Article recommendation system sorting method based on large language model Zero-Shot

Abstract

The invention discloses a big language model Zero-Shot-based article recommendation system sorting method, which comprises the following steps of: constructing a user historical sequence interaction article set and a sequence candidate article set, constructing a big language model Zero-Shot sorting instruction template, generating corresponding cue words of a big language model, and sorting the cue words of the big language model according to the user historical sequence interaction article set and the sequence candidate article set. The big language model extracts article preferences of the user through historical interaction sequence information of the user, sorts a sequence candidate article set, extracts a sorting result Su corresponding to the user from an output result of the big language model, verifies the integrity of the sorting result Su, and sends the sorting result Su to the user; using an attention mechanism to sort the embedding vectors of the user with the context and the item candidate set to obtain a sorting result Au; and averaging the sorting result corresponding to each article of the sorting result Su and the sorting result Au, and then sorting to obtain a recommendation result of the user. According to the method, semantic information of the large language model is increased, and recommendation can be carried out more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of item recommendation, and in particular to an item recommendation system ranking method based on a large language model Zero-Shot. Background Art

[0002] A large amount of information and items are generated on the Internet every day. How to better push this information and items to users who care about them is a hot issue that needs to be solved urgently. As a technology that matches information and items with users, the recommendation system extracts possible information and items from users through historical interaction information, such as information and items that users have browsed, collected, and shared, and analyzes and extracts user intentions, thereby recommending information and items from the recommendation platform to users. Therefore, by designing effective recommendation technology, the recommendation platform can use more accurate recommendation technology to recommend information and items that users care about, which can effectively improve the utilization and value of information and items;

[0003] Among the existing recommendation system ranking technologies based on traditional deep learning, most are built on collaborative filtering and its derivative technologies. This type of collaborative filtering technology uses the historical interaction information of users to obtain more accurate embedding vectors of users and items, thereby predicting the possibility of interaction between users and items for ranking, and plays an important role in the current recommendation system technology. However, this type of technology does not mine text information. Collaborative filtering technology only uses the interaction information between users and items, and lacks text semantic information between users and items. However, users generally make choices based on item semantic information when using recommendation platforms on a daily basis, so the collaborative filtering technology of existing recommendation systems has certain limitations. Summary of the invention

[0004] In order to overcome the defects and shortcomings of the prior art, the present invention provides an item recommendation system ranking method based on a large language model Zero-Shot. The present invention starts from the current pre-trained large language model, inputs item information and user information into the large language model for processing, increases the semantic information of the large language model, and can make recommendations more accurately.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides an item recommendation system ranking method based on a large language model Zero-Shot, comprising the following steps:

[0007] Obtain the user's historical interaction items to obtain a historical interaction item set, sort the historical interaction item set according to the user's interaction time, and obtain the user's historical sequential interaction item set;

[0008] Obtain the user's candidate item set, sort the candidate item set according to the appearance time, and obtain a sequential candidate item set;

[0009] Build a large language model Zero-Shot sorting instruction template, fill in the user's historical sequential interaction records and the information of the sequential candidate item set, and generate the corresponding large language model prompt words;

[0010] The prompt word is input into the large language model. The large language model extracts the user's item preferences through the user's historical interaction sequence information, and sorts the sequential candidate item set. The output result of the large language model extracts the user's corresponding sorting result S u ;

[0011] Verify the sorting result S u integrity;

[0012] Use the attention mechanism to sort the embedding vectors of the user and item candidate sets with context, and get the sorting result A u ;

[0013] Sorting results S u And sort result A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result, which is recommended as the user's recommendation result.

[0014] As a preferred technical solution, the historical interaction item set is sorted according to the user interaction time to obtain the user historical sequential interaction item set, which is specifically expressed as:

[0015]

[0016] Among them, T u represents the user interaction time, H u Represents a collection of historical interactive items, TimeSorted represents a time sorting function, Represents a collection of items that a user has interacted with in order.

[0017] As a preferred technical solution, the step of obtaining a user's candidate item set specifically includes:

[0018] Get the user set U, item set I and historical interaction information H output by the item recommendation system, and map each user u and item i to get the corresponding embedding vector e u , embedding vector e i ;

[0019] For the embedding vector e u and the embedding vector e i Calculate the inner product and get the preference score between user u and item i, expressed as:

[0020]

[0021] in, represents the preference score of user u for item i;

[0022] The optimization is performed through the Bayesian personalized loss function, which is expressed as:

[0023]

[0024] in, is the activation function, ln(x) is the logarithmic function with base e, H u represents the set of historical interaction items, j represents the items that the user has not interacted with, represents the preference score of user u for item j;

[0025] The optimization goal is:

[0026]

[0027] The optimal embedding vector is obtained after optimization and embedding vector The optimized user u's preference score for the item is calculated as:

[0028]

[0029] By sorting the preference scores, select the first n candidate items for user u as CI u .

[0030] As a preferred technical solution, the candidate item set is sorted according to the appearance time to obtain a sequential candidate item set, which is specifically expressed as:

[0031]

[0032] Among them, C u represents a set of sequential candidate items, TimeSorted represents a time sorting function, CI u represents the set of candidate items, Indicates the appearance time of the candidate item.

[0033] As an optimal technical solution, a large language model Zero-Shot sorting instruction template is constructed, which is specifically expressed as follows:

[0034] "The items that the user has interacted with include [the user's historical sequential interaction records]. Now there are n candidate items, which are [the sorted candidate item set]. Predict the most likely interaction result of the user through the historical sequential interaction information and sort these n items. Please output the numerical result of the sorting. Please note that items outside the candidate set should not be generated."

[0035] As a preferred technical solution, the large language model extracts the user's item preferences through the user's historical interaction sequence information, and sorts the sequence candidate item set, and extracts the user's corresponding sorting result S from the large language model output result u , specifically expressed as:

[0036] S u =Extract(P u )

[0037]

[0038] Among them, r u is the ranking result output by the large language model, G is the large language model, T is the instruction template of the large language model, is the historical sequential interaction information corresponding to user u, C u is a set of sequential candidate items, S u It represents the sorting result after sequence text extraction, and Extract represents the text extraction function.

[0039] As a preferred technical solution, the verification sorting result S u Integrity, including:

[0040] Verify the integrity of the sorting results. If duplicate items are found, remove them and only keep the first items.

[0041] If the sorting result S is verified u If there is an item that does not belong to the set of sequential candidate items, delete the item;

[0042] If it is verified that there is no item in the sequence candidate item set, add it to the sorting result S u ;

[0043] By verifying the integrity of the sorting results, the sorting result S u Is equal to the elements of the set of sequential candidate items.

[0044] As a preferred technical solution, the attention mechanism is used to sort the embedding vectors of the user and item candidate sets with context, and the sorting result A is obtained. u , specifically including:

[0045] Using the large language model encoder, we extract the embedding vectors of the candidate sets of contextual users and items, expressed as:

[0046]

[0047] Among them, T is the large language model instruction template, is the historical sequential interaction information corresponding to user u, C u is the set of sequential candidate items, E u The information with historical sequential interactive items reflects the historical item interests of user u. Indicates the spatial location of the text representing the candidate item;

[0048] Through the attention mechanism, we can get the user U’s attention to the item C in the candidate item set. ui The score is:

[0049]

[0050] Where sum(Score)=1, d represents the dimension of the embedding vector;

[0051] Sort the candidate items according to the attention scores and get the sorting result A u :

[0052] A u =Sorted(Score)

[0053] Among them, Sorted represents the sorting function.

[0054] As a preferred technical solution, the sorting result S u And sort result A u The sorting results corresponding to each item are averaged and sorted to obtain the final sorting result, which is expressed as:

[0055] FR u =Stored_Avg(S u ,A u )

[0056] Among them, the Stored_Avg function indicates that for the two sorting results S u ,A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result FR u .

[0057] The present invention also provides an item recommendation system sorting system based on a large language model Zero-Shot, comprising: a user history order interaction item set construction module, a sequence candidate item set construction module, a sorting instruction template construction module, a sorting result Su Extraction module, sorting results S u Integrity verification module, sorting result A u Construction module, recommendation result output module;

[0058] The user historical sequential interaction item set construction module is used to obtain the user's historical interaction items to obtain a historical interaction item set, and sort the historical interaction item set according to the user's interaction time to obtain the user's historical sequential interaction item set;

[0059] The sequential candidate item set building module is used to obtain a user's candidate item set, sort the candidate item set according to the appearance time, and obtain a sequential candidate item set;

[0060] The sorting instruction template construction module is used to construct a large language model Zero-Shot sorting instruction template, fill in the user's historical sequence interaction record and the information of the sequence candidate item set, and generate the corresponding large language model prompt word;

[0061] The sorting result S u The extraction module is used to input the prompt words into the large language model. The large language model extracts the user's item preferences through the user's historical interaction sequence information, and sorts the sequential candidate item set. The output result of the large language model extracts the user's corresponding sorting result S u ;

[0062] The sorting result S u The integrity verification module is used to verify the sorting result S u integrity;

[0063] The sorting result A u The construction module is used to use the attention mechanism to sort the embedding vectors of the user and item candidate sets with context, and obtain the sorting result A u ;

[0064] The recommendation result output module is used to sort the result S u And sort result A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result, which is recommended as the user's recommendation result.

[0065] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0066] (1) The present invention uses Zero-Shot and instruction template methods to enable the large language model to complete the corresponding semantic analysis capabilities without fine-tuning the large language model. That is, it extracts the user's long-term and short-term interest preferences for the sequential items that the user has interacted with, increases the semantic information of the large language model, and enables more accurate recommendations.

[0067] (2) The present invention uses a large language model encoder to obtain contextual embedding vectors of users and candidate item sets, and uses the Attention mechanism to sort the items to obtain sorting results with text embedding information.

[0068] (3) The present invention combines the sorting order obtained by the large language model and the context embedding vector obtained by the large language encoder in the sorting result of the attention mechanism to obtain the final sorting result. The result takes into account the semantic understanding and contextual text information of the large language model, thereby enabling more accurate recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of the method for ranking item recommendation system based on large language model Zero-Shot of the present invention;

[0070] Figure 2 A schematic diagram of the training process of collecting user candidate items in the present invention;

[0071] Figure 3 Schematic diagram of the input prompt word template of the large language model of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0073] Example 1

[0074] like Figure 1 As shown, this embodiment provides an item recommendation system ranking method based on a large language model Zero-Shot, comprising the following steps:

[0075] S1: Processing of user historical interaction items;

[0076] Processing the user's historical interaction items, specifically including: collecting the user's historical interaction items to obtain a historical interaction item set, and sorting them according to the user's historical interaction order;

[0077] Step S1 processes the user's historical interactive items, which specifically includes the following sub-steps:

[0078] S1.1: Collect user historical interaction items;

[0079] Collect the historical interaction items of user u to obtain the user's historical interaction item set

[0080] S1.2: Sort user historical interactions by interaction time;

[0081] According to the user interaction time T u For the historical interactive item set H u Sort the items to get the user's historical order of interaction:

[0082]

[0083] in, TimeSorted represents the time sorting function;

[0084] S2: User candidate item set processing;

[0085] The user's candidate item set is processed, specifically including collecting the user's candidate item set through a recommendation system, and sorting the candidate item set according to the appearance time to obtain a sorted candidate item set.

[0086] Specifically, collect n candidate item sets CI of user u u And the corresponding appearance time The time sorting by appearance time is:

[0087]

[0088] Get the sequential candidate item set C u .

[0089] like Figure 2 As shown in the figure, when training the model to collect user candidate item sets, for a given item recommendation system, there are user set U, item set I and historical interaction information H, and the corresponding embedding vector is mapped to each user u and item i. Where d is the dimension of the embedding vector. Through the inner product, we can get the preference score between user u and item i:

[0090]

[0091] in, It represents the preference score of user u for item i. The larger the score, the higher the preference, and the smaller the score, the lower the preference.

[0092] Optimization is performed through the Bayesian personalized loss function:

[0093]

[0094] in, is the activation function, ln(x) is the logarithmic function with base e. There is an optimization goal:

[0095]

[0096] After optimization, the optimal embedding vector expression is The optimized user u's preference score for the item is calculated as By sorting the preference scores, select the first n candidate items for user u as CI u .

[0097] S3: Build a large language model Zero-Shot sorting instruction template;

[0098] By designing a Zero-Shot sorting template for the large language model, the user's historical interaction item information and the user's candidate item information are filled in the corresponding positions in sequence according to the instruction template, and the prompt words corresponding to the user's item sorting recommendation large language model are constructed and passed to the large language model.

[0099] Specifically, by designing the instruction template of the large language model, data can be processed in batches efficiently. And the sorted candidate item set C u ,By embedding relevant information into the instruction template, the large language model makes ,judgment on it by processing the semantic information.

[0100] like Figure 3 As shown, the instruction template of the large language model is represented as:

[0101] "The items that the user has interacted with are [the user's historical sequential interaction records]. Now there are n candidate items, which are [the sorted candidate item set]. Predict the most likely interaction result of the user based on the historical sequential interaction information and sort these n items. Please output the sorted numerical result. Please note that items outside the candidate set should not be generated."

[0102] For each user u, the template is used to fill in the information of "[user's historical sequential interaction record]" and "[sorted candidate item set]" to generate the prompt words of the large language model corresponding to the user u.

[0103] S4: Use the large language model to output the ranking results;

[0104] In this embodiment, the prompt words are input into the large language model, which understands the user's item preferences through the user's historical interaction sequence information, predicts the user's possible preference for the candidate set items, sorts the items, and outputs the sorting results.

[0105] In a recommendation scenario, time is critical information when extracting user historical interaction sequence information. For example, if a user buys books for a long time, bought snacks a long time ago, and bought clothes in the short term, then the recommendation system should consider that books and clothes are what the user considers in the long term and in the short term. That is, the item sequence information contains the user's long-term and short-term interest information. If the sequence information is missing, the recommendation accuracy will be reduced.

[0106] Specifically, the prompt word is input into the large language model, which extracts the item preferences of user u through the user's historical interaction sequence information, and sorts the sequential candidate item set as follows:

[0107]

[0108] Among them, R u is the ranking result output by the large language model, G is the large language model, T is the instruction template of the large language model, is the historical sequential interaction information corresponding to user u, C u is a set of sequential candidate items.

[0109] By outputting the result R to the large language model u Perform sequence text extraction and extract the sorting results corresponding to user u, which are:

[0110] S u =Extract(R u )

[0111] Among them, S u It represents the sorting result after sequence text extraction, and Extract represents the text extraction function;

[0112] S5: Verify the sorting result S u Integrity;

[0113] Verify the integrity of the sorting results. Since the large language model has hallucinations and will output results that are not in the candidate set, it is necessary to verify the integrity of the sorting results. If duplicate items are found, they need to be excluded and only one item is left. If the content in the original candidate set does not exist, it is excluded. If the items that are not selected are verified, they are randomly selected to supplement the sorting results. Finally, the sorting results are output to the user.

[0114] Specifically, the obtained sorting result S uVerify item integrity. u There are three main situations for matching items in:

[0115] If duplicate items are found, remove them and only keep the first items;

[0116] Found the sorting result S u There are items that do not belong to the sequential candidate set C u Items in the box are directly removed;

[0117] Discover the set of candidate items in order C u The item does not exist, and is added directly to the sorting result S u ;

[0118] Finally, the sorting result S is guaranteed u and the sequential candidate item set C u The elements of are equal.

[0119] S6: Use the attention mechanism to sort the embedding vectors of the user and item candidate sets with context;

[0120] Using the large language model encoder, the embedding vectors of the candidate sets of contextual users and items are extracted:

[0121]

[0122] Among them, T is the large language model instruction template, is the historical sequential interaction information corresponding to user u, C u is a set of sequential candidate items, Where d represents the dimension of the embedding vector, E u The information with historical sequential interactive items reflects the historical item interests of user u. Indicates the spatial position of the candidate item text.

[0123] Through the attention mechanism, we can get the user U’s attention to the items in the candidate item set. The scores are:

[0124]

[0125] in, sum(Score)=1, Score i Indicates the attention score of candidate item i. The candidate items are sorted according to the attention score:

[0126] A u =Sorted(Score)

[0127] Among them, Au is the final sorting result, representing the user u with contextual order item information and the candidate item set C with text information u Attention score, Sorted represents the sorting function;

[0128] S7: Integrate the two ranking results as the user's recommendation result;

[0129] Sorting results S u And the ranking result A obtained by the large language model encoder and attention score u The sorting is:

[0130] FR u =Stored_Avg(S u ,A u )

[0131] Among them, the Stored_Avg function indicates that for the two sorting results S u ,A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result FR u , the final sorting result FR u Recommended as the recommendation result for user u.

[0132] Example 2

[0133] This embodiment provides an item recommendation system sorting system based on a large language model Zero-Shot, which is used to implement the item recommendation system sorting method based on a large language model Zero-Shot in the above embodiment 1. The system includes: a user history sequence interaction item set construction module, a sequence candidate item set construction module, a sorting instruction template construction module, and a sorting result S u Extraction module, sorting results S u Integrity verification module, sorting result A u Construction module, recommendation result output module;

[0134] In this embodiment, the user historical sequential interaction item set construction module is used to obtain the user's historical interaction items to obtain a historical interaction item set, and sort the historical interaction item set according to the user's interaction time to obtain the user's historical sequential interaction item set;

[0135] In this embodiment, the sequential candidate item set construction module is used to obtain the user's candidate item set, sort the candidate item set according to the appearance time, and obtain the sequential candidate item set;

[0136] In this embodiment, the sorting instruction template construction module is used to construct a large language model Zero-Shot sorting instruction template, fill in the user's historical sequence interaction records and the information of the sequence candidate item set, and generate the corresponding large language model prompt words;

[0137] In this embodiment, the sorting result S u The extraction module is used to input the prompt words into the large language model. The large language model extracts the user's item preferences through the user's historical interaction sequence information, and sorts the sequential candidate item set. The output result of the large language model extracts the user's corresponding sorting result S u ;

[0138] In this embodiment, the sorting result S u The integrity verification module is used to verify the sorting result S u integrity;

[0139] In this embodiment, the sorting result A u The construction module is used to use the attention mechanism to sort the embedding vectors of the user and item candidate sets with context, and obtain the sorting result A u ;

[0140] In this embodiment, the recommendation result output module is used to sort the result S u And sort result A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result, which is recommended as the user's recommendation result.

[0141] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for ranking item recommendation systems based on a large language model Zero-Shot, characterized in that: The steps include: Obtain the user's historical interaction items to obtain a historical interaction item set, sort the historical interaction item set according to the user's interaction time, and obtain the user's historical sequential interaction item set; Obtain the user's candidate item set, sort the candidate item set according to the appearance time, and obtain a sequential candidate item set; Build a large language model Zero-Shot sorting instruction template, fill in the user's historical sequential interaction records and the information of the sequential candidate item set, and generate the corresponding large language model prompt words; The prompt word is input into the large language model. The large language model extracts the user's item preferences through the user's historical interaction sequence information, and sorts the sequential candidate item set. The output result of the large language model extracts the user's corresponding sorting result S u ; Verify the sorting result S u integrity; Use the attention mechanism to sort the embedding vectors of the user and item candidate sets with context, and get the sorting result A u ; Sorting results S u And sort result A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result, which is recommended as the user's recommendation result.

2. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: The historical interaction item set is sorted according to the user interaction time to obtain the user historical sequential interaction item set, which is specifically expressed as: Among them, T u represents the user interaction time, H u Represents a collection of historical interactive items, TimeSorted represents a time sorting function, Represents a collection of items that a user has interacted with in order.

3. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: The step of obtaining a user's candidate item set specifically includes: Get the user set U, item set I and historical interaction information H output by the item recommendation system, and map each user u and item i to get the corresponding embedding vector e u , embedding vector e i ; For the embedding vector e u and the embedding vector e i Calculate the inner product and get the preference score between user u and item i, expressed as: in, represents the preference score of user u for item i; The optimization is performed through the Bayesian personalized loss function, which is expressed as: in, is the activation function, ln(x) is the logarithmic function with base e, H u represents the set of historical interaction items, j represents the items that the user has not interacted with, represents the preference score of user u for item j; The optimization goal is: The optimal embedding vector is obtained after optimization and embedding vector The optimized user u's preference score for the item is calculated as: By sorting the preference scores, select the first n candidate items for user u as CI u .

4. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: The candidate item set is sorted according to the appearance time to obtain a sequential candidate item set, which is specifically expressed as: Among them, C u represents a set of sequential candidate items, TimeSorted represents a time sorting function, CI u represents the set of candidate items, Indicates the appearance time of the candidate item.

5. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: Construct a large language model Zero-Shot sorting instruction template, which is specifically expressed as follows: "The items that the user has interacted with include [the user's historical sequential interaction records]. Now there are n candidate items, which are [the sorted candidate item set]. Predict the most likely interaction result of the user through the historical sequential interaction information and sort these n items. Please output the numerical result of the sorting. Please note that items outside the candidate set should not be generated." 6. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: The large language model extracts the user's item preferences through the user's historical interaction sequence information, and sorts the sequential candidate item set. The large language model output results are used to extract the user's corresponding sorting result S u , specifically expressed as: S u =Extract(R u ) Among them, R u is the ranking result output by the large language model, G is the large language model, T is the instruction template of the large language model, is the historical sequential interaction information corresponding to user u, C u is a set of sequential candidate items, S u It represents the sorting result after sequence text extraction, and Extract represents the text extraction function.

7. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: The verification sorting result S u Integrity, including: Verify the integrity of the sorting results. If duplicate items are found, remove them and only keep the first items. If the sorting result S is verified u If there is an item that does not belong to the set of sequential candidate items, delete the item; If it is verified that there is no item in the sequence candidate item set, add it to the sorting result S u ; By verifying the integrity of the sorting results, the sorting result S u Is equal to the elements of the set of sequential candidate items.

8. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: Use the attention mechanism to sort the embedding vectors of the user and item candidate sets with context, and get the sorting result A u , specifically including: Using the large language model encoder, we extract the embedding vectors of the candidate sets of contextual users and items, expressed as: Among them, T is the large language model instruction template, is the historical sequential interaction information corresponding to user u, C u is the set of sequential candidate items, E u The information with historical sequential interactive items reflects the historical item interests of user u. Indicates the spatial location of the text representing the candidate item; Through the attention mechanism, we can get the user U’s attention to the items in the candidate item set. The score is: Where sum(Score)=1, d represents the dimension of the embedding vector; Sort the candidate items according to the attention scores and get the sorting result A u : A u =Sorted(Score) Among them, Sorted represents the sorting function.

9. The method for ranking item recommendation systems based on large language model Zero-Shot according to claim 1, characterized in that: Sorting results S u And sort result A u The sorting results corresponding to each item are averaged and sorted to obtain the final sorting result, which is expressed as: FR u =Stored_Avg(S u ,A u ) Among them, the Stored_Avg function indicates that for the two sorting results S u ,A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result FR u .

10. An item recommendation system ranking system based on a large language model Zero-Shot, characterized in that: include: User history sequential interaction item set building module, sequential candidate item set building module, sorting instruction template building module, sorting result S u Extraction module, sorting results S u Integrity verification module, sorting result A u Construction module, recommendation result output module; The user historical sequential interaction item set construction module is used to obtain the user's historical interaction items to obtain a historical interaction item set, and sort the historical interaction item set according to the user's interaction time to obtain the user's historical sequential interaction item set; The sequential candidate item set building module is used to obtain a user's candidate item set, sort the candidate item set according to the appearance time, and obtain a sequential candidate item set; The sorting instruction template construction module is used to construct a large language model Zero-Shot sorting instruction template, fill in the user's historical sequence interaction record and the information of the sequence candidate item set, and generate the corresponding large language model prompt word; The sorting result S u The extraction module is used to input the prompt words into the large language model. The large language model extracts the user's item preferences through the user's historical interaction sequence information, and sorts the sequential candidate item set. The output result of the large language model extracts the user's corresponding sorting result S u ; The sorting result S u The integrity verification module is used to verify the sorting result S u integrity; The sorting result A u The construction module is used to use the attention mechanism to sort the embedding vectors of the user and item candidate sets with context, and obtain the sorting result A u ; The recommendation result output module is used to sort the result S u And sort result A u The ranking results corresponding to each item are averaged and then sorted to obtain the final ranking result, which is recommended as the user's recommendation result.

Images