Cold start product click-through rate prediction method and device, and electronic equipment

Abstract

This application relates to a method, apparatus, and electronic device for predicting the click-through rate (CTR) of a cold-start product. The method includes: obtaining the ranking result of the cold-start product in a cold-start model; obtaining a target hot-start feature vector after mapping the cold-start feature vector of the cold-start product to a hot-start model, wherein the cold-start feature vector of the cold-start product refers to the feature vector of the cold-start product in the cold-start model, and the target hot-start feature vector refers to the feature vector of the cold-start product in the hot-start model after mapping, and the target hot-start feature vector is used to characterize the user interaction features of the cold-start product after mapping; and inputting the ranking result and the target hot-start feature vector into the hot-start model to predict the CTR of the cold-start product. This improves the accuracy of the prediction results and avoids the problem of large errors in the prediction results due to the lack of key user interaction features of the cold-start product.

Description

Technical Field

[0001] This application relates to the field of intelligent recommendation technology, and in particular to a method, apparatus and electronic device for predicting the click pass rate of a cold start product. Background Technology

[0002] In real-world recommendation systems, new users and products are constantly being added. Because these new users and products lack sufficiently rich historical behavioral data, they often cannot receive accurate recommendations or be accurately recommended to the right users. This is known as the cold start problem in recommendation systems.

[0003] Currently, existing methods for predicting the click-through rate (CTR) of cold-start products mainly rely on newly trained cold-start models. However, considering the lack of exposure samples for cold-start products, in order for the cold-start model to accurately predict the CTR of cold-start products and other relevant products, existing cold-start models need to perform feature enhancement through feature engineering, meta-learning, and other methods. However, regardless of the feature enhancement methods, cold-start products still lack key user interaction features, resulting in significant errors in the CTR prediction results.

[0004] Therefore, improving the accuracy of click pass rate prediction results for cold start products has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a method, apparatus, and electronic device for predicting the click pass rate of cold start products, in order to solve the problem that existing methods for predicting the click pass rate of cold start products have large errors in the prediction results.

[0006] On one hand, embodiments of this application provide a method for predicting the click-through rate of a cold-start product, the method comprising:

[0007] Obtain the ranking results of the cold start products to be predicted in the cold start model;

[0008] Obtain the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model. The cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after it is mapped to a hot start product. The target hot start feature vector is used to characterize the user interaction features of the cold start product after it is mapped to a hot start product.

[0009] The sorting results and the target hot start feature vector are input into the hot start model to predict the click pass rate of the cold start product.

[0010] Optionally, obtaining the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model includes:

[0011] Map N cold start feature vectors to N hot start feature vectors, where N is a positive integer greater than 1;

[0012] Based on the preset optimization equation, M cold start feature vectors are selected from the N cold start feature vectors, where M is a positive integer greater than 1 and less than N, and the M cold start feature vectors are the M feature vectors that are similar to the cold start feature vectors of the cold start product among the N cold start feature vectors.

[0013] The target hot start feature vector is determined based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors, and the N hot start feature vectors include the M hot start feature vectors.

[0014] Optionally, mapping the N cold-start feature vectors to N hot-start feature vectors includes:

[0015] Using a preset iterative algorithm, the target cold start feature vector is iterated multiple times from the feature vector space of the cold start model to the feature vector space of the hot start model to obtain the iteration error after each iteration. The target cold start feature vector is any one of the N cold start feature vectors.

[0016] If a preset condition is met, the iteration stops, and the feature vector generated by the iteration in the feature vector space of the hot start model is determined as the hot start feature vector corresponding to the target cold start feature vector. The preset condition means that the iteration error is less than a preset error threshold.

[0017] Optionally, the preset iterative algorithm is:

[0018] z i,l =z i,l-1 +εφ p (z i.l-1 )

[0019] Among them, z i,l z represents the feature vector obtained by the l-th iteration of the i-th cold start feature vector in the feature vector space of the cold start model. i,l-1 This represents the feature vector obtained by the (l-1)th iteration of the i-th cold start feature vector in the feature vector space of the cold start model, where i is any positive integer from 1 to N, l is any positive integer greater than 1, ε represents the step size of each iteration, and φ pThe gradient represents the probability distribution obtained based on the feature vector space of the cold start model.

[0020] Optionally, the preset condition is:

[0021]

[0022] Where i represents the i-th cold start feature vector among the N cold start feature vectors, and i is any positive integer from 1 to N, φ p The gradient ε represents the gradient obtained from the probability distribution of the feature vector space based on the cold start model. z This represents the preset error threshold.

[0023] Optionally, determining the target hot start feature vector based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors includes:

[0024] Calculate the M similarities between the M cold start feature vectors and the cold start feature vectors of the cold start product, and use the M similarities as M weight values;

[0025] Calculate the M differences between the M cold start feature vectors and their corresponding hot start feature vectors;

[0026] The target hot-start feature vector is determined based on the M weight values ​​and the M differences.

[0027] Optionally, the formula for calculating the target hot-start feature vector is as follows:

[0028]

[0029] Where v′ represents the target hot-start feature vector, v represents the cold-start feature vector of the cold-start product, and m j Let m' represent the j-th cold start feature vector among the M cold start feature vectors. j Let ψ(m) represent the hot start feature vector corresponding to the j-th cold start feature vector. j (v) represents the similarity between the j-th cold start feature vector and the cold start feature vector of the cold start product, m′ j -m j This represents the difference between the j-th hot-start feature vector and the j-th cold-start feature vector.

[0030] On the other hand, embodiments of this application also provide a click pass rate prediction device for cold start products, the device comprising:

[0031] The first acquisition module is used to acquire the ranking results of the cold start products to be predicted in the cold start model;

[0032] The second acquisition module is used to acquire the target hot start feature vector after the cold start feature vector of the cold start product is mapped to the hot start model. The cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after it is mapped to the hot start product. The target hot start feature vector is used to characterize the user interaction features of the cold start product after it is mapped to the hot start product.

[0033] The prediction module is used to input the ranking results and the target hot start feature vector into the hot start model to predict the click pass rate of the cold start product.

[0034] On the other hand, embodiments of this application also provide an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0035] Memory, used to store computer programs;

[0036] When a processor executes a program stored in memory, it implements the steps of the click pass rate prediction method for a cold start product as described in any embodiment of the first aspect.

[0037] On the other hand, embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the click pass rate prediction method for cold start products as described in any embodiment of the first aspect.

[0038] In this embodiment, the ranking result of the cold start product to be predicted in the cold start model is obtained; the target hot start feature vector after the cold start feature vector of the cold start product is mapped to the hot start model is obtained, wherein the cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after being mapped to the hot start product, and the target hot start feature vector is used to characterize the user interaction features of the cold start product after being mapped to the hot start product; the ranking result and the target hot start feature vector are input into the hot start model to predict the click pass rate of the cold start product. The above method can map the cold start feature vector of a cold start product to the hot start model, thereby obtaining the target hot start feature vector of the cold start product in the hot start model. Then, the hot start model can be used to predict the target hot start feature vector to obtain the click pass rate of the cold start product. Since the obtained target hot start feature vector can reflect the user interaction characteristics of the cold start product after it is mapped to the hot start product, the accuracy of the prediction results can be improved, thus avoiding the problem of large errors in the prediction results caused by the lack of key user interaction characteristics of the cold start product. Attached Figure Description

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating a method for predicting the click pass rate of a cold start product, provided in an embodiment of this application;

[0042] Figure 2 This is a schematic diagram of the model structure of a cold start model provided in an embodiment of this application;

[0043] Figure 3 This is a schematic diagram of the structure of a hot-start model provided in an embodiment of this application;

[0044] Figure 4 An approximate Stan path diagram provided for embodiments of this application;

[0045] Figure 5A schematic diagram illustrating a click pass rate prediction process for a cold start product, provided as an embodiment of this application;

[0046] Figure 6 This is a schematic diagram of a test effect provided in an embodiment of this application;

[0047] Figure 7 This is another schematic diagram illustrating the test results provided in an embodiment of this application;

[0048] Figure 8 This is another schematic diagram of the test effect provided in the embodiments of this application;

[0049] Figure 9 This is another schematic diagram illustrating the test results provided in an embodiment of this application;

[0050] Figure 10 A schematic diagram of a click pass rate prediction device for a cold start product provided in an embodiment of this application;

[0051] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] See Figure 1 , Figure 1 This is a flowchart illustrating a method for predicting the click-through rate of a cold-start product, as provided in an embodiment of this application. Figure 1 As shown, the click-through rate prediction method for this cold start product includes:

[0054] Step 101: Obtain the ranking results of the cold start products to be predicted in the cold start model.

[0055] It should be noted that the click-through rate prediction method for cold-start products provided in this application embodiment can be applied to any product recommendation system, such as an insurance recommendation system, a product recommendation system, a short video recommendation system, etc. This product recommendation system includes a pre-trained cold-start model and a warm-start model. The model structure of the cold-start model is as follows: Figure 2As shown, the main inputs to the cold start model are user features and product features, excluding user interaction features with the cold start product (such as browsing history and purchase behavior). User features include gender and age, which can be constructed using libsvm's feature standard; for example, gender (male) can be converted to the form (gender_male:1). Product features can include product price, type, and attribute information. After feature construction, the constructed user and product features are input into a multi-hot feature compression layer to obtain compressed multi-hot features. These compressed features are then input into a deep neural network to obtain feature vectors in the feature vector space of the cold start model (i.e., cold start feature vectors). Finally, the ranking result is obtained using activation functions, position biases, and feature vectors. The position bias can be the position bias of each product obtained based on the previous ranking result, which can be used to correct the current ranking result. The ranking result refers to the position of each product in the recommendation list; the higher the product's ranking position, the greater its exposure probability. The model structure of the hot start model is as follows: Figure 3 As shown, the difference between the warm start model and the cold start model lies in that the warm start model's input also includes the user's interaction features with the product. The construction method for these interaction features can still refer to the libsvm feature standard. For example, assuming the user browses products A and B, the corresponding browsing history feature can be constructed as (Browsing History_Product A:1, Browsing History_Product B:1). It is important to emphasize that the interaction features input to the warm start model do not include the user's interaction features with the cold start insurance product. Besides the different types of input features, the warm start model and the cold start model also differ in the range of products covered by the product features. Here, we use one cold start product a. C As an example, it is assumed that the product covered by the hot-start model is I. M Therefore, the cold start model covers product I. M +a C The hot start model is used for product I. M The system sorts products and generates a corresponding recommendation list based on the sorting results; the cold start model is used to sort product I... M +a C The data is sorted, and then a corresponding recommendation list is generated based on the sorting results.

[0056] Specifically, the cold start product to be predicted can be one or more, and this embodiment does not impose a specific limitation. The cold start product to be predicted here refers to a product that does not possess user interaction characteristics, such as a product that has not been clicked, converted, or viewed by a user. The ranking result mentioned above refers to the recommended position of the cold start product in the recommendation list of the cold start model, that is, the position of the cold start product in product pool I. M +aC The exposure position in the image.

[0057] In this step, to obtain the ranking results of the cold start product to be predicted in the cold start model, the following can be used: Figure 2 The cold start model shown is used for prediction. Specifically, after feeding user and product features into the cold start model, it predicts the click-through rate (CTR) of each product in the product pool, thus obtaining the ranking of the cold start product in the recommendation list and its exposure position in the product pool. After obtaining the exposure position, it can be saved. Later, after mapping the cold start product to a hot start product, the hot start model can be used to further refine the CTR (i.e., the exposure position) of the cold start product.

[0058] Step 102: Obtain the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model. The cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after mapping to the hot start product. The target hot start feature vector is used to characterize the user interaction features of the cold start product after mapping to the hot start product.

[0059] In this step, the Stan path calculation algorithm can be used to establish several mapping relationships from cold start feature vectors to hot start feature vectors. Then, from these established cold start feature vectors, multiple cold start feature vectors similar to the cold start feature vector of the cold start product are identified as auxiliary points. Finally, based on these auxiliary points, the mapping path from the cold start model to the hot start model of the cold start product is calculated, ultimately yielding the target hot start feature vector. This method transforms the cold start product into a hot start product, obtaining the user interaction features after the cold start product is mapped to the hot start product, thus avoiding the problem of sparse conversion samples for the cold start product, which requires a long waiting time for the cold start process.

[0060] Step 103: Input the sorting results and the target hot start feature vector into the hot start model to predict the click pass rate of the cold start product.

[0061] In this step, after obtaining the ranking results and the target hot-start feature vector, these results and the target hot-start feature vector can be input into the hot-start model. Then, using the activation function in the hot-start model, the click-through rate (pCTR) of the cold-start product can be calculated. Specifically, the formula for calculating the pCTR of the cold-start product is as follows:

[0062] pCTR=A M (,v′)(1)

[0063] Among them, A M denoted by , where POS represents the sorting result and v′ represents the target hot-start feature vector.

[0064] In this embodiment, the cold-start feature vector of a cold-start product can be mapped to a warm-start model to obtain the target warm-start feature vector of the cold-start product in the warm-start model. Then, the warm-start model can be used to predict the click-through rate of the cold-start product. Since the obtained target warm-start feature vector reflects the user interaction characteristics of the cold-start product after mapping to a warm-start product, the accuracy of the prediction results can be improved, thus avoiding the problem of large errors in the prediction results caused by the lack of key user interaction characteristics in the cold-start product.

[0065] Further, step 102 above, obtaining the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model, includes:

[0066] Map N cold start feature vectors to N hot start feature vectors, where N is a positive integer greater than 1;

[0067] Based on the preset optimization equation, M cold start feature vectors are selected from N cold start feature vectors, where M is a positive integer greater than 1 and less than N, and the M cold start feature vectors are the M feature vectors that are similar to the cold start feature vectors of the cold start product among the N cold start feature vectors.

[0068] Based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors, the target hot start feature vector is determined, and the N hot start feature vectors contain the M hot start feature vectors.

[0069] Specifically, the aforementioned N cold start feature vectors can be N cold start feature vectors randomly selected from the feature vector space of the cold start model. The aforementioned M cold start feature vectors are M feature vectors from the N cold start feature vectors that are similar to the cold start feature vectors of the cold start product. The values ​​of N and M can be set according to actual needs, and this application embodiment does not impose specific limitations.

[0070] In one embodiment, when obtaining the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model, N cold start feature vectors can be mapped into N hot start feature vectors first. Then, based on a preset optimization equation, M cold start feature vectors are selected from the N cold start feature vectors. Finally, the target hot start feature vector is determined according to the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors.

[0071] It should be noted that since there are usually multiple products and multiple users in the cold start and warm start models, and the product characteristics of each product and the user characteristics of each user are different, the products recommended to different users will naturally be different. To extract the feature information of each user + product pair, a feature vector corresponding to each user + product pair can be generated in the feature vector space of the cold start and warm start models. When obtaining the warm start feature vector mapped from each cold start feature vector, it is necessary to calculate the Stan path for the user + product pair corresponding to that cold start feature vector. Calculating the Stan path for each user + product pair would result in a significant delay, severely impacting the user experience. Therefore, in this embodiment, M feature vectors similar to the cold start feature vector of the cold start product can be selected from the N cold start feature vectors as auxiliary points. These auxiliary points are then used to calculate the target warm start feature vector of the user + cold start product pair in the warm start model, thereby accelerating the mapping speed of the cold start feature vector of the cold start product from the cold start model feature space to the warm start model feature space.

[0072] Furthermore, the above steps, mapping N cold-start feature vectors to N warm-start feature vectors, include:

[0073] Using a pre-defined iterative algorithm, the target cold start feature vector is iterated multiple times from the feature vector space of the cold start model to the feature vector space of the hot start model to obtain the iteration error after each iteration. The target cold start feature vector is any one of the N cold start feature vectors.

[0074] If the preset conditions are met, the iteration stops, and the feature vector generated by the iteration in the feature vector space of the hot start model is determined as the hot start feature vector corresponding to the target cold start feature vector. The preset conditions refer to the iteration error being less than the preset error threshold.

[0075] Specifically, the iterative algorithm described above is used to iteratively calculate the warm-start feature vector in the target domain (i.e., the feature vector space of the cold-start model) for each cold-start feature vector in the source domain (i.e., the feature vector space of the warm-start model). The preset error threshold can be set according to actual needs and is not specifically limited in this embodiment. It should be noted that this iterative calculation process usually requires many iterations to map the cold-start feature vector from the source domain to the target domain, thus consuming a considerable amount of time.

[0076] In one embodiment, for any cold-start feature vector among the N cold-start feature vectors, a preset iterative algorithm can be used to iterate the cold-start feature vector from the feature vector space of the cold-start model to the feature vector space of the hot-start model multiple times, obtaining the iteration error after each iteration. Then, it is determined whether the iteration error after each iteration satisfies a preset condition. If the preset condition is met, the iteration stops, and the feature vector generated by the iteration in the feature vector space of the hot-start model is determined as the hot-start feature vector corresponding to the cold-start feature vector. Similarly, through the same method described above, the hot-start feature vector corresponding to each of the N cold-start feature vectors can be obtained, thereby facilitating the subsequent acquisition of M cold-start feature vectors as auxiliary points from the N cold-start feature vectors with established mapping relationships.

[0077] Furthermore, the preset iterative algorithm is as follows:

[0078] z i,l = i,l-1 +φ p ( i.l-1 (2)

[0079] Among them, z i,l Let z represent the feature vector obtained by the l-th iteration of the i-th cold start feature vector in the feature vector space of the cold start model. i,l-1 Let represent the feature vector obtained by the (l-1)th iteration of the i-th cold start feature vector in the feature vector space of the cold start model, where i is any positive integer from 1 to N, l is any positive integer greater than 1, ε represents the step size of each iteration, and φ p The gradient represents the probability distribution of the feature vector space based on the cold start model.

[0080] It should be noted that the gradient φ p It can be calculated using the following formula:

[0081]

[0082] Where, k(z,z) i.0 )=exp(-|zz i.0 |2 / σ 2),k(z,z i.0 ) represents the Gaussian kernel function, z i.0 Let z represent the i-th cold-start feature vector in the source domain, z is any test vector in the source domain, and i is any positive integer from 1 to N. P(z) represents the probability distribution of the source domain.

[0083] In this embodiment, by using the above formulas (2) and (3), each cold start feature vector in the N cold start feature vectors can be iteratively calculated in the cold start model feature vector space, and finally the iteration error after each iteration is obtained, thereby improving the accuracy of the iterative calculation.

[0084] Furthermore, the preset conditions are:

[0085]

[0086] Where i represents the i-th cold start feature vector among N cold start feature vectors, and i is any positive integer from 1 to N, φ p The gradient ε represents the gradient obtained from the probability distribution of the feature vector space based on the cold start model. z This indicates the preset error threshold.

[0087] It should be noted that, according to existing research, it can be proven that the z-based iteration eventually converges, i.e., it satisfies the following formula:

[0088]

[0089] Based on the above formula (5), it can be proven that by iterating z continuously, the cold-start feature vector can be mapped from the source domain to the target domain. That is, the mapping method of any cold-start feature vector from the cold-start model feature space to the hot-start model feature space can be obtained, i.e., the Stan path. When the iteration error is less than the error threshold ε... z When the condition is met (i.e., when formula (4) above is satisfied), the iteration can be stopped, and the feature vector generated by the iteration in the feature vector space of the hot-start model can be determined as the hot-start feature vector corresponding to the target cold-start feature vector. In this way, it can be accurately confirmed whether the cold-start feature vector has been mapped from the source domain to the target domain, so as to accurately determine whether it is necessary to continue iterative calculation.

[0090] Further, the above steps, based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors, determine the target hot start feature vector, including:

[0091] Calculate the M similarities between the M cold start feature vectors and the cold start feature vectors of the cold start product, and use the M similarities as M weight values;

[0092] Calculate the M differences between the M cold start feature vectors and their corresponding hot start feature vectors;

[0093] The target hot-start feature vector is determined based on M weight values ​​and M differences.

[0094] In one embodiment, when determining the target hot-start feature vector, M similarities between M cold-start feature vectors and the cold-start feature vectors of the cold-start product can be calculated first, as well as M differences between the M cold-start feature vectors and their respective corresponding hot-start feature vectors. Then, based on the M similarities and M differences, the target hot-start feature vector is calculated. Figure 4 As shown, rectangles represent cold-start feature vectors in the cold-start model feature vector space, and circles represent hot-start feature vectors in the hot-start model feature vector space. Assuming four cold-start feature vectors are selected as auxiliary points (A1, A2, A3, and A4) from the cold-start model feature vector space, their corresponding four hot-start feature vectors (B1, B2, B3, and B4) can be obtained using the aforementioned iterative algorithm. Then, the similarity (ψ1, ψ2, ψ3, and ψ4) between these four auxiliary points and the cold-start feature vector of the cold-start product (target point A5) can be calculated. Finally, based on the similarity of these four auxiliary points and the differences between them and their respective hot-start feature vectors, the target hot-start feature vector B5 is calculated.

[0095] It should be noted that when calculating the similarity between the M cold-start feature vectors and the cold-start feature vectors of the cold-start product, a preset similarity algorithm can be used. This similarity algorithm can include, but is not limited to, cosine similarity algorithm, Pearson correlation coefficient algorithm, Euclidean distance algorithm, etc. As an optional implementation, the cosine similarity algorithm can be used to calculate the similarity between the M cold-start feature vectors and the cold-start feature vectors of the cold-start product.

[0096] In this embodiment, auxiliary points can be used to calculate the target hot start feature vector of the user + cold start product pair in the hot start model, thereby accelerating the mapping speed of the cold start feature vector of the cold start product from the cold start model feature space to the hot start model feature space.

[0097] Optionally, the formula for calculating the target hot-start feature vector is as follows:

[0098]

[0099] Where v′ represents the target hot-start feature vector, v represents the cold-start feature vector of the cold-start product, and m j Let m' represent the j-th cold start feature vector among the M cold start feature vectors.j Let ψ(m) represent the hot start feature vector corresponding to the j-th cold start feature vector. j (v) represents the similarity between the j-th cold start feature vector and the cold start feature vector of the cold start product, m′ j -m j This represents the difference between the j-th hot-start feature vector and the j-th cold-start feature vector.

[0100] It should be noted that m in the above formula (6) j It can be obtained from the following optimization equation:

[0101]

[0102] Where M represents the number of auxiliary points, Ψ = {Ψ ij} represents the similarity matrix, z i.0 Let m represent the i-th cold start feature vector among N cold start feature vectors. j Ψ represents the cold-start feature vector corresponding to the j-th auxiliary point out of M auxiliary points. ij logΨ ij Let represent the regularization term, and α represent the weight of the regularization term. The above optimization equation can be solved using the Expectation-Maximization Algorithm (EM algorithm). The solution to the above optimization equation is the auxiliary point set m = {m j That is, to select M cold start feature vectors from N cold start feature vectors.

[0103] In this embodiment, the set of auxiliary points can be obtained through the optimization equation of the above formula (7). Then, the cold start feature vector corresponding to the cold start product can be mapped to the target hot start feature vector in the hot start model through the above formula (6) and each auxiliary point in the set of auxiliary points, thereby improving the accuracy of the prediction result of the click pass rate of the cold start product.

[0104] In one embodiment, the prediction process for the click pass rate of a cold-start product is as follows: Figure 5 As shown, specifically, the products to be predicted as cold-start products are first sorted based on the cold-start model to obtain the sorting results. Then, the Stan approximation path algorithm is used to map the cold-start products to hot-start products, obtaining the cold-start feature vectors corresponding to the cold-start products and mapping them to the target hot-start feature vectors in the hot-start model. After the mapping is completed, the target hot-start feature vectors and the sorting results are fed into the hot-start model to finally obtain the CTR prediction value of the cold-start products.

[0105] Four widely used ranking models were tested on two public datasets (Tenrec and Kuairec): Deep Neural Networks (DNN), Wide & Deep (WD), Deep & Cross Network (DCN), and DeepFM. The CTR prediction errors of these four models for cold-start product + user pairs were compared with and without the Stan approximation path mapping. Figures 6 to 9 As shown, where, Figures 6 to 9 Columns 1 and 3 represent the prediction error without the Stan approximation path mapping. Figures 6 to 9 Columns 2 and 4 represent the prediction error after adding the Stan approximation path mapping. Compared to the prediction without this algorithm, the prediction error decreased by 10-16% after adding the algorithm, proving the effectiveness of this algorithm.

[0106] See Figure 10 , Figure 10 This is a schematic diagram of a click pass rate prediction device for a cold start product provided in an embodiment of this application. Figure 10 As shown, the device 1000 includes:

[0107] The first acquisition module 1001 is used to acquire the ranking results of the cold start products to be predicted in the cold start model;

[0108] The second acquisition module 1002 is used to acquire the target hot start feature vector after the cold start feature vector of the cold start product is mapped to the hot start model. The cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after it is mapped to the hot start product. The target hot start feature vector is used to characterize the user interaction features after the cold start product is mapped to the hot start product.

[0109] The prediction module 1003 is used to input the ranking results and the target hot start feature vector into the hot start model to predict the click pass rate of the cold start product.

[0110] Furthermore, the second acquisition module 1002 includes:

[0111] The acquisition submodule is used to map N cold start feature vectors into N hot start feature vectors, where N is a positive integer greater than 1;

[0112] The selection submodule is used to select M cold start feature vectors from N cold start feature vectors based on a preset optimization equation, where M is a positive integer greater than 1 and less than N, and the M cold start feature vectors are the M feature vectors that are similar to the cold start feature vectors of the cold start product among the N cold start feature vectors.

[0113] The determination submodule is used to determine the target hot start feature vector based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors. The N hot start feature vectors contain the M hot start feature vectors.

[0114] Furthermore, the acquisition sub-modules include:

[0115] The iterative unit is used to iterate the target cold start feature vector from the feature vector space of the cold start model to the feature vector space of the hot start model multiple times using a preset iterative algorithm, and obtain the iteration error after each iteration. The target cold start feature vector is any one of the N cold start feature vectors.

[0116] The first determining unit is used to stop the iteration when a preset condition is met, and to determine the feature vector generated by the iteration in the feature vector space of the hot start model as the hot start feature vector corresponding to the target cold start feature vector. The preset condition means that the iteration error is less than a preset error threshold.

[0117] Furthermore, the preset iterative algorithm is as follows:

[0118] z i,l =z i,l-1 +εφ p (z i.l-1 )

[0119] Among them, z i,l Let z represent the feature vector obtained by the l-th iteration of the i-th cold start feature vector in the feature vector space of the cold start model. i,l-1 Let represent the feature vector obtained by the (l-1)th iteration of the i-th cold start feature vector in the feature vector space of the cold start model, where i is any positive integer from 1 to N, l is any positive integer greater than 1, ε represents the step size of each iteration, and φ p The gradient represents the probability distribution of the feature vector space based on the cold start model.

[0120] Furthermore, the preset conditions are:

[0121]

[0122] Where i represents the i-th cold start feature vector among N cold start feature vectors, and i is any positive integer from 1 to N, φ pThe gradient ε represents the gradient obtained from the probability distribution of the feature vector space based on the cold start model. z This indicates the preset error threshold.

[0123] Furthermore, the sub-modules are identified as including:

[0124] The first calculation unit is used to calculate the M similarities between the M cold start feature vectors and the cold start feature vectors of the cold start product, and to use the M similarities as M weight values.

[0125] The second calculation unit is used to calculate the M differences between the M cold start feature vectors and their corresponding hot start feature vectors.

[0126] The second determining unit is used to determine the target hot start feature vector based on M weight values ​​and M differences.

[0127] Furthermore, the formula for calculating the target hot-start feature vector is as follows:

[0128]

[0129] Where v′ represents the target hot start feature vector, v represents the cold start feature vector of the cold start product, and m j Let m represent the j-th cold start feature vector among M cold start feature vectors. j ' represents the hot start feature vector corresponding to the j-th cold start feature vector, ψ(m j ,) represents the similarity between the j-th cold start feature vector and the cold start feature vector of the cold start product, m j ′- j This represents the difference between the j-th hot-start feature vector and the j-th cold-start feature vector.

[0130] It should be noted that the device 1000 can implement the steps of the click pass rate prediction method for cold start products provided in any of the aforementioned method embodiments, and achieve the same technical effect, which will not be described in detail here.

[0131] like Figure 11 As shown in the figure, this application embodiment also provides an electronic device, including a processor 1111, a communication interface 1112, a memory 1113, and a communication bus 1114, wherein the processor 1111, the communication interface 1112, and the memory 1113 communicate with each other through the communication bus 1114.

[0132] Memory 1113 is used to store computer programs;

[0133] In one embodiment of this application, when the processor 1111 executes the program stored in the memory 1113, it implements the click pass rate prediction method for cold start products provided in any of the foregoing method embodiments, including:

[0134] Obtain the ranking results of the cold start products to be predicted in the cold start model;

[0135] Obtain the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model. The cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after being mapped to the hot start product. The target hot start feature vector is used to characterize the user interaction features of the cold start product after being mapped to the hot start product.

[0136] The ranking results and the target hot start feature vector are input into the hot start model to predict the click pass rate of cold start products.

[0137] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the click pass rate prediction method for cold start products as provided in any of the foregoing method embodiments.

[0138] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0139] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for predicting the click-through rate of a cold-start product, characterized in that, The method includes: Obtain the ranking results of the cold start products to be predicted in the cold start model; Obtain the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model. The cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after it is mapped to a hot start product. The target hot start feature vector is used to characterize the user interaction features of the cold start product after it is mapped to a hot start product. The sorting results and the target hot start feature vector are input into the hot start model to predict the click pass rate of the cold start product; The step of obtaining the target hot start feature vector after mapping the cold start feature vector of the cold start product to the hot start model includes: Map N cold start feature vectors to N hot start feature vectors, where N is a positive integer greater than 1; Based on the preset optimization equation, M cold start feature vectors are selected from the N cold start feature vectors, where M is a positive integer greater than 1 and less than N, and the M cold start feature vectors are the M feature vectors that are similar to the cold start feature vectors of the cold start product among the N cold start feature vectors. The target hot start feature vector is determined based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors, and the N hot start feature vectors include the M hot start feature vectors; Both the cold start model and the hot start model include a multi-hot feature compression layer, a deep neural network, and an activation function. The multi-hot feature compression layer is used to transform input features into compressed multi-hot features. The deep neural network is used to transform the compressed multi-hot features into feature vectors in the feature vector space. The activation function, together with the position bias and the feature vectors in the feature vector space, determines the ranking result of each product in the recommendation list. The input features of the cold start model include user features and product features, and the input features of the hot start model include user features, product features, and user-product interaction features.

2. The method according to claim 1, characterized in that, The process of mapping N cold-start feature vectors to N hot-start feature vectors includes: Using a preset iterative algorithm, the target cold start feature vector is iterated multiple times from the feature vector space of the cold start model to the feature vector space of the hot start model to obtain the iteration error after each iteration. The target cold start feature vector is any one of the N cold start feature vectors. If a preset condition is met, the iteration stops, and the feature vector generated by the iteration in the feature vector space of the hot start model is determined as the hot start feature vector corresponding to the target cold start feature vector. The preset condition means that the iteration error is less than a preset error threshold.

3. The method according to claim 2, characterized in that, The preset iterative algorithm is as follows: in, This indicates that the i-th cold start feature vector is subjected to the i-th cold start feature vector space in the cold start model. The feature vector obtained from the nth iteration This indicates that the i-th cold start feature vector is subjected to the i-th cold start feature vector space of the cold start model. The eigenvectors obtained from the nth iteration, where i is any positive integer from 1 to N. For any positive integer greater than 1, This indicates the step size for each iteration. The gradient represents the probability distribution obtained based on the feature vector space of the cold start model.

4. The method according to claim 3, characterized in that, The preset conditions are: Where i represents the i-th cold start feature vector among the N cold start feature vectors, and i is any positive integer from 1 to N. The gradient represents the probability distribution obtained based on the feature vector space of the cold start model. This represents the preset error threshold.

5. The method according to claim 1, characterized in that, The step of determining the target hot start feature vector based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors includes: Calculate the M similarities between the M cold start feature vectors and the cold start feature vectors of the cold start product, and use the M similarities as M weight values; Calculate the M differences between the M cold start feature vectors and their corresponding hot start feature vectors; The target hot-start feature vector is determined based on the M weight values ​​and the M differences.

6. The method according to claim 5, characterized in that, The formula for calculating the target hot-start feature vector is as follows: in, This represents the target hot-start feature vector. This represents the cold start feature vector of the cold start product. This represents the j-th cold start feature vector among the M cold start feature vectors. This represents the warm start feature vector corresponding to the j-th cold start feature vector. This represents the similarity between the j-th cold start feature vector and the cold start feature vector of the cold start product. This represents the difference between the j-th hot-start feature vector and the j-th cold-start feature vector.

7. A click pass rate prediction device for a cold start product, characterized in that, The device includes: The first acquisition module is used to acquire the ranking results of the cold start products to be predicted in the cold start model; The second acquisition module is used to acquire the target hot start feature vector after the cold start feature vector of the cold start product is mapped to the hot start model. The cold start feature vector of the cold start product refers to the feature vector of the cold start product in the cold start model, and the target hot start feature vector refers to the feature vector of the cold start product in the hot start model after it is mapped to the hot start product. The target hot start feature vector is used to characterize the user interaction features of the cold start product after it is mapped to the hot start product. The prediction module is used to input the ranking results and the target hot start feature vector into the hot start model to predict the click pass rate of the cold start product; The second acquisition module includes: The acquisition submodule is used to map N cold start feature vectors into N hot start feature vectors, where N is a positive integer greater than 1; The selection submodule is used to select M cold start feature vectors from the N cold start feature vectors based on a preset optimization equation, where M is a positive integer greater than 1 and less than N, and the M cold start feature vectors are the M feature vectors from the N cold start feature vectors that are similar to the cold start feature vectors of the cold start product. The determination submodule is used to determine the target hot start feature vector based on the M cold start feature vectors and the M hot start feature vectors corresponding to the M cold start feature vectors, wherein the N hot start feature vectors include the M hot start feature vectors; Both the cold start model and the hot start model include a multi-hot feature compression layer, a deep neural network, and an activation function. The multi-hot feature compression layer is used to transform input features into compressed multi-hot features. The deep neural network is used to transform the compressed multi-hot features into feature vectors in the feature vector space. The activation function, together with the position bias and the feature vectors in the feature vector space, determines the ranking result of each product in the recommendation list. The input features of the cold start model include user features and product features, and the input features of the hot start model include user features, product features, and user-product interaction features.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of the click pass rate prediction method for a cold start product as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the click pass rate prediction method for cold start products according to any one of claims 1-6.

Images