Vehicle-mounted personalized recommendation sorting system based on large language model

Abstract

The invention discloses a vehicle-mounted personalized recommendation sorting system based on a large language model, and relates to the field of large language models. Comprising the following steps: generating material keywords based on a large language model through classification labels, description and abstract contents of offline data, and constructing a material attribute index table; constructing user behavior characteristics through the desensitized historical behavior data of the user in the vehicle-mounted environment; outputting a user feature sequence through an adaptive feature extraction module by combining the obtained user behavior features with spatial position information; the recommendation materials and the cold start materials of the user recall layer are combined with the material attribute index table, screening is conducted through the selection module, and a candidate set sequence is generated. The zero-sample recommendation sorting capacity of a large language model is fully mined, the user is helped to rapidly obtain needed content in the driving environment, and the user experience is improved. Therefore, the vehicle-mounted information entertainment experience is obviously optimized, and the satisfaction degree of the user and the intelligent degree of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of large language models, and in particular to an in-vehicle personalized recommendation ranking system based on a large language model. Background Art

[0002] With the rapid development of Internet of Vehicles technology, in-vehicle infotainment systems have become an important platform for improving driving experience and providing personalized services. In the in-vehicle scenario, personalized recommendation sorting systems can provide drivers and passengers with music, radio, information and other content that suits their preferences, helping users quickly obtain the required information in an environment with information overload. However, traditional recommendation systems face unique challenges in the in-vehicle environment.

[0003] Traditional recommendation methods, such as collaborative filtering (NCF), graph collaborative filtering (NGCF), and SASRec based on sequence modeling, mainly rely on users' historical behavior data for modeling. These methods have two major limitations: first, it is difficult to fully capture users' complex interests and intentions based on historical behavior alone, especially in car scenarios, where users' needs may change dynamically due to factors such as geographic location and driving situation; second, these models usually lack the ability to process background knowledge and common sense reasoning, which limits the recommendation performance and makes it difficult to provide high-quality recommendations that meet the user's current situation.

[0004] In order to improve the performance and interactivity of recommendations, researchers have begun to explore the introduction of pre-trained language models (PLMs) into recommendation systems. PLMs have the ability to mine user information and can model user preferences more deeply. However, existing methods usually require a lot of fine-tuning on specific task data, which limits the adaptability of the model in diverse recommendation tasks and increases computational overhead. Large language models (LLMs) such as GPT-4 demonstrate powerful natural language understanding and reasoning capabilities and have been tried as rankers for recommendation systems. By narrowing down the candidate set and then ranking it, it shows the potential for zero-shot ranking, even surpassing traditional models.

[0005] Despite initial progress, several challenges remain. Standard LLMs are mainly trained based on general world knowledge and cannot effectively capture the specific user behavior patterns and nuances of the content to be sorted in the vehicle scenario. In addition, natural language prompts make it difficult to realize the potential of LLMs, affecting the reliability of recommendation sorting.

[0006] In view of the above problems, the present invention proposes an in-vehicle personalized recommendation ranking system based on a large language model. Summary of the invention

[0007] The purpose of the present invention is to solve the shortcomings of the prior art and to propose an in-vehicle personalized recommendation ranking system based on a large language model.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] An in-vehicle personalized recommendation ranking system based on a large language model includes the following steps:

[0010] S1: Through the classification labels, descriptions, summaries and other contents of offline data (such as albums, broadcasts, etc.), and based on the large language model, material keywords are generated to build a material attribute index table; this index table can provide a structured description of the material and provide a basis for subsequent material screening and recommendation sorting.

[0011] S2: Construct user behavior characteristics through desensitized historical behavior data of users in the vehicle environment;

[0012] S3: Outputting the user feature sequence through the adaptive feature extraction module by combining the obtained user behavior features with the spatial location information;

[0013] S4: Combine the recommended materials and cold-start materials in the user recall layer with the material attribute index table, and filter them through the selection module to generate a candidate set sequence. This candidate set sequence covers materials that meet the user's potential needs and current context, ensuring the comprehensiveness and diversity of the recommendations;

[0014] S5: Construct structured prompt words through user behavior sequences and material candidate set sequences, so that the model can understand the user's intention more accurately and enhance the relevance of the reasoning results to the user's current situation;

[0015] S6: Generate material ranking results for different users through the reasoning ability of LLMs;

[0016] S7: Regularly update the user material ranking list through continuous user interaction and feedback data. Improve the intelligence level of the recommendation system and user satisfaction in the long run.

[0017] In the step S2, the historical behavior data includes user search terms, recent favorites and listening history.

[0018] Preferably: in the step S3, the working logic of the adaptive feature extraction module is:

[0019] S31: First, preprocess the user behavior data, especially the user search terms. Since these data may contain a lot of noise (such as irrelevant words, repeated words, spelling errors, stop words, etc.), data cleaning must be performed to improve the quality and effectiveness of the data. After completing the data cleaning, associate the attributes in the material index table (classification labels, descriptions, summaries, etc.) with the user's historical behavior data to generate a list of attributes of the user's recent behavior;

[0020] S32: In the scenario of in-vehicle recommendation, text importance evaluation is performed based on TF-IDF to screen out important keywords. The algorithm is as follows:

[0021]

[0022] Among them, TF represents word frequency, t represents a certain word, and document refers to the user's behavior record (such as search terms, historical listening materials).

[0023]

[0024] Among them, IDF is the inverse document frequency, N is the total number of documents (that is, the total number of user interaction behaviors), TF-IDF ( t ) =TF P t ) ×IDF ( t ) ;

[0025] S33: For users with few interactive behaviors, fill the keyword sequence through the supplement mechanism;

[0026] S34: Adaptively generate a user feature sequence according to the time sequence of the user's behavior.

[0027] Preferably: in the step S33, the supplement mechanism uses the popular material content of the entire platform as a part of the user characteristics or combines the user's spatial location information to provide local popular or time-sensitive material content.

[0028] Preferably: in the step S4, the recommended material is a material based on historical preference.

[0029] Preferably: in the step S4, the cold start material is newly launched content or insufficiently exposed material.

[0030] Preferably: the step S4 comprises the following steps:

[0031] S41: In order to solve the problem of long token limitation of large models, a selection module is used to generate a smaller candidate pool for further processing by LLMs. The selection module mainly performs weighted fusion on the materials in the recall layer and the cold start materials. First, the materials in the recall layer of each user are obtained. The materials in the recall layer have a weight information, which indicates the score of relevance. The materials with larger weights in the recall layer are more likely to be selected.

[0032] S42: weighted fusion of the cold start materials and the recalled materials;

[0033] S43: Associate the attributes (keywords, etc.) in the material index table with the materials, perform random selection based on the probability of each material, select some materials into the candidate set, and generate a candidate set sequence for each user in the order of user interaction time.

[0034] Preferably: in the step S42, the weighted fusion and selection algorithm is as follows:

[0035] For user u, the set of items in the recall layer is M recall ={m1,m2,…,m n}, each material m i There is a corresponding weight w i , its score is:

[0036] S recall (m i )=w i

[0037] Assume that the cold start material set is M cold ={m c1 ,m c2 ,…,m ck}, their selection probability can be expressed as:

[0038]

[0039] Where k is the number of cold start materials, p cold is an adjustable coefficient that controls the probability of cold start selection.

[0040] In the weighted fusion stage, the material weights of the recall layer and the selection probabilities of the cold start materials are fused.

[0041]

[0042] Normalized weighted score

[0043]

[0044] The beneficial effects of the present invention are:

[0045] 1. The present invention utilizes the desensitized user behavior information and spatial location information to construct an adaptive feature extraction module, a selection module and structured prompt words.

[0046] 2. The present invention fully exploits the ability of large language models to perform zero-sample recommendation and sorting, helping users to quickly obtain required content in a driving environment, thereby significantly optimizing the in-vehicle infotainment experience and improving user satisfaction and the intelligence of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1This is a logic diagram of an in-vehicle personalized recommendation sorting system based on a large language model proposed by the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is further described in detail below in conjunction with specific implementation methods.

[0049] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", and "set" should be understood in a broad sense, for example, they can be fixedly connected or set, or detachably connected or set, or integrally connected or set. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0050] Embodiment 1:

[0051] An in-vehicle personalized recommendation ranking system based on a large language model, such as Figure 1 As shown, it includes the following steps:

[0052] Step A: Obtain the classification tags, descriptions, summaries, and other contents of offline data (such as albums, broadcasts, and other materials) through MySQL, generate material keywords using large language models (LLMs), and build a structured material attribute index table.

[0053] Step A-1: ​​Write a Python program to obtain the classification label, material description, summary, etc. of valid material information (album, broadcast) through the MySQL query interface, and build a material attribute index object.

[0054] Step A-2: Write the corresponding prompt words and Python program, and generate up to 10 keywords for each material offline through the large model (LLMs). These generated keywords are stored persistently as a new attribute of the corresponding material. Finally, they form a material index table with the existing attribute information.

[0055] Step B: Obtain user behavior characteristics and spatial characteristics through the big data center platform. Preprocess the user behavior data and associate it with material attributes to clean up noise data, improve data quality, and generate user feature sequences for recent users through the adaptive feature extraction module.

[0056] Step B-1: Call the big data API interface, write a Python program, obtain all user behavior data within 10 days, and receive the data using a pandas object. Due to the huge amount of user data, query and store in batches to improve access efficiency. User behavior data involves multiple data reporting events, and each event is queried and processed one by one.

[0057] Step B-2: Associate the material data queried by the user with the material index table, index the material attributes by ID, and obtain the material classification label, description, summary and other information. Then clean the user behavior data (such as user search data), remove stop words from the data through the stop word list of Harbin Institute of Technology, and remove noise data such as irrelevant words and repeated words to obtain the original attributes of user behavior characteristics.

[0058] Step B-3: Adaptive feature extraction is then performed for different users through the adaptive feature extraction module. On the one hand, for users with a large number of behaviors, the TF-IDF algorithm is used to evaluate the importance of the text to filter out important keywords. On the other hand, for users with few interactive behaviors, popular materials on the entire platform and popular or timely materials in the geographic location are used to fill the gaps in the low interactive behaviors.

[0059] Step B-4: Generate user feature sequences in sequence according to the time sequence of user behaviors, and store the user feature sequences in batches.

[0060] Step C: Use the selection module to generate a smaller candidate pool through the MySQL cold-start material table and the Redis user recall layer materials. And associate the material ID with the material index table to finally generate a candidate set sequence that meets the user's potential needs and current situation. The whole process needs to solve the problem of long token limitations of large models and maintain the comprehensiveness and diversity of recommendations.

[0061] Step C-1: Write a python program to connect to the MySQL database, query the cold start material table, and obtain the material ID list and its related attributes. In addition, the material data of the user recall layer is stored in Redis. By connecting to the Redis database, the materials and corresponding weights of the user recall layer are obtained.

[0062] Step C-2: Write a weighted fusion selection algorithm for the selection module. The selection module first calculates the selection probability of the cold start material and uses an adjustable coefficient to control its influence. Secondly, the weight of the recall layer material is weighted and fused with the selection probability of the cold start material. Finally, the fused score is normalized to obtain the probability of each material being selected. According to the normalized probability, a certain number of materials are randomly selected from the material pool to generate a candidate set.

[0063] Step C-3: Use the material index table in step A to associate with the materials in the candidate set, and use the material attributes (such as keywords) to supplement to ensure the comprehensiveness and diversity of the recommendation. Finally, according to the user's interaction time sequence, generate a candidate set sequence that meets the user. Similar to the user behavior sequence in step B, the user candidate set sequence is stored in batches.

[0064] Step D: Based on the user behavior sequence in step B and the candidate set sequence in step C, generate structured prompt words so that the large language model (LLMs) can better understand the relationship between user behavior and materials.

[0065] Step D-1: Build a structured prompt word template, define Profile, Role, skills, Rules, Workflow, Initialization in sequence, and use markdown syntax to structure the prompt words. Structured prompt words are highly readable, can improve semantic cognition, can be directed to awaken the deep capabilities of large models, and are highly reusable.

[0066] Step D-2: Read the user feature sequence and the candidate set sequence in batches, modify the Initialization part in the template through the structured prompt word template, and generate user-specific prompt word templates in sequence.

[0067] Step E: Based on the prompt word template in step D, input it into the large language model through the large language model API interface, and output the ranking of the candidate set in sequence according to the rules defined in the template. Finally, the sorted material list is stored in redis to facilitate the backend to obtain the data of the corresponding user sorting stage.

[0068] Step E-1: Write a Python program, call the large language model API interface, input the user-specific prompt word templates in step D into the large language model in sequence, and output the candidate set material ranking in sequence according to the rules in the template. Among them, the large language model selected models such as chatgpt-4o and Alibaba Cloud Tongyi Qianwen-MAX.

[0069] Step E-2: Look up the ID of the sorted material list through the material attribute index table, and store the ID list and user ID in the redis database. The storage format in Redis uses the user ID as the key and the material ID list as the value.

[0070] Step E-3: Regularly update the user material ranking list through continuous user interaction and feedback data.

[0071] During the update phase, the user feature sequence and candidate set sequence will also change dynamically. In addition, the rules in the structured prompt word template will also change accordingly, and the current recommendation time will also be considered. Regular update tasks are completed by setting timed tasks.

[0072] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An in-vehicle personalized recommendation ranking system based on a large language model, characterized in that: The following steps are involved: S1: Generate material keywords based on the classification labels, descriptions, and summary content of offline data and the large language model to build a material attribute index table; S2: Construct user behavior characteristics through desensitized historical behavior data of users in the vehicle environment; S3: Outputting the user feature sequence through the adaptive feature extraction module by combining the obtained user behavior features with the spatial location information; S4: Combine the recommended materials and cold start materials in the user recall layer with the material attribute index table, filter them through the selection module, and generate a candidate set sequence; S5: Construct structured prompt words through user behavior sequence and material candidate set sequence; S6: Generate material ranking results for different users through the reasoning ability of LLMs; S7: Regularly update the user material sorting list through continuous user interaction and feedback data.

2. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 1, characterized in that: In the step S2, the historical behavior data includes user search terms, recent favorites and listening history.

3. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 1, characterized in that: In the step S3, the working logic of the adaptive feature extraction module is: S31: First, the user behavior data is cleaned. After the data cleaning is completed, the attributes in the material index table are associated with the user's historical behavior data to generate an attribute list of the user's recent behavior; S32: In the scenario of in-vehicle recommendation, text importance evaluation is performed based on TF-IDF to screen out important keywords; S33: For users with only a few interactive behaviors, fill in the keyword sequence through a supplementary mechanism. S34: Adaptively generate a user feature sequence according to the time sequence of the user's behavior.

4. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 3, characterized in that: In the step S32, the algorithm model is as follows: Among them, TF represents word frequency, t represents a certain word, and document refers to the user's behavior record; Where IDF is the inverse document frequency, N is the total number of documents TF-IDF ( t ) =TF ( t ) ×IDF ( t ) .

5. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 4, characterized in that: In the step S33, the supplement mechanism uses the popular material content of the entire platform as part of the user characteristics or combines the user's spatial location information to provide local popular or time-sensitive material content.

6. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 1, characterized in that: In the step S4, the recommended materials are materials based on historical preferences.

7. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 6, characterized in that: In the step S4, the cold start material refers to newly launched content or material that has not been fully exposed.

8. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 7, characterized in that: The step S4 comprises the following steps: S41: First, the materials in the recall layer of each user are obtained. The materials in the recall layer are all provided with a weight information, which indicates the score of relevance. The materials with larger weight in the recall layer are more likely to be selected. S42: weighted fusion of the cold start materials and the recalled materials; S43: Associating the attributes in the material index table with the materials, performing random selection based on the probability of each material, selecting some materials to enter the candidate set, and generating a candidate set sequence for each user in the order of user interaction time.

9. The in-vehicle personalized recommendation ranking system based on a large language model according to claim 8, characterized in that: In step S42, the weighted fusion and selection algorithm is as follows: For user u, the set of items in the recall layer is M recall ={m1,m2,…,m n }, each material m i There is a corresponding weight w i , its score is: S recall (m i )=w i Assume that the cold start material set is M cold = { m c1 ,m c2 ,…,m ck} , their selection probability can be expressed as: Where k is the number of cold start materials, p cold is an adjustable coefficient that controls the probability of cold start selection. In the weighted fusion stage, the material weights of the recall layer and the selection probabilities of the cold start materials are fused. Normalized weighted score,

Images