Recommendation system-oriented spatio-temporal data enhancement method

Abstract

The invention discloses a spatio-temporal data enhancement method oriented to a recommendation system. According to the method, the spatial and temporal characteristics of the user and the article are obviously summarized, a large model is used as an encoder, and embedding is unified in a single vector space; and then the feature dimensions are aligned by using a linear layer to be fused into a recommendation model, so that the problem of insufficient feature fusion is solved. According to the method, the interaction history and the candidate set of the user are understood by using a large model, the preference of the user is reasoned, positive and negative sample pairs are generated, and the generated sample set and the original sample set are combined as a final training set. Benefited from the excellent reasoning ability and natural language understanding ability of the large model, the training sample set is expanded, and the noise problem is relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of information recommendation, and specifically relates to a spatio-temporal data augmentation method for a recommendation system. Background Art

[0002] Spatio-temporal recommendation systems greatly expand the application boundary of recommendation systems by integrating users' geographical location and time information. Such systems not only consider users' historical behaviors and preferences, but also use a variety of technologies to process time features and space features.

[0003] To improve recommendation performance, existing work has made many efforts. Some traditional recommendation models improve recommendation performance by modeling the interaction relationship between users and items. They mainly rely on collaborative filtering technology, have limited processing ability for complex features, and cannot use spatio-temporal attributes to assist recommendations. In addition, they also face the problem of insufficient supervision signals (feedback data). Other models use spatio-temporal sequence information to assist in modeling. However, they do not explicitly encode users' preferences and item features and effectively integrate them into the model, resulting in a shallow understanding of users' preferences and item features.

[0004] Data augmentation generates new training samples by performing various transformations on the original data. For spatio-temporal data, previous studies mostly used traditional methods for augmentation. They have made some progress, but still have not solved the inherent problems of recommendation systems. Specifically, they do not directly augment user-item interactions and still have the problem of low-quality user feedback data.

[0005] In recent years, large models have been booming and have achieved remarkable success in many tasks and fields. Initially, large models were developed as pre-trained language-based models to solve different natural language tasks. However, over time, the number of parameters of large models has increased sharply, which enables them to show powerful natural language understanding ability, learn complex semantics and knowledge representations from large-scale text corpora, and complete various reasoning tasks. Some recent studies have tried to use large models to assist recommendations. But they only augment ordinary features and do not consider spatio-temporal features.

[0006] Based on the deficiencies of existing research, it can be seen that current recommendation systems mainly face two challenges. The first is that spatio-temporal features are not explicitly summarized and encoded into the recommendation model. Due to the limitations of privacy protection in collecting user location information and behavior time, there may be insufficient data to mine users' spatio-temporal preferences, summarize them in text form, and finally encode and integrate them into the recommendation system. The second is the low-quality user feedback data, including data sparsity and noise problems. In terms of data sparsity, due to time and space limitations, users can often only interact with a small part of the items in the system, which leads to sparse feedback data. For example, in a food recommendation system, limited by working hours, users may only be active on holidays. In terms of the noise problem, users may interact with items influenced by popular trends or social circles, which does not represent their true preferences. For example, a user chooses a restaurant because of the high rating in the recommendation system, but actually does not like it.

[0007] In summary, the current research on spatio-temporal recommendation systems has problems such as insufficient feature fusion and low-quality user feedback data. At this time, it is urgent to find a solution that can solve the above problems and improve the performance of the spatio-temporal recommendation model. Summary of the Invention

[0008] In view of the deficiencies of the existing technology, the present invention proposes a spatio-temporal enhancement recommendation method based on a large model. By using the excellent natural language processing and reasoning capabilities of the large model, the large model summarizes the features of users and items and encodes and integrates them into the recommendation system to solve the problem of insufficient feature fusion in the spatio-temporal recommendation system.

[0009] In addition, the present invention uses the large model to generate positive and negative sample pairs for each user, expands the number of training samples, and reduces the noise caused by random sampling of the Bayesian Personalized Ranking (BPR) algorithm to solve the problem of low-quality user feedback data.

[0010] The present invention provides a spatio-temporal data enhancement method for a recommendation system, including the following steps:

[0011] Step 1: Use the large model to enhance the time features of items to obtain enhanced item time features;

[0012] Step 2: Use the reasoning ability of the large model to summarize the general features, time features, and spatial features of users according to the interaction history of users to obtain enhanced user features;

[0013] Step 3: Use the large model to understand the interaction history of users and the candidate set, and infer the items that users are interested in and not interested in as positive and negative sample pairs to expand the training sample set;

[0014] Step 4: Input the enhanced item time features, the original item general features and spatial features in the dataset, and the enhanced user features into the encoding layer to generate high-dimensional vectors, and then use a linear layer for dimensionality reduction;

[0015] Step 5: Input the dimensionality-reduced feature vectors and the id embedding vectors generated by the basic recommendation model into the feature fusion layer, and obtain a comprehensive representation through weighted fusion;

[0016] Step 6: Train the recommendation model using the BPR loss function. Input the comprehensive representation into the BPR loss function without sample pruning, and combine the enhanced samples and the original training set as the final training set to obtain the first loss function;

[0017] Step 7: For the time features and spatial features in the dimensionality-reduced item and user features, input them into the BPR layer with pruning, and adopt a sample pruning strategy to obtain the second loss function;

[0018] Step 8: Fuse the first loss function and the second loss function according to a certain weight to obtain the total loss function, and use this total loss function to optimize the model training.

[0019] The present invention also provides a recommendation system integrated with the above spatio-temporal data enhancement method.

[0020] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0021] 1. The present invention uses a large model to enhance spatio-temporal data and applies it to a recommendation system. Compared with traditional spatio-temporal recommendation methods, the present invention explicitly summarizes the spatio-temporal features of users and items, uses a large model as the encoding layer to unify the embeddings in a single vector space; then uses a linear layer to align the feature dimensions for fusion into the recommendation model, solving the problem of insufficient feature fusion.

[0022] 2. Sparse and noisy user feedback is an inherent problem in recommendation systems. The present invention uses a large model to understand the user's interaction history and candidate set, infer the user's preferences and generate positive and negative sample pairs, and combines the generated sample set and the original sample set as the final training set. Thanks to the excellent reasoning ability and natural language understanding ability of the large model, this not only expands the training sample set but also alleviates the noise problem.

[0023] 3. In order to emphasize relevant supervision signals and prevent incorrect gradient descent, the present invention adopts a pruning strategy. Use spatio-temporal features for sample pruning, removing a part of easily distinguishable training sample pairs to make the optimization more stable and effective.

[0024] 4. The spatio-temporal data enhancement method proposed by the present invention is pluggable and can be integrated into various recommendation models to improve their recommendation effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art.

[0026] Figure 1 It is a schematic diagram of the framework of the offline stage of a spatio-temporal enhanced recommendation technology based on a large model according to an embodiment of the present application.

[0027] Figure 2 It is a schematic diagram of the framework of the online stage of a spatio-temporal enhanced recommendation technology based on a large model according to an embodiment of the present application.

[0028] Figure 3 It is the structure of the BPR layer with pruning in an embodiment of the present application.

[0029] Figure 4 It is a schematic diagram of three prompt templates in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] The following will illustrate the specific implementation manners of the present invention in conjunction with the drawings and embodiments. However, the following embodiments are only used to illustrate the present invention in detail and do not limit the scope of the present invention in any way.

[0031] Overall concept of the present application: The present application can be divided into two stages, namely an offline stage of enhancing data using a large model and an online stage of training a recommendation model using the enhanced data. The offline stage can be further divided into three modules: item enhancement, user enhancement, and user-item enhancement. In the item enhancement module, the large model is used to transform the time features of items. In the user enhancement module, the large model infers the preference information of users. In the user-item enhancement module, the large model processes the interaction history of users and candidate sets to generate positive and negative sample pairs. In the online stage, in order to fuse the enhanced features into the recommendation model, these text features are encoded and dimensionality-reduced. Finally, the present application uses the BPR loss function to optimize the model and uses spatio-temporal features to prune samples.

[0032] An embodiment of the present application discloses a spatio-temporal data enhancement method based on a large model, Figure 1 and Figure 2 respectively gives the schematic diagrams of the frameworks of the offline stage and the online stage of this embodiment. The following will describe the present application in detail in conjunction with the embodiments, specifically including the following steps:

[0033] Step 1: Use a large model to enhance the time features (business hours) of items.

[0034] Concatenate the specific business hours of the merchant (item) with the item prompt template. The item prompt template includes three parts: task description, specific business hours, and output format description. See Figure 4, input the concatenated prompt into the large model to generate enhanced time features.

[0035] Furthermore, the features of merchants can be divided into general features C i (category), spatial features S i (location) and time features T i (business hours). They all exist in the original dataset and are stored using the nested dictionary items. The keys in the first layer are the three types of features, and the second-layer dictionary stores the corresponding relationship between merchant IDs and text features. Among them, the business hours of merchants are too specific and have a fine granularity, which may lead to large errors in the user time features inferred by the large model and make it difficult to capture the user-merchant collaboration signal. Therefore, this step is used to enhance the time features of merchants. Specifically, concatenate the specific business hours of merchants with the item prompt template and then input it into the large model , to obtain enhanced time features T i , and replace the corresponding content in the dictionary items. The formula is as follows:

[0036]

[0037] Among them, P i is the item prompt template, and T i is the feature in text form; after enhancement, in this implementation, these three types of features are input into the encoding layer for further processing.

[0038] Step 2: Utilize the inference ability of the large model to summarize the general features (preferred categories), time features (preferred dining times), and spatial features (preferred dining locations) of the user based on the user's interaction history.

[0039] First, replace the specific business hours of each merchant in the user's interaction history with the enhanced result of Step 1, then concatenate the interaction history with the user prompt template, and enhance it through the large model. The user prompt template includes three parts: task description, user interaction history, and output format description. See Figure 4 .

[0040] Then call the split function in Python to split the enhanced result to obtain the general features C u 、time features T u and spatial features S u of each user. Store them using the nested dictionary users. The keys in the first layer are the three types of features, and the second-layer dictionary stores the corresponding relationship between user IDs and text features. The formula is as follows:

[0041]

[0042] Among them, P u is the user prompt template, and C u, S u , T u These are text - form features. After enhancement, in this implementation, these three features are input into the encoding layer for further processing.

[0043] Step 3: Since BPR random sampling can lead to problems of false positive samples (noise) and false negative samples (non - interactions). Therefore, it is necessary to use a large - model to understand the user's interaction history and candidate set, and infer the items that the user is interested in and not interested in as positive and negative sample pairs, so as to achieve the purpose of expanding the training sample set and reducing noise.

[0044] Specifically, the user's interaction history and candidate set are concatenated with the user - item prompt template and input into the large - model. Since the large - model has a token number limit when processing input, in this implementation, the basic recommendation model scores all items and ranks them in descending order, and a small part of them is selected as the positive - sample candidate set and negative - sample candidate set (five each). In this embodiment, the top five ranked items are selected as the positive - sample candidate set.

[0045] The user - item prompt template includes four parts: task description, interaction history, candidate set, and output - format description, as shown in Figure 4 . The candidate set is obtained by screening the entire item set using the basic recommendation model.

[0046] Selecting the negative - sample candidate set is relatively difficult. On the one hand, to avoid being a positive sample, their rankings cannot be too high. On the other hand, to stimulate and improve the performance of the recommendation model, these negative - sample candidate sets should have a certain correlation with the positive samples. Therefore, their rankings cannot be too low.

[0047] After many experiments, it is found that screening five items from the items ranked in the top 10% as the negative - sample candidate set can achieve the best results. The large - model outputs enhanced positive and negative sample pairs <p, n>, and the formula is as follows:

[0048]

[0049] where P ui is the user - item prompt template, and p and n represent the positive sample and negative sample generated by the large - model respectively.

[0050] The enhanced implicit feedback set S A consists of paired training triples <u, p, n>, where u represents a certain user. In this implementation, S A is merged with the original training set S in a certain proportion as the final training set.

[0051] Step 4: Input the enhanced item time features generated in Step 1, the original item general features (categories), spatial features (locations) in the dataset, and the enhanced user features generated in Step 2 into the encoding layer to generate high-dimensional vectors, and then use a linear layer to reduce the dimension.

[0052] The item features and user features output by the large model are in text form. To be able to fuse with traditional recommendation models, these texts need to be encoded into vector representations in this implementation. Since the large model has excellent context awareness ability and can learn rich semantic information, the large model is used as the encoding layer in this implementation. It is represented formulaically as follows:

[0053]

[0054]

[0055] Among them, c i , s i , t i represent the three categories of encoded item features, and c u , s u , t u represent the three categories of encoded user features, which are embedding vectors of shape 1×h.

[0056] Since the vectors encoded by the large model have too high a dimension and cannot be directly fused with the recommendation model. Therefore, this implementation uses a linear layer with dropout to reduce its dimension. While capturing key information, it can also reduce the risk of overfitting. The formula is as follows:

[0057]

[0058]

[0059] Among them, , , represent the three categories of item features after dimension reduction, represents the three categories of user features after dimension reduction. They are embedding vectors of shape 1×l, where l << h.

[0060] Step 5: Input the feature vectors after dimension reduction obtained in Step 4 and the id embedding vectors generated by the basic recommendation model into the feature fusion layer, and obtain a comprehensive representation through weighted fusion.

[0061] Traditional recommendation models, such as NGCF and GCMC, use collaborative filtering mechanisms and multi-layer neural networks to aggregate information, and can obtain the item id embedding vector Id i and the user id embedding vector Id u through multiple-layer propagation mechanisms.

[0062] This implementation inputs the item features, user features, and Id after dimensionality reduction i and Id u into the feature fusion layer together, and obtains a comprehensive representation through weighted fusion. The specific calculation formula is as follows:

[0063]

[0064]

[0065] where, respectively represent the comprehensive item representation and user representation. This implementation uses to adjust the fusion weight of spatio-temporal features and to adjust the fusion weight of ordinary features. In addition, this implementation uses L2 normalization to process the low-dimensional feature vectors to eliminate the scale differences between different features and improve numerical stability.

[0066] Step 6: Train the recommendation model using the BPR loss function. Since not all the enhanced sample pairs generated in Step 3 are reliable interactions and may contain noise, this implementation uses a parameter to control its fusion ratio with the original training set to generate the final training set.

[0067] Input the comprehensive representation generated in Step 5 into the BPR loss function without sample pruning, and merge the enhanced samples generated in Step 3 and the original training set as the final training set. This step will generate Loss1, and the specific calculation formula is as follows:

[0068]

[0069] where, the training triple <u, p, n> is selected from S∪S A . Since the samples generated by the user-item enhancement module are not all reliable interactions and may still contain noise, using all of them for model training will damage the performance. Therefore, this implementation needs to control the number of samples in S A .

[0070] Specifically, . Where, represents the enhanced sample fusion rate, which controls the proportion of enhanced samples in the final BPR training set, B represents the batch size, which controls the number of samples used for training the model in each iteration. . Where, is the comprehensive user representation calculated in Step 5, is the comprehensive item representation of the positive sample, is the comprehensive item representation of the negative sample, respectively represent the scores of the user for the positive sample and the negative sample, The non-linear activation function is sigmoid. To prevent overfitting, this implementation uses to control the L2 regularization strength .

[0071] Step 7: For the dimension-reduced item and user features generated in Step 4, select the time features and spatial features among them, input them into the BPR layer, and adopt a sample pruning strategy to emphasize the relevant supervision signals, as shown in Figure 3 .

[0072] Specifically, to calculate Loss2, this implementation takes the time feature of the item , the spatial feature , the time feature of the user , and the spatial feature and inputs them into a BPR layer with sample pruning. It contains four BPR loss functions, which are used to receive four combinations of spatio-temporal features respectively.

[0073] To emphasize the relevant supervision signals, this implementation uses spatio-temporal features to prune the samples and remove some easily distinguishable training triples, which can make the optimization more stable and effective. The calculation formula is as follows:

[0074]

[0075] Among them, the function sorts the losses of each training triple in ascending order and selects the top M items. is the pruning rate. When ; When ; When When .

[0076] Each BPR loss function calculates the loss of a spatio-temporal combination, and Loss2 is obtained by adding the losses of the four BPR loss functions. The specific calculation formula is:

[0077]

[0078] Step 8: Fuse the loss calculated in Step 6 and the loss calculated in Step 7 according to a certain weight to obtain the final total loss . Use this loss function to optimize the model training. The specific calculation formula is:

[0079]

[0080] Among them For control In The proportion in

[0081] Verification example:

[0082] The dataset used in this example is the Yelp dataset provided by Yelp. Since the original dataset is too large, this example selects the interaction data in 2018, filters out merchants and users with less than 6 interactions, and regards all user-merchant interactions as implicit feedback.

[0083] This example adopts three commonly used evaluation metrics for recommendation algorithms. Among them, R@20 is the proportion of actually relevant results among the top 20 recommended results in all relevant results in the test set, P@20 is the proportion of actually relevant results among the top 20 recommended results in the top 20 recommended results, and N@20 measures the ranking quality of the top 20 results of the recommended results.

[0084] Table 1 shows the performance improvement after combining two traditional recommendation models with the enhanced method of this embodiment.

[0085] Table 1 Overall recommendation performance improvement after different models are combined with this embodiment

[0086]

[0087] This embodiment improves the performance of these two models, proving the effectiveness and applicability of this embodiment. The improvement of each of their indicators exceeds 20%. Among them, GCN is better, and the improvement of each indicator exceeds 25%.

[0088] Table 2 shows the impact of the ranking of the negative sample candidate set in the user-enhanced module on the performance of the recommendation model. The recommendation model here takes (After being enhanced using this embodiment ) as an example for illustration. The ranking refers to the ranking of the negative sample candidate set in the entire item set. For example, if the dataset has 9591 items, ranking = 0.1% means selecting 5 items around the 10th place.

[0089] Table 2 Analysis of the ranking of the negative sample candidate set

[0090]

[0091] It can be seen from Table 2 that: the results with ranking = 0.1% are the worst because their positions are too far forward, which will lead to deviation in model learning. As the ranking increases, the recommendation performance first rises and then falls. Ranking = 10.0% achieves the best performance. Too large a ranking will lead to a decline in performance because their positions are too far back, with low similarity to the positive samples and it is difficult to promote model optimization.

[0092] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, without departing from the principle and spirit of the present invention, various changes, modifications, substitutions, and variations to these embodiments including components still fall within the protection scope of the present invention.

Claims

1. A spatiotemporal data enhancement method for recommendation systems, characterized in that: The following steps are involved: Step 1: Use the large model to enhance the time characteristics of the object to obtain enhanced time characteristics of the object; Step 2: Using the reasoning ability of the large model, summarize the user's general features, temporal features, and spatial features based on the user's interaction history to obtain enhanced user features; Step 3: Use the large model to understand the user's interaction history and candidate set, infer the items that the user is interested in and not interested in as positive and negative sample pairs, and expand the training sample set; Step 4: Input the enhanced item temporal features, the original item common features and spatial features in the data set, and the enhanced user features into the encoding layer to generate a high-dimensional vector, and then use the linear layer to reduce the dimension; Step 5: Input the reduced feature vector and the ID embedding vector generated by the basic recommendation model into the feature fusion layer, and obtain a comprehensive representation through weighted fusion; Step 6: Use the BPR loss function to train the recommendation model, input the comprehensive representation into the BPR loss function without sample pruning, and combine the enhanced samples and the original training set as the final training set to obtain the first loss function; Step 7: Input the time features and spatial features of the reduced-dimensional item and user features into the BPR layer with pruning, adopt the sample pruning strategy, and obtain the second loss function; Step 8: Fuse the first loss function and the second loss function according to certain weights to obtain a total loss function, and use the total loss function to optimize model training.

2. The method for spatiotemporal data enhancement for recommendation systems according to claim 1, characterized in that: In step 1, enhancing the time feature of the item is specifically as follows: splicing the specific business hours of the item with an item reminder template, wherein the item reminder template includes a task description, specific business hours, and output format instructions; The concatenated prompts are fed into the large model to generate enhanced temporal features and replace the temporal features of the corresponding items in the original dataset.

3. The method for spatiotemporal data enhancement for recommendation systems according to claim 1 or 2, characterized in that: In step 2, the user characteristics are summarized as follows: Replace the specific business hours of each item in the user interaction history with the enhanced time feature in step 1; splicing the replaced interaction history with a user prompt template, wherein the user prompt template includes a task description, a user interaction history, and an output format description; The spliced ​​cues are fed into the large model to generate enhanced general, temporal and spatial features.

4. The method for spatiotemporal data enhancement for recommendation systems according to claim 1, characterized in that: In step 3, generating positive and negative sample pairs is specifically as follows: splicing the user's interaction history and candidate set with a user-item prompt template, wherein the user-item prompt template includes a task description, an interaction history, a candidate set, and an output format description; Use the basic recommendation model to score all items and rank them in descending order, select the top five items as the positive sample candidate set, and select several items from the top 10% of the items as the negative sample candidate set; The concatenated prompts are fed into the large model to generate positive and negative sample pairs.

5. The method for spatiotemporal data enhancement for recommendation systems according to claim 1, characterized in that: In step 4, encoding and dimensionality reduction are specifically as follows: Use a large model as the encoding layer to encode enhanced text features into high-dimensional vectors; Use a linear layer with dropout to reduce the dimensionality of the high-dimensional vector to obtain a low-dimensional feature vector.

6. The method for spatiotemporal data enhancement for recommendation systems according to claim 1, characterized in that: The step 5 also includes using L2 normalization to process the low-dimensional feature vector to eliminate scale differences between different features.

7. The method for spatiotemporal data enhancement for recommendation systems according to claim 6, characterized in that: In step 6, the BPR loss function is specifically used as follows: Input the comprehensive representation generated in step 5 into the BPR loss function without sample pruning; Combine the enhanced samples and the original training set in a certain ratio as the final training set; Use control to enhance the proportion of samples in the training set to prevent the influence of noise on the training of the recommendation model.

8. The method for spatiotemporal data enhancement for recommendation systems according to claim 1, characterized in that: In step 7, the sample pruning strategy is as follows: Input the reduced temporal and spatial features of items and users into the BPR layer with pruning; Use spatiotemporal features to prune samples and remove some easily distinguishable training triplets; Four different spatiotemporal feature combinations are processed respectively through four BPR loss functions. After each BPR loss function is processed, a sample pruning strategy is adopted to obtain the second loss function.

9. The method for spatiotemporal data enhancement for recommendation systems according to claim 1, characterized in that: In step 8, the fusion loss function is specifically: The first loss function calculated in step 6 and the second loss function calculated in step 7 are combined according to certain weights to obtain a total loss function; Use the total loss function to optimize the recommendation model training to improve the performance of the recommendation system.

10. A recommendation system, characterized in that: The spatiotemporal data enhancement method according to any one of claims 1 to 9 is integrated.

Images