An adaptive exponential decay method for crowdsourcing task recommendation

Abstract

The present invention discloses an adaptive exponential decay method for decaying the offline model prediction value of the recommended task for the worker according to the real-time competition situation in the crowdsourcing scenario. Based on the distribution difference of the prediction value of different offline prediction models, the present invention designs an algorithm for adaptively adjusting the exponential decay parameter according to the local distribution and global distribution of the offline prediction value. In this model, the global distribution determines the upper limit of the exponential decay intensity, and the local distribution determines whether this upper limit can be reached, wherein the absolute distribution of the local distribution is used to determine the increase or decrease of the initial local distribution intensity, and the relative distribution determines the intensity of the increase or decrease. After adjusting the hyperparameters, the model has the ability to generalize between different offline models. This model stabilizes the problem of abnormal fluctuations in the indicators related to the recommendation list caused by the non-independence between workers in the crowdsourcing scenario.

Description

1. Technical Field

[0001] The present invention relates to the field of recommendation algorithms, in particular to task recommendation under crowdsourcing competition, and specifically to an adaptive parameter exponential decay method based on offline prediction value distribution. 2. Background Technology

[0002] Crowdsourcing, as an innovative business model, promotes open innovation within groups. The recent growth of the sharing economy has also fueled the rapid development of crowdsourcing. Applying recommendation algorithms to crowdsourcing can help contractors (often called workers) select bids for tasks with a high probability of crowdsourcing, reduce screening costs for publishers, and improve the user experience for both workers and contractors. However, recommendation algorithms in other scenarios cannot be directly applied to crowdsourcing. Crowdsourcing platforms are unique in their multi-role user relationships, cascading user behaviors, task timeliness, and user non-independence. The key challenge lies in addressing the issue of abnormal fluctuations in recommendation algorithm metrics caused by user non-independence, which primarily manifests itself in competition between workers. However, existing algorithms fail to account for user non-independence in crowdsourcing tasks. Existing recommendation algorithm frameworks can be categorized based on input relationships into single-point learning-to-rank methods, pairing methods, and list methods. In single-point learning-to-rank methods, each training example is simply a pairing of a query keyword (i.e., a user in the recommendation algorithm) and a document (i.e., an item in the recommendation algorithm). Whether users are related to each other has nothing to do with other documents or other query keywords. In crowdsourcing scenarios, both workers and publishers are assumed to be independent. The pairing method compares samples pairwise, constructs partially ordered document pairs, learns ranking from the comparisons, and combines any two documents to form a document pair as an input sample. However, the pairing method considers the correlation between document pairs (in this case, tasks), not between users. The list method, on the other hand, considers the difference between the current document ranking and the optimal ranking, also regarding document pairs (in this case, tasks), not between users. Therefore, the three existing ranking learning methods cannot address the competition problem between workers. Existing attenuation applications all improve the basic attenuation function based on the scenario. This requires incorporating scenario factors into the attenuation function based on different known quantities (time offset) or different known data forms (seismic waves, water waves). However, the usage scenarios differ significantly from the crowdsourcing scenario of the present invention, and no previous work has constructed an attenuation function for such scenarios. Therefore, after obtaining a prediction using an offline model, the present invention uses an adaptive parameter attenuation algorithm in the online component to determine the attenuation strength based on the competition intensity to obtain an online prediction. This method not only leverages the excellent representation and learning capabilities of neural networks but also captures real-time online competition.

[0003] Before machine learning and deep learning were widely used in recommendation algorithms, some studies used decay functions to capture the changes in user interests or item attributes over time, as a correction or supplementary means, combined with traditional methods such as user scoring and collaborative filtering to model and achieve recommendations. Medo M et al. proposed an adaptive model that combines the similarity of user rating patterns with the popular spread of news on the evolving network, measures the similarity between users through positive and negative feedback from users on news, and achieves novel news recommendations by subsequently introducing the duration decay of the recommendation score (Medo M, Zhang Y, Zhou T, et al. Adaptive model for recommendation of news [J]. EPL, 2009, 88 (3)). Yang Li et al. Considering that user preferences change over time, they introduced time decay and preference fluctuation to jointly obtain user preferences. The decay function is defined by the decay rate and decay increment (Yang Li, Hu Yunhong, Shao Guirong. Collaborative preference acquisition method integrating time decay and preference fluctuation [J]. Computer Applications, 2016, 36 (07): 2011-2015.). Anelli VM et al. introduced item popularity into collaborative filtering algorithms and proposed the TimePop algorithm. This algorithm, which considers only the popularity of items among users' neighbors, rather than global popularity, avoids using time windows when selecting candidates and avoids selecting candidates with a fixed number of tags (Anelli VW, Di Noia T, Di Sciascio E, et al. Local popularity and timein top-n recommendation [C] / / European Conference on Information Retrieval. Springer, Cham, 2019: 861-868.). In recent years, a small number of studies have combined machine learning and decay functions for recommendation. Liu H et al. analyzed user tagging behavior, verifying whether the frequency, time, and sequential position of tags reflect user interests and how they change with these three characteristics. To reveal the decay pattern of user interests, it is assumed that user activities are periodic and user interests are both stable and variable. Nonlinear least squares was used to fit the decay process of each user, and five decay functions were tried (Liu H. Resource recommendation via user tagging behavior analysis [J]. Cluster Computing, 2019: 1-10.). Zheng Q et al. proposed an improved collaborative filtering algorithm based on expert trust and time decay.The predicted value of a target user is obtained by combining the user's weighted rank and a time decay factor, which is used to reflect changes in user interests (Zheng Q. An Improved Collaborative Filtering Algorithm Based on Expert Trust and Time Decay [C] / / 2018 11th International Symposium on Computational Intelligence and Design (ISCID). IEEE, 2018, 2:12-15.). Peng D et al. also improved a single forgetting function algorithm by assuming that user interests in social networks constantly change (Peng D, Li Y, Zhou H, et al. Time Decay Friend Recommender System for Social Network [C]. testbeds and research infrastructures for the development of networks and communities, 2018.). However, the present invention, based on existing prediction results, attenuates or amplifies the predicted value according to real-time competition. Previous work has not yet constructed a decay function for this scenario. Therefore, the present invention will also improve the basic exponential decay function based on the basic decay function and crowdsourcing scenarios and known predicted values. 3. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of abnormal fluctuations in related indicators of the recommendation list caused by user non-independence in crowdsourcing task recommendation. On the basis of obtaining offline prediction values ​​from an offline model, the attenuation factor of an exponential decay function is determined according to the real-time competition situation, that is, the global and local distribution of offline prediction values ​​in the recommendation list, and the offline prediction values ​​are attenuated to obtain online prediction values.

[0005] The object of the present invention is achieved in this way:

[0006] When performing attenuation, the attenuation or gain parameter is determined from two perspectives: the global distribution of predicted values ​​in bid records that are higher than the current record, and the local distribution. The global distribution refers to the predicted value distribution for the entire list when generating a recommendation list for a worker, while the local distribution refers to the predicted value distribution for a specific task in the recommendation list. Since only the current and bid predicted values ​​are known when generating the recommendation list, competition is reflected only based on bid records. Finally, the attenuation parameter is corrected using the predicted values ​​in bid records that are lower than the current record.

[0007] The specific steps are:

[0008] The present invention relates to the same statistical concept as Figure 1 As shown, in summary, the attenuation strength is determined by the global distribution and local distribution of the predicted value. The global distribution determines the upper bound of the strength and whether the local distribution can reach this upper bound. The absolute distribution is used to determine the increase or decrease in the initial local distribution strength, and the relative distribution determines the strength of the increase or decrease. Depend on get.

[0009] (1) Global attenuation parameters

[0010] As a global attenuation factor, it is the influence of the distribution of predicted values ​​in the entire recommendation list on the attenuation strength. Because the purpose of attenuation is to adjust the predicted value s of each task in a worker's score according to the competition situation. l The relative position ranking is adjusted, and α global As a global attenuation factor, it determines the upper limit of the attenuation strength of the change in relative ranking position, so α global Determined by the overall distribution of the score, when the score value is more discrete, a greater attenuation strength α is required global In order to affect the relative position ranking in the score.

[0011] Here we use the score quartile Q = Q1-Q3 as α global The interquartile range is mainly used to measure the dispersion of sequential data, i.e., the predicted values ​​of a worker's recommendation list. The interquartile range reflects the dispersion of the middle 50% of the data and is the difference between the upper quartile (Q3, i.e., at 75%) and the lower quartile (Q1, i.e., at 25%). The interquartile range can avoid the influence of extreme values ​​because after prediction by the neural network offline model, the minimum and maximum extreme values ​​in each recommendation list tend to be 0 or 1, which does not reflect the dispersion of the data. global It can make the task l The maximum offset of the list position caused by attenuation is 50%, that is, attenuation from Q3 to Q1. Here we define α global The intensity required to reduce Q3 to Q1 using exponential decay is given by As a result, a is the suppression factor as a hyperparameter, when the predicted value is increased, α global Suppression is performed to prevent the interquartile range of the predicted values ​​generated by the model from being too large. a remains unchanged after the initial adjustment within all epochs of a model.

[0012] (2) Local attenuation parameters

[0013] d local,lAs a local attenuation factor in each task, it represents the predicted score of other workers in the lth task l =[s l ,s l1 ,s l2 ,...,s lm ] on the attenuation intensity, m is the predicted value of the bidding records of other workers in the lth task. local,l By d local,l =f(l,d,d EM ) is determined by the parameters l, d and d EM The function of will be described below. The local attenuation in the present invention is for s lm >s l The present invention measures the degree of competition within the task from two perspectives to determine the local attenuation parameter d local,l , (1) The number of bids l is obtained by parameter Δ l ,(2) The absolute distribution of the bid prediction value within the task and the relative distribution of the prediction value between the lists determine the parameters d and d EM , to correct Δ l The purpose of local decay is to make d local,l Tends to 1 to achieve the position of the current task in the recommendation list from Q 3 Offset to Q 1 attenuation strength.

[0014] The number of bids determines the parameter Δ l :

[0015] First of all, regardless of score l =[s l ,s l1 ,s l2 ,...,s lm ]The size and distribution of the score value have an impact on s l The impact of a task is greater than s l The number of bidding behaviors of the predicted value will intensify the competition of the current task, which will lead to the l Attenuation, since the upper limit of the attenuation intensity is determined by the global distribution A global , so we hope that the Δ determined by the bid quantity l When the competition is the strongest, it can approach 1, so that the final attenuation intensity is between 0 and α global Between. Δ l By Δ l =tanh(l0×l×l offset ) get, l offset It is used to map the bid number to a smaller interval, which is determined by the characteristics of the dataset, that is, the range of the bid number. offsetThe magnitude and initial value l0. l0 is a hyperparameter defined according to the magnitude of the dataset bid to map l to a range that makes tanh meaningful.

[0016] Parameters d, d EM :

[0017] In addition to the effect of the number of bids on the attenuation intensity, score l =[s l ,s l1 ,s l2 ,...,s lm ] will also affect the current attenuation strength and be used to correct l0. Under the same number of bids, score l Different distributions will also result in different attenuation strengths.

[0018] First, we need to calculate the predicted value distribution and the current predicted value s l The distance determines the parameter d, where d represents s l Distance score l The distance between two standard deviations d∈(-1,1), such as Figure 2 As shown, d is determined by score l The degree of centralization and decentralization jointly determine the Figure 2 It is obvious that the degree of competition caused by the right part is greater than the predicted value of the left part, so d2>0>d1 is obtained.

[0019] In addition to the distribution and the current predicted value s l In addition to the distance d, the shape of the distribution itself will also cause differences in competition intensity. However, d only includes the mean and variance of the data and cannot reflect the distribution shape. It is possible that the mean and variance are the same but the competition intensity is different, because the predicted value does not necessarily conform to the Gaussian distribution, such as Figure 3 (This is just a schematic diagram, not the actual value). The distributions have the same mean and variance, that is, d is the same, but the distribution to the right obviously indicates greater competition.

[0020] Therefore, it is necessary to strengthen or suppress d according to the distribution form. In order to judge whether to strengthen or suppress and conduct quantitative analysis, it is necessary to use a certain distribution as a benchmark. Here, the current recommendation list is used in the score l The predicted value distribution score between extreme values list [minima:maxima] is the benchmark, Figure 5 and Figure 6 It can be seen that the distribution of the recommendation list is approximately the same as the distribution of the current model at its epoch, and thus the stability within the model can be used as a benchmark. Then the EMD distance is used to calculate the distribution within the task, score l to score list The distance from Get, Π(score l ,score list ) is the set of all possible joint distributions of the two distributions. For each possible joint distribution γ, sample (x, y):γ is used to obtain the sample x, y and calculate the distance between the two samples, as well as the expectation E of all sample distances. (x,y)γ [||xy||], the lower bound of this expected value in all possible joint distributions is score l to score list EM distance.

[0021] In order to avoid the influence of the predicted value dimension on W, score is first calculated before W is calculated. l and score list Normalization is performed. Because the dimension of the local prediction value has nothing to do with the attenuation intensity caused by the distribution of the prediction value, for example, score l Including l Together 1 -3 Magnitude and score l Including l Together 1 -1 Magnitude, s l The intensity of competition is the same.

[0022] Subsequently, since the EM distances are all greater than zero, it is impossible to determine whether to strengthen or suppress d. Here, score l and score list The mean difference is used as the criterion for strengthening or suppressing, positive means strengthening and negative means suppressing, that is, Figure 4 m1-m and m2-m. Finally, the attenuation caused by the distribution of the predicted value Finally, by d and d EM Joint decision offset , The improved sigmoid function is used here.

[0023] The correction to l0 is determined by d, whether it is strengthened or weakened. On this basis, d EM The sigmoid function can be used to avoid the occurrence of abnormal values ​​that may cause l0 to change too much. EM )∈(0.5,1.5). μ1,μ2 are used as hyperparameters to scale and correct the overall strength.

[0024] Finally, we use the symmetric method to calculate s lm l The gain part here is not the gain of s l rather than improving lm ​>s l The obtained d local,l Weaken, for example, even if the competition intensity is the same, s l Ranking in the top 10% and bottom 10% in the task, s l The probability of winning the bid is obviously different, that is, s should be used lm >s l Part of d local,l Adjust the gain to d local,l And the present invention believes that the importance of attenuation is greater than that of gain in competition, so the local distribution gain parameter g is obtained in the symmetric method local,l Finally, add the penalty ratio.

[0025] Get g local,l After using g local,l with d local,l The ratio r l Right local,l To weaken, use the sigmoid function to l After smoothing, cosine decay is used to obtain the final local attenuation parameter d local,l , d local,l =d local,l ×cos(2×(1 / (1+(e -rl ) β )-0.5)) where β is the hyperparameter control gain part of d local,l The purpose of using the cosine function is to weaken to the maximum, that is, r l =1 when d local,l It will not be weakened to 0, to ensure that it is mainly attenuated during competition.

[0026] The positive effects of the present invention are:

[0027] 1. Utilize offline predictions from neural networks or machine learning, and then attenuate them online based on real-time competition intensity. This approach leverages the neural network's excellent representation and learning capabilities while capturing real-time competition online.

[0028] 2. The attenuation intensity of the exponential decay of offline prediction values ​​is determined by the global and local distribution of offline prediction values, which alleviates the abnormal fluctuation problem of indicators related to the recommendation list caused by the non-independence of workers.

[0029] 3. After adjusting the hyperparameters, it has a certain generalization ability between different offline prediction models. IV. Description of the Figures

[0030] Figure 1 It is the overall definition diagram of the attenuation factor of the present invention.

[0031] Figure 2 is the absolute distribution of offline predicted values ​​of bids within a task.

[0032] Figure 3 They are different distribution forms with the same mean and variance.

[0033] Figure 4 is the absolute distribution of offline predicted values ​​of bids within a task.

[0034] Figure 5 It is the distribution of prediction values ​​of different offline models in each Epoch.

[0035] Figure 6 It is the offline prediction value distribution of the same offline model among recommendation lists.

[0036] Figure 7 This is a comparison of the HR@10 index of the offline model OPCE-CF before and after using the present invention.

[0037] Figure 8 This is a comparison of the MRR indicators of the offline model OPCE-CF before and after using the present invention.

[0038] Figure 9 This is a comparison of the HR@10 index of the offline model NAIS before and after using the present invention.

[0039] Figure 10 This is a comparison of the MRR indicators of the offline model NAIS before and after using the present invention.

[0040] Figure 11 This is a comparison of the HR@10 index of the offline model NAIS+Attentive pooling network before and after using the present invention.

[0041] Figure 12 This is a comparison of the MRR indicators of the offline model NAIS+Attentive pooling network before and after using the present invention.

[0042] Figure 13 It is the overall process of exponential decay of offline model prediction values. V. Specific Implementation Methods

[0043] The present invention quantitatively analyzes real-time competition intensity by analyzing the local and global distribution of offline prediction values ​​within tasks and recommendation lists, and introduces local and global attenuation factors to exponentially decay the offline prediction values. This divides the crowdsourcing task into an offline prediction phase using neural networks or machine learning and an online prediction phase using the present invention, resolving the problem that existing ranking recommendation algorithms cannot handle the non-independence between workers. After adjusting hyperparameters, the present invention demonstrates a certain degree of generalization across different offline models. The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0044] In the first step, the set of candidate tasks that are valid at the login time is determined by the worker login time as an unsorted candidate recommendation list, and the offline prediction value of the current bid is obtained according to the login time. The offline prediction value score of the current worker in the candidate recommendation list is [s1,s1...,s l ], l is the number of tasks in the list.

[0045] In the second step, the global attenuation factor is obtained using the interquartile range of the score. The interquartile range is mainly used to measure the discreteness of sequential data, that is, the predicted value of a worker's recommendation list, and determines the upper limit of the exponential decay strength.

[0046] The third step is to use the task lm >s l The number of bids determines the attenuation strength benchmark of the local attenuation factor, and maps the number of bids l to the range where the tanh function is effective, d local,l As a local attenuation factor in each task, it represents the predicted score of other workers in the lth task l =[s l ,s l1 ,s l2 ,...,s lm ] on the attenuation intensity, and m is the predicted value of the bidding records of other workers in the lth task.

[0047] The fourth step is to modify l0 using the absolute and relative distributions of the offline prediction values ​​within the task. Let d represent s l Distance score l The distance between two standard deviations is d∈(-1,1). In addition to the distribution and the current predicted value s l In addition to the distance d, the distribution shape itself will also cause differences in competition intensity, and d only includes the mean and variance of the data and cannot reflect the distribution shape. Therefore, it is necessary to strengthen or suppress d according to the distribution shape. In order to judge whether to strengthen or suppress and conduct quantitative analysis, it is necessary to use a certain distribution as a benchmark. Here, the current recommendation list is used in the score l The predicted value distribution score between extreme values list [minima:maxima] is used as the benchmark, and the EMD distance is used to calculate the distribution score within the task l to score list The distance W. Finally, d and d EM Together they determine the local attenuation strength.

[0048] Step 5: Calculate s using a symmetrical method lm l ​The gain part, and the present invention believes that the importance of attenuation is greater than the gain in the competition, so the local distribution gain parameter g is obtained in the symmetric method local,l Finally, add the penalty ratio.

[0049] The sixth step is to search for hyperparameters for different offline models, including hyperparameter a for suppressing global attenuation and β for controlling the gain part to d. local,l The maximum suppression ratio, l0 is used to map the number of bids, and μ1, μ2 are used as hyperparameters to scale the overall correction strength.

[0050] Figure 7 and Figure 8 The figures are respectively the effects of HR@10 and MRR of the offline model OPCA-CF before and after using the present invention. It can be seen that the present invention can significantly alleviate the problem of abnormal fluctuation of related indicators of the recommendation list. Figure 9 and Figure 10 The results of HR@10 and MRR of the offline model NAIS before and after adjusting the hyperparameters. Figure 11 and Figure 12 The HR@10 and MRR results of the offline model NAIS+Attentive Pooling Network before and after adjusting the hyperparameters are shown. It can be seen that the present invention has a generalization effect among offline models after adjusting the hyperparameters. Figure 13 It is the overall process of exponential decay of offline model prediction values.

[0051] Examples of user usage scenarios of the present invention:

[0052] Scenario 1: When recommending suitable tasks to task undertakers in a crowdsourcing scenario, in addition to the undertaker's own capabilities and interests, the real-time competition within the task can also determine whether the undertaker's bid will be successful. In this scenario, the present invention can be used to quantify the real-time competition intensity.

[0053] Scenario 2: In a recommendation scenario where there is competition among users, the prediction value generated by the neural network model is attenuated according to the real-time competition intensity.

[0054] In the above scenario, the method of the present invention achieves good results.

Claims

1. An adaptive exponential decay method for crowdsourcing task recommendation, characterized by: In crowdsourcing scenarios, we use offline model predictions and real-time bidding to quantitatively analyze the competition intensity of the current task. We use the global and local distributions of the predictions to derive the attenuation parameter to attenuate the existing predictions. Attenuation and gain are two symmetrical processes, and the attenuation intensity is determined by both the global and local distributions of the predictions. The global distribution determines the upper bound of the intensity, while the local distribution determines whether this upper bound can be reached. The absolute distribution is used to determine the increase or decrease in the initial local distribution intensity, while the relative distribution determines the intensity of the increase or decrease. Finally, we impose a certain penalty on the attenuation parameter and apply it to the gain calculation. This is because whether a worker wins the bid is mainly affected by the degree of competition, and the attenuation parameter is a key indicator reflecting the current real-time competition situation. The specific approach is as follows: In the first step, the set of candidate tasks valid at the login time is determined by the worker login time as the unsorted candidate recommendation list. The offline prediction value of the current bid is obtained according to the login time. The offline prediction value of the current worker in the candidate recommendation list is score = [s1, s1..., s l ], l is the number of tasks in the list; the global attenuation factor α is obtained by using the quartile difference of score global , α global Indicates that the task can be used to l The maximum offset of the list position caused by attenuation is 50%; The second step is to divide the matrix into six types of planes and determine the key that the node needs to store based on the coordinates of the node. These six types of planes are: XOY type plane, XOZ type plane, YOZ type plane, diagonal plane passing through the Z axis, diagonal plane passing through the Y axis, and diagonal plane passing through the X axis. Then, the plane where the node is located is determined based on the node's coordinates, and the corresponding plane key and a pseudo-random function with six parameters are stored. The third step is to use the task lm >s l The number of bids determines the attenuation strength benchmark of the local attenuation factor, and maps the number of bids l to the range where the tanh function is effective, d local,l As a local attenuation factor in each task, it represents the predicted score of other workers in the lth task l =[s l ,s l1 ,s l2 ,...,s lm ] on the attenuation intensity, m is the predicted value of the bidding records of other workers in the lth task; The fourth step is to modify l0 using the absolute and relative distributions of the offline prediction values ​​within the task, and use d to represent s l Distance score l The distance between two standard deviations d∈(-1,1) is used to calculate the distribution score within the task using EMD distance l to score list The distance W, before calculating W, first calculate the score l and score list Normalize the score. l and score list The mean difference is used as the standard for strengthening or suppressing, positive means strengthening and negative means suppressing, d EM =sigmoid(mean(score l )-mean(score list ))×e W(scorel,scorelist) ; Step 5: d and d EM Joint decision offset , Step 6: Calculate s using the symmetric method from step 1 to step 4 lm l The gain part g local,l， Get g local,l After using g local,l with d local,l The ratio r l Right local,l To weaken, use the sigmoid function to l After smoothing, cosine decay is used to obtain the final local attenuation parameter d local,l， d local,l =d local,l ×cos(2×(1 / (1+(e -rl ) β )-0.5));​ In the seventh step, the hyperparameter a is used to suppress the global attenuation, and β is used to control the gain part to d local,l The maximum suppression ratio, l0 is used to map the number of bids, and μ1, μ2 are used as hyperparameters to scale the overall correction strength.

2. The adaptive exponential decay method for crowdsourcing task recommendation according to claim 1, wherein: The global attenuation factor is obtained by using the offline prediction value distribution in the recommendation list, and the local attenuation factor is obtained by using the offline prediction value distribution within the task: The global attenuation factor αglobal uses the score quartile Q = Q1-Q3 as the benchmark for αglobal. αglobal can reduce the number of tasks within the task. l The maximum offset of the list position caused by attenuation is 50%, The number of bids in the local attenuation factor determines the parameter Δl. The Δl determined by the number of bids can approach 1 when the competition is the strongest. Δl = tanh(l0×l×l offset ), l offset It is used to map the bid number to an interval that makes the calculated final attenuation strength between 0 and αglobal, which is determined by the characteristics of the dataset, that is, the range of the bid number l offset The magnitude and initial value l0 of the dataset; l0 is a hyperparameter defined according to the magnitude of the dataset bid to map l to a range that makes tanh meaningful; The absolute distribution of the bid prediction value within the task in the local decay factor determines the parameter d, which is determined by the distribution of the prediction value and the current prediction value s l The distance determines the parameter d, where d represents s l Distance score l The distance between two standard deviations d∈(-1,1); The relative distribution of predicted values ​​between lists in the local attenuation factor determines the parameter d EM , according to the distribution of the predicted value within the task, d is strengthened or suppressed, and the EMD distance is used to calculate the distribution within the task, score l to score list The distance from Calculated, and in order to avoid the impact of the predicted value dimension on W, score is first calculated before calculating W l and score list Normalization is performed. Since the EM distance is greater than zero, it is impossible to determine whether to strengthen or suppress d. Here, score l and score list The mean difference is used as the standard for strengthening or suppressing. If it is positive, it will strengthen, and if it is negative, it will suppress. Finally, d EM =sigmoid(mean(score l )-mean(score list ))×e W(scorel,scorelist) .

3. The method according to claim 2, characterized in that The attenuation or gain parameters are determined from the distribution of predicted values ​​from two perspectives: the global distribution and the local distribution of predicted values. The global distribution is the distribution of predicted values ​​for the entire list when generating a recommendation list for a worker, and the local distribution is the distribution of predicted values ​​for a certain task in the recommendation list. When generating the recommendation list, since only the current and bid predicted values ​​are known, the competition situation is reflected only based on the bid records. The global distribution determines the upper bound of the strength, and the local distribution determines whether this upper bound can be reached. The absolute distribution is used to determine the increase or decrease in the strength of the initial local distribution, and the relative distribution determines the strength of the increase or decrease.

Images