scoring model is introduced based on the semantic scoring model, and predictable fusion is achieved through rigorous source domain calibration. By combining query local temporal evidence with the interaction attributes of source domain definition and text conditional transfer, a modular, reusable, and easily deployable basic solution for zero-shot cross-domain sequence recommendation can be provided without developing using target domain interaction corpora.

[0103] like Figure 3 As shown, this embodiment of the invention also provides a behavior-enhanced semantic recommendation device 300 for zero-sample cross-domain sequence recommendation, comprising: The acquisition module 301 is used to acquire a target query in the target domain. The target query includes a historical interaction sequence and multiple candidate items. The historical interaction sequence includes multiple historical items and interaction times that the target user interacted with before the current predicted position. The target domain includes multiple target domain items, multiple target domain users, and target domain interaction data between the target domain items and the target domain users. The historical items and the candidate items both belong to the target domain items, and the target user belongs to the target domain users. The semantic scoring module 302 is used to determine the semantic candidate score of each candidate item based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item through a semantic scoring model. The semantic scoring model is pre-trained based on data from the source domain. The source domain includes multiple source domain items, multiple source domain users, and source domain interaction data between the source domain items and the source domain users. The first determining module 303 is used to determine the time characteristics corresponding to the historical item based on the time information in the historical interaction sequence; The behavior scoring module 304 is used to determine the behavior candidate score for each candidate item based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embedding corresponding to the candidate items, through a behavior scoring model. The interaction attributes include popularity-related proxies and co-occurrence information proxies. The interaction attribute mapper includes a popularity mapper and a co-occurrence mapper. The popularity mapper is used to determine the popularity-related proxies based on the text semantic embedding, and the co-occurrence mapper is used to determine the co-occurrence information proxies based on the text semantic embedding. The popularity mapper and the co-occurrence mapper are pre-trained based on data from the source domain. The popularity-related proxies are used to represent the proxy results of the relative popularity of items learned from the source domain in the target domain, and the co-occurrence information proxies are used to represent the proxy results of the collaborative association features between items learned from the source domain in the target domain. The fusion module 305 is used to fuse the semantic candidate score and the behavior candidate score to obtain a fused candidate score for each candidate item; The second determining module 306 is used to determine the recommendation result corresponding to the target query based on the fusion candidate score.

[0104] Optionally, the time feature includes a time interval feature and a relative time position feature; The first determining module 303 is specifically used for: Based on the time information in the historical interaction sequence, the time interval feature and relative time position feature corresponding to each historical item are determined respectively, wherein the time interval feature corresponding to the j-th historical item is... for: ; ; ; The relative time location feature of the j-th historical item for: ; in, Let j be the time interval corresponding to the j-th historical item. Let j be the interaction time corresponding to the j-th historical item. Let j-1 be the interaction time corresponding to the historical item. The quantity of the historical items, For a fixed time scale, and These are all preset constants used to prevent division by zero. Used to determine the maximum value among multiple values ​​within parentheses.

[0105] Optionally, the semantic scoring module 302 includes: The first mapping processing unit is used to map the text semantic embedding corresponding to the candidate item through the candidate side item mapping function to obtain the candidate item representation corresponding to the candidate item; The second mapping processing unit is used to map the text semantic embedding corresponding to the historical item through the historical item mapping function to obtain the historical item representation corresponding to the historical item. The processing unit is used to encode and read out multiple historical item representations arranged according to the historical interaction sequence to obtain the sequence representation corresponding to the historical interaction sequence. The first determining unit is used to determine the semantic candidate score corresponding to each candidate item based on the sequence representation and the representation of each candidate item.

[0106] Optionally, the behavior-enhanced semantic recommendation device 300 for zero-sample cross-domain sequence recommendation further includes: The third determining module is used to determine the popularity attribute and co-occurrence attribute of the source domain item based on the source domain interaction data. The popularity attribute is used to characterize the relative popularity of the source domain item among the multiple source domain items. The co-occurrence attribute is used to characterize the association features formed between the source domain item and other source domain items based on user interaction behavior. The input module is used to input the text semantic embedding corresponding to the source domain item into the popularity mapper to obtain the popularity-related proxy corresponding to the source domain item, and to input the text semantic embedding corresponding to the source domain item into the co-occurrence mapper to obtain the co-occurrence information proxy corresponding to the source domain item. The training module is used to adjust the parameters of the popularity mapper based on the difference between the popularity-related agent and the popularity attribute to obtain a trained popularity mapper, and to adjust the parameters of the co-occurrence mapper based on the difference between the co-occurrence information agent and the co-occurrence attribute to obtain a trained co-occurrence mapper.

[0107] Optionally, determining the popularity attribute corresponding to the source domain item based on the source domain interaction data includes: The source domain interaction data is divided into multiple time periods according to a preset time period, and the number of interactions of the source domain items in each time period is counted. Based on the number of interactions and the preset attenuation coefficient, the attenuation status of the source domain item in each time period is determined; For source domain items with non-zero decay state, the average parallel rank is assigned, and then normalized after being divided by the target value to obtain the normalized rank value of the source domain item in each time period. The target value is the maximum average rank plus one. The popularity attribute corresponding to the source domain item is determined based on the normalized rank value; Among them, source domain items Corresponding popularity attribute for: ; ; in, This indicates the last time period mentioned. Indicates the number of time periods at the end of the aggregation. Source Domain Items In the The normalized rank value corresponding to each time period Source Domain Items In the The decay state over a time period, Source Domain Items In the Number of interactions per time period The preset attenuation coefficient is denoted as .

[0108] Optionally, determining the co-occurrence attribute of the source domain item based on the source domain interaction data includes: Construct a source domain user-item association matrix and an initial co-occurrence matrix based on the source domain interaction data; After removing self-connections from the initial co-occurrence matrix, the co-occurrence values ​​between different source domain items are subjected to logarithmic transformation and symmetric normalization to obtain a symmetric normalized matrix. The symmetric normalized matrix elements in Used to characterize source domain items and source domain items Co-occurrence weights between them; For a target source domain item, calculate the reweighting coefficient between the target source domain item and its co-occurring neighbors. The co-occurring neighbors are source domain items whose co-occurrence weight with the target source domain item meets a preset requirement. The target source domain item is any one of the source domain items. Based on the reweighting coefficient, the text semantic embeddings corresponding to all the co-occurring neighbors are weighted and aggregated to obtain the neighborhood embedding of the target source domain item; The co-occurrence attribute of each source domain item is determined based on the neighborhood embedding; Among them, source domain items Corresponding co-occurrence attributes for: ; ; in, Represents a set of source domain items with non-zero neighborhood embeddings. for The mean of the middle neighborhood embedding, In order to be in The principal component transformation matrix fitted on top, Source Domain Items Corresponding co-occurring neighbors The reweighting coefficients, Indicates source domain items Corresponding text semantic embedding, Indicates co-occurring neighbors Corresponding text semantic embedding, This indicates the preset weight parameters. Indicates source domain items The mean of the cosine similarity with all source domain items. Used to convert input values ​​into modulation coefficients between 0 and 1.

[0109] Optionally, the behavior scoring module 304 includes: The construction unit is used to construct the historical features corresponding to the historical items based on the time features and the popularity-related proxies corresponding to the historical items; The first determining unit is used to determine the behavioral location input corresponding to the historical item based on the historical features and the co-occurrence information agent corresponding to the historical item; The second determining unit is used to perform sequence encoding on the behavior position input corresponding to the historical item according to the interaction order of the historical item in the historical interaction sequence, and determine the historical behavior representation corresponding to the historical interaction sequence based on the encoding state corresponding to the last valid historical position. The third determining unit is used to determine the candidate behavior representation corresponding to the candidate item based on the text semantic embedding corresponding to the candidate item and the co-occurrence information proxy corresponding to the candidate item; The fourth determining unit is used to determine the behavioral candidate score corresponding to each candidate item based on the degree of matching between the historical behavioral representation and the candidate behavioral representation.

[0110] The behavior-enhanced semantic recommendation device 300 for zero-sample cross-domain sequence recommendation provided in this application embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0111] It should be noted that the division of units in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods. Furthermore, the functional units in each embodiment of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0112] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method described in each embodiment of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] like Figure 4 As shown, this application embodiment provides an electronic device 400, including: a memory 402, a processor 401, and a program stored in the memory 402 and executable on the processor 401; the processor 401 is used to read the program in the memory 402 to implement the steps in the behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation as described above.

[0114] This application also provides a readable storage medium storing a program that, when executed by a processor, implements each process of the above-described embodiment of the behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation, and achieves the same technical effect. To avoid repetition, it will not be described again here. The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (such as compact disks (CDs), digital video discs (DVDs), Blu-ray discs (BDs), high-definition universal discs (HVDs), etc.), and semiconductor storage (such as read-only memory (ROMs), erasable programmable read-only memory (EPROMs), electrically erasable programmable read-only memory (EEPROMs), non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0115] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of this application.

[0117] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of this application without departing from its spirit, and all of these modifications are within the scope of protection of this application.

Claims

1. A behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation, characterized in that, include: Obtain a target query in the target domain. The target query includes a historical interaction sequence and multiple candidate items. The historical interaction sequence includes multiple historical items and interaction times that the target user interacted with before the current predicted position. The target domain includes multiple target domain items, multiple target domain users, and target domain interaction data between the target domain items and the target domain users. The historical items and candidate items both belong to the target domain items, and the target user belongs to the target domain users. Based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item, a semantic candidate score is determined for each candidate item through a semantic scoring model. The semantic scoring model is pre-trained based on data from the source domain, which includes multiple source domain items, multiple source domain users, and source domain interaction data between the source domain items and the source domain users. Based on the time information in the historical interaction sequence, the time characteristics corresponding to the historical item are determined; Based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embeddings corresponding to the candidate items, a behavior candidate score is determined for each candidate item through a behavior scoring model. The interaction attributes include popularity-related proxies and co-occurrence information proxies. The interaction attribute mapper includes a popularity mapper and a co-occurrence mapper. The popularity mapper is used to determine the popularity-related proxies based on text semantic embeddings, and the co-occurrence mapper is used to determine the co-occurrence information proxies based on text semantic embeddings. The popularity mapper and the co-occurrence mapper are pre-trained based on data from the source domain. The popularity-related proxies are used to represent the proxy results of the relative popularity of items learned from the source domain in the target domain, and the co-occurrence information proxies are used to represent the proxy results of the collaborative association features between items learned from the source domain in the target domain. The semantic candidate score and the behavioral candidate score are fused to obtain a fused candidate score for each candidate item. The recommended result corresponding to the target query is determined based on the fusion candidate score.

2. The behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation according to claim 1, characterized in that, The time features include time interval features and relative time position features; Based on the time information in the historical interaction sequence, the time characteristics corresponding to the historical item are determined, including: Based on the time information in the historical interaction sequence, the time interval feature and relative time position feature corresponding to each historical item are determined respectively, wherein the time interval feature corresponding to the j-th historical item is... for: ； ； ； The relative time location feature of the j-th historical item for: ； in, Let j be the time interval corresponding to the j-th historical item. Let j be the interaction time corresponding to the j-th historical item. Let j-1 be the interaction time corresponding to the historical item. The quantity of the historical items, For a fixed time scale, and These are all preset constants used to prevent division by zero. Used to determine the maximum value among multiple values ​​within parentheses.

3. The behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation according to claim 1, characterized in that, The step of determining the semantic candidate score for each candidate item using a semantic scoring model based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item includes: The candidate item representation is obtained by mapping the text semantic embedding corresponding to the candidate item through the candidate item mapping function; The text semantic embedding corresponding to the historical item is mapped by the historical item mapping function to obtain the historical item representation corresponding to the historical item. Encode and read out multiple historical item representations arranged according to the historical interaction sequence to obtain the sequence representation corresponding to the historical interaction sequence; The semantic candidate score for each candidate item is determined based on the sequence representation and the representation of each candidate item.

4. The behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation according to claim 1, characterized in that, Before determining the behavioral candidate score for each candidate item using a behavior scoring model based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embedding corresponding to the candidate items, the method further includes: Based on the source domain interaction data, the popularity attribute and co-occurrence attribute of the source domain item are determined. The popularity attribute is used to characterize the relative popularity of the source domain item among the multiple source domain items. The co-occurrence attribute is used to characterize the association features between the source domain item and other source domain items based on user interaction behavior. The text semantic embedding corresponding to the source domain item is input into the popularity mapper to obtain the popularity-related proxy corresponding to the source domain item. The text semantic embedding corresponding to the source domain item is input into the co-occurrence mapper to obtain the co-occurrence information proxy corresponding to the source domain item. The parameters of the popularity mapper are adjusted based on the difference between the popularity-related agent and the popularity attribute to obtain a trained popularity mapper. Similarly, the parameters of the co-occurrence mapper are adjusted based on the difference between the co-occurrence information agent and the co-occurrence attribute to obtain a trained co-occurrence mapper.

5. The behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation according to claim 4, characterized in that, Determining the popularity attribute of the source domain item based on the source domain interaction data includes: The source domain interaction data is divided into multiple time periods according to a preset time period, and the number of interactions of the source domain items in each time period is counted. Based on the number of interactions and the preset attenuation coefficient, the attenuation status of the source domain item in each time period is determined; For source domain items with non-zero decay state, the average parallel rank is assigned, and then normalized after being divided by the target value to obtain the normalized rank value of the source domain item in each time period. The target value is the maximum average rank plus one. The popularity attribute corresponding to the source domain item is determined based on the normalized rank value; Among them, source domain items Corresponding popularity attribute for: ； ； in, This indicates the last time period mentioned. Indicates the number of time periods at the end of the aggregation. Source Domain Items In the The normalized rank value corresponding to each time period Source Domain Items In the The decay state over a time period, Source Domain Items In the Number of interactions per time period The preset attenuation coefficient is denoted as .

6. The behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation according to claim 4, characterized in that, Determining the co-occurrence attributes of the source domain items based on the source domain interaction data includes: Construct a source domain user-item association matrix and an initial co-occurrence matrix based on the source domain interaction data; After removing self-connections from the initial co-occurrence matrix, the co-occurrence values ​​between different source domain items are subjected to logarithmic transformation and symmetric normalization to obtain a symmetric normalized matrix. The symmetric normalized matrix elements in Used to characterize source domain items and source domain items Co-occurrence weights between them; For a target source domain item, calculate the reweighting coefficient between the target source domain item and its co-occurring neighbors. The co-occurring neighbors are source domain items whose co-occurrence weight with the target source domain item meets a preset requirement. The target source domain item is any one of the source domain items. Based on the reweighting coefficient, the text semantic embeddings corresponding to all the co-occurring neighbors are weighted and aggregated to obtain the neighborhood embedding of the target source domain item; The co-occurrence attribute of each source domain item is determined based on the neighborhood embedding; Among them, source domain items Corresponding co-occurrence attributes for: ； ； in, Represents a set of source domain items with non-zero neighborhood embeddings. for The mean of the middle neighborhood embedding, In order to be in The principal component transformation matrix fitted on top, Source Domain Items Corresponding co-occurring neighbors The reweighting coefficients, Indicates source domain items Corresponding text semantic embedding, Indicates co-occurring neighbors Corresponding text semantic embedding, This indicates the preset weight parameters. Indicates source domain items The mean of the cosine similarity with all source domain items. Used to convert input values ​​into modulation coefficients between 0 and 1.

7. The behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation according to claim 1, characterized in that, The step of determining the behavioral candidate score for each candidate item based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embedding corresponding to the candidate items, through a behavior scoring model, includes: Based on the time features and the popularity-related proxies corresponding to the historical items, construct the historical features corresponding to the historical items; Based on the historical features and the co-occurrence information agent corresponding to the historical items, determine the behavioral location input corresponding to the historical items; According to the interaction order of the historical items in the historical interaction sequence, the behavior position input corresponding to the historical item is sequence encoded, and the historical behavior representation corresponding to the historical interaction sequence is determined based on the encoding state corresponding to the last valid historical position. Based on the text semantic embedding corresponding to the candidate item and the co-occurrence information proxy corresponding to the candidate item, the candidate behavior representation corresponding to the candidate item is determined. Based on the degree of matching between the historical behavior representation and the candidate behavior representation, a behavior candidate score is determined for each candidate item.

8. A behavior-enhanced semantic recommendation device for zero-sample cross-domain sequence recommendation, characterized in that, include: The acquisition module is used to acquire a target query in the target domain. The target query includes a historical interaction sequence and multiple candidate items. The historical interaction sequence includes multiple historical items and interaction times that the target user interacted with before the current predicted position. The target domain includes multiple target domain items, multiple target domain users, and target domain interaction data between the target domain items and the target domain users. The historical items and the candidate items all belong to the target domain items, and the target user belongs to the target domain users. The semantic scoring module is used to determine the semantic candidate score of each candidate item based on the historical interaction sequence, the text semantic embedding corresponding to the historical item, and the text semantic embedding corresponding to the candidate item, through a semantic scoring model. The semantic scoring model is pre-trained based on data from the source domain, and the source domain includes multiple source domain items, multiple source domain users, and source domain interaction data between the source domain items and the source domain users. The first determining module is used to determine the time characteristics corresponding to the historical item based on the time information in the historical interaction sequence; The behavior scoring module is used to determine the behavior candidate score for each candidate item based on the time features, the interaction attributes corresponding to the historical items, the interaction attributes corresponding to the candidate items, and the text semantic embeddings corresponding to the candidate items, through a behavior scoring model. The interaction attributes include popularity-related proxies and co-occurrence information proxies. The interaction attribute mapper includes a popularity mapper and a co-occurrence mapper. The popularity mapper is used to determine the popularity-related proxies based on the text semantic embeddings, and the co-occurrence mapper is used to determine the co-occurrence information proxies based on the text semantic embeddings. The popularity mapper and the co-occurrence mapper are pre-trained based on data from the source domain. The popularity-related proxies are used to represent the proxy results of the relative popularity of items learned from the source domain in the target domain, and the co-occurrence information proxies are used to represent the proxy results of the collaborative association features between items learned from the source domain in the target domain. The fusion module is used to fuse the semantic candidate score and the behavior candidate score to obtain a fused candidate score for each candidate item; The second determining module is used to determine the recommendation result corresponding to the target query based on the fusion candidate score.

9. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor, characterized in that the processor is configured to read the program from the memory to implement the steps in the behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation as described in any one of claims 1 to 7.

10. A readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the steps in the behavior-enhanced semantic recommendation method for zero-sample cross-domain sequence recommendation as described in any one of claims 1 to 7.

Images