An information-based commodity recommendation method

Abstract

The application relates to a kind of information-based commodity recommendation method, comprising: obtaining historical data, using correlation level marking mode, early marking is carried out to historical data;The labeled grade data combination is used as training target, and a plurality of initial recommendation models are obtained by model training;Based on score matrix, the loss function corresponding to a plurality of initial recommendation models is calculated respectively, and the initial recommendation model with the minimum loss function value is screened out as a verification recommendation model;The verification recommendation model is verified correspondingly, and the final recommendation model is obtained;The actual information name and all commodity data are input into the final recommendation model, and the corresponding recommended commodity is output. Compared with the prior art, the application can effectively improve the labeling efficiency and the richness of the correlation, thereby ensuring the accuracy of the constructed model, i.e. ensuring the accuracy of subsequent commodity recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of commodity recommendation, and in particular to a commodity recommendation method based on information. BACKGROUND

[0002] Improving conversion rate is one of the core indicators of a media platform, and commodity recommendation is a powerful tool to achieve this goal. Generally, it is divided into explicit recommendation and implicit recommendation. Among them, explicit recommendation is based on the premise of platform's own high-quality commodities, which are concentrated and packaged, and then recommended to customers by adding dimensions such as commodity freshness, popularity, price discount, and commodity quality. This way of demand positioning is not direct, and the conversion rate is relatively low. Implicit recommendation is based on the premise of the user's potential demand, and different commodities are pushed to different users by extracting corresponding commodities. This way of demand positioning is direct, and the conversion rate is relatively high. However, with the emergence of massive information on various platforms, how to quickly and accurately mine the commodities of interest to users is a problem that needs to be solved.

[0003] The existing commodity recommendation for information is mainly cold start, which needs to manually determine the information and commodities, and mark the commodities with high similarity to the information as positive examples and the commodities with low similarity as negative examples. The model is trained according to the combination of information feature vectors, commodity feature vectors, and artificial labels, the similarity between information and commodities is calculated, and cross-entropy is used as the model loss function. This positive and negative example marking method has the problem of low efficiency. For example, if a certain article and 10 commodities are marked, only 10 sample data

article + commodity + positive and negative labels

[0004] In addition, there is also a problem of poor richness of related relationships. Marking commodities as positive and negative examples has a large granularity of related relationships, and cannot reflect the differences between positive examples. For example, for an information related to a mobile phone, according to the positive and negative example marking rule, commodities of the intelligent electrical appliance category will be marked as positive examples, and the differences in related relationships between the mobile phone, computer, and washing machine and the target information cannot be reflected.

[0005] The above will result in inaccurate final recommended commodities, which are difficult to accurately adapt to user demand. SUMMARY

[0006] The present application relates to the technical field of commodity recommendation, and in particular to a commodity recommendation method based on information.

[0007] The purpose of the present application can be achieved by the following technical scheme: a commodity recommendation method based on information, comprising the following steps:

[0008] S1, acquire historical data, adopt a correlation level marking method, and mark the historical data in the early stage;

[0009] S2, combine the marked level data as a training target, train the model, and obtain a plurality of initial recommendation models;

[0010] S3, based on the score matrix, calculate the loss function corresponding to the plurality of initial recommendation models respectively, and select the initial recommendation model with the minimum loss function value as the verification recommendation model;

[0011] S4, verify the verification recommendation model to obtain the final recommendation model;

[0012] S5, input the actual information name and all commodity data into the final recommendation model, and output the corresponding recommended commodities.

[0013] Further, the step S1 is specifically for the information name in the historical data, and the correlation level marking is performed on different commodities in the historical data.

[0014] Further, the correlation level marking includes:

[0015] 0, meaning unrelated commodities;

[0016] 1, meaning relatively related commodities;

[0017] 2, meaning highly related commodities.

[0018] Further, the step S2 combines the marked level data, specifically: information vector + positive example commodity vector + negative example commodity vector.

[0019] Further, the positive example commodity and the negative example commodity have a relative relationship. For any two commodity combinations, the one with a higher correlation level marking value is the positive example commodity, and the one with a lower correlation level marking value is the negative example commodity.

[0020] Further, the step S3 specifically includes the following steps:

[0021] S31, calculate the score matrix corresponding to the plurality of initial recommendation models respectively through a semantic model;

[0022] S32, further calculate the loss function of the corresponding initial recommendation model according to the score matrix;

[0023] S33, select the initial recommendation model with the minimum loss function value as the verification recommendation model.

[0024] Further, the specific process of the step S31 is:

[0025] S311, embedding of the information title, the positive example commodity title and the negative example commodity title is calculated respectively through the semantic model;

[0026] S312, information matrix, positive example commodity matrix and negative example commodity matrix are constructed respectively;

[0027] S313, the positive example commodity matrix and the negative example commodity matrix are combined into a unified commodity matrix, and the inner product between the information matrix and the unified commodity matrix is calculated, that is, a score matrix is obtained.

[0028] Further, the step S312 is specifically combining information into an information matrix, splicing positive example commodities into a positive example commodity matrix, and splicing negative example commodities into a negative example commodity matrix.

[0029] Further, the positive example commodities and the negative example commodities are spliced in a horizontal splicing mode to correspondingly splice the positive example commodity matrix and the negative example commodity matrix.

[0030] Further, the calculation formula of the loss function in the step S32 is:

[0031]

[0032] Wherein, sim(h, h + ) is the inner product similarity of information embedding and positive example commodity embedding, and sim(h, h - ) is the inner product similarity of information embedding and negative example commodity embedding.

[0033] Compared with the prior art, when the pre-labeling is performed, the present application adjusts from positive and negative example labeling to correlation level labeling, when the data accumulation is insufficient, the labeling mode of the correlation level is used to reflect the difference between the correlation relationships, and the combination of "information vector + positive example commodity + negative example commodity" is used as the model training target, which can effectively improve the labeling efficiency and the richness of the correlation relationship, and further ensure the accuracy of the constructed model, that is, the accuracy of the subsequent commodity recommendation.

[0034] The present application adopts the correlation level labeling mode, which can effectively improve the labeling efficiency, compared with the existing labeling mode, under the condition of the same labeling amount, the actual training sample amount of the present application presents exponential growth, and the correlation relationship is more rich; the present application determines the training target in the form of positive and negative example combination, which can fully ensure the balance of the positive and negative proportion of the training sample; the present application trains the model through the triple group, compared with the traditional binary group training, the triple group integrates the relationship between the positive and negative examples, which can directly learn the difference between the positive and negative examples, through the reasonable combination of the positive and negative examples, the relative relationship between the samples can be further learned, so that the model can deduce the correlation, and effectively improve the accuracy of the recommendation model.

[0035] This invention combines information into an information matrix, horizontally concatenates positive example products into a positive example product matrix, and horizontally concatenates negative example products into a negative example product matrix. Then, it combines the positive and negative example product matrices into a unified product matrix, calculates the inner product between the information matrix and the unified product matrix to obtain a score matrix. Based on the score matrix, a loss function is calculated, and the initial recommendation model with the smallest loss function value is selected as the validation recommendation model. This approach effectively measures the model loss from the perspective of positive and negative example combinations, ensuring the model's accuracy. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating the process of calculating the fraction matrix in the embodiment. Detailed Implementation

[0038] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0039] Example

[0040] like Figure 1 As shown, an information-based product recommendation method includes the following steps:

[0041] S1. Obtain historical data and perform preliminary labeling on the historical data using a relevant level labeling method. Specifically, this involves labeling different products in the historical data with relevant levels based on the information names within the historical data. In this embodiment, the relevant level labeling includes:

[0042] 0 means unrelated product;

[0043] 1. Meaning relatively related goods;

[0044] 2. Meaning highly relevant goods;

[0045] S2. Using the labeled level data combination as the training target, multiple initial recommendation models are obtained through model training. The labeled level data combination is specifically: information vector + positive example product vector + negative example product vector. Positive example products and negative example products are relative. For any two product combinations, the one with the higher related level label value is the positive example product, and the one with the lower related level label value is the negative example product.

[0046] S3. Based on the score matrix, calculate the loss function for each of the multiple initial recommendation models, and select the initial recommendation model with the smallest loss function value to be used as the validation recommendation model. Specifically:

[0047] First, using a semantic model, score matrices corresponding to multiple initial recommendation models are calculated separately—the embeddings of news titles, positive product titles, and negative product titles are calculated separately, and news matrices, positive product matrices, and negative product matrices are constructed separately (the news is combined into a news matrix, the positive products are horizontally concatenated into a positive product matrix, and the negative products are horizontally concatenated into a negative product matrix). Then, the positive product matrix and the negative product matrix are combined into a unified product matrix, and the inner product between the news matrix and the unified product matrix is ​​calculated, which yields the score matrix.

[0048] Then, based on the score matrix, the loss function corresponding to the initial recommendation model is calculated:

[0049]

[0050] In the formula, sim(h, h) + ) represents the inner product similarity between the information embedding and the positive example product embedding, sim(h, h) - The inner product similarity between the information embedding and the negative example product embedding is denoted as .

[0051] The initial recommendation model with the smallest loss function will then be used as the validation recommendation model.

[0052] S4. Perform corresponding validations on the validation recommendation model to obtain the final recommendation model;

[0053] S5. Input the actual information name and all product data into the final recommendation model, and output the corresponding recommended products.

[0054] This embodiment applies the above technical solution and mainly includes the following:

[0055] 1. When making initial annotations, change the annotation from positive and negative examples to relevant level annotations (e.g., 0: irrelevant products, 1: somewhat relevant products, 2: highly relevant products), as shown in Table 1.

[0056] Table 1

[0057]

[0058] Second, use a combination of labeled graded data: [information vector + positive example product vector + negative example product vector] as the training target. It should be noted that positive and negative examples are relative; for any two product combinations, the one with the higher relevance grade is designated as the positive example product, and the one with the lower grade as the negative example product. This labeling method greatly improves labeling efficiency.

[0059] For example, when a news article is labeled with 10 irrelevant products, 10 somewhat relevant products, and 10 highly relevant products, the traditional modeling method can only generate 30 training data points, while the positive and negative example comparison method can generate 300 training data points. This not only simplifies the labeling process but also enhances the richness of the correlations between training samples. Examples are shown in Table 2.

[0060] Table 2

[0061]

[0062] III. Figure 2 As shown, the embeddings of the information title, positive product title, and negative product title are calculated using a semantic model. The information is then combined into an information matrix. First, the positive products are horizontally concatenated to form a positive product matrix. Then, the negative products are horizontally concatenated to form a negative product matrix. Finally, the positive and negative product matrices are combined into a unified product matrix. The inner product between the information matrix and the unified product matrix is ​​calculated to obtain the score matrix.

[0063] IV. Calculate the loss function using the fractional matrix, as shown in the following formula:

[0064]

[0065] In the formula, sim(h, h) + ) represents the inner product similarity between the information embedding and the positive example product embedding, sim(h, h) - The loss function is the inner product similarity between the information embedding and the negative example product embedding. This loss function can effectively measure the model loss from the perspective of positive and negative example combinations.

[0066] In summary, this technical solution uses a relevance-level annotation method to reflect the differences between relevance relationships, and uses a combination of [information vector + positive example product + negative example product] as the model target to train and build the recommendation model. This effectively improves annotation efficiency, and under the same annotation volume, the actual training sample size grows exponentially, resulting in richer relevance relationships. It can fully guarantee sample balance: through the combination of positive and negative examples, it ensures a balanced ratio of positive and negative training samples. It has relational equivalence: compared with traditional binary training, triplet training integrates the relationship between positive and negative examples, and can directly learn the differences between positive and negative examples. Through reasonable combinations of positive and negative examples, it can also learn the relative relationships between samples. For example, for an article introducing computers, computers are positive examples relative to other digital products, and other digital products are positive examples relative to food. The recommendation model built by this technical solution can deduce the relevance: computers > other digital products > food.

Claims

1. An information-based commodity recommendation method, characterized by, The method comprises the following steps: S1, obtaining historical data, and performing preliminary labeling on the historical data by using a correlation level labeling method; S2, combining the labeled level data as a training target, and obtaining a plurality of initial recommendation models through model training; S3, calculating the loss function corresponding to the plurality of initial recommendation models based on a score matrix, and selecting an initial recommendation model with the minimum loss function value as a verification recommendation model; S4, verifying the verification recommendation model to obtain a final recommendation model; S5, inputting actual information names and all commodity data into the final recommendation model, and outputting corresponding recommended commodities; Step S1 is specifically directed to information names in historical data, and different commodities in the historical data are labeled with correlation levels; The labeled level data combination in step S2 is: information vector + positive example commodity vector + negative example commodity vector, the positive example commodity and the negative example commodity have a relative relationship, and for any two commodity combinations, the one with a higher correlation level labeling value is the positive example commodity, and the one with a lower correlation level labeling value is the negative example commodity.

2. The information-based product recommendation method according to claim 1, wherein The correlation level labeling includes: 0, meaning unrelated commodities; 1, meaning more related commodities; 2, meaning highly related commodities.

3. The information-based product recommendation method according to claim 1, wherein Step S3 specifically includes the following steps: S31, calculating the score matrix corresponding to the plurality of initial recommendation models by using a semantic model; S32, further calculating the loss function of the corresponding initial recommendation model according to the score matrix; S33, selecting the initial recommendation model with the minimum loss function value as the verification recommendation model.

4. The information-based product recommendation method according to claim 3, wherein The specific process of step S31 is: S311, calculating the embedding of the information title, the positive example commodity title, and the negative example commodity title by using a semantic model; S312, constructing information matrix, positive example commodity matrix, and negative example commodity matrix respectively; S313, combining the positive example commodity matrix and the negative example commodity matrix into a unified commodity matrix, calculating the inner product between the information matrix and the unified commodity matrix, and obtaining the score matrix.

5. The information-based product recommendation method according to claim 4, wherein Step S312 specifically combines information into an information matrix, splices positive example commodities into a positive example commodity matrix, and splices negative example commodities into a negative example commodity matrix.

6. The information-based product recommendation method according to claim 5, wherein The positive example commodities and the negative example commodities are both spliced by a horizontal splicing method to obtain the positive example commodity matrix and the negative example commodity matrix.

7. The information-based product recommendation method according to claim 4, wherein The calculation formula of the loss function in step S32 is: ， wherein sim(h, h + ) is the inner product similarity of the information embedding and the positive example commodity embedding, and sim(h, h - ) is the inner product similarity of the information embedding and the negative example commodity embedding.

Images