Selecting a training strategy for training a machine learning model
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a machine learning model. One of the methods includes performing experiments to select a training strategy for use in training the model on a particular data set. The selected training strategy includes a binning strategy for binning the raw feature vectors before the raw feature vectors are provided to the predictive model.
1. A method performed by one or more computers, the method comprising:
receiving instructions to train a predictive model on a particular data set, wherein the particular data set includes a plurality of raw feature vectors;
performing experiments to select a training strategy for use in training the predictive model on the particular data set, wherein the selected training strategy includes a binning strategy for binning the raw feature vectors before the raw feature vectors are provided to the predictive model, and wherein performing experiments comprises:
selecting a plurality of training strategies, wherein each of the plurality of training strategies includes a respective binning strategy;
performing a respective training experiment for each of the plurality of training strategies by training the model using the training strategy;
identifying a best performing training strategy from the plurality of training strategies based on results of the respective training experiments for each of the plurality of training strategies; and
generating a first new training strategy from the best performing training strategy by adjusting the binning strategy included in the best performing training strategy based on a measure of entropy of one or more features during the training of the model using the best performing training strategy, comprising, for a particular feature in the raw feature vectors, determining whether to decrease a number of bins for the particular feature based on an absolute value of a measure of entropy of the particular feature during the training of the model using the best performing training strategy; and
training the predictive model on the particular data set using the selected training strategy.
2. The method of claim 1 , wherein performing experiments further comprises:
determining that termination criteria are not satisfied after performing the respective training experiment for each of the plurality of training strategies;
in response to determining that the termination criteria are not satisfied, generating one or more new training strategies including the first new training strategy based on the best performing training strategy; and
performing a respective training experiment for each of the new training strategies by training the model using the new training strategy.
3. The method of claim 2 , wherein performing experiments further comprises:
identifying a best performing new training strategy;
determining that the termination criteria are satisfied; and
selecting the best performing new training strategy as the training strategy for use in training the predictive model on the particular data set.
4. The method of claim 2 , wherein generating new training strategies based on the best performing training strategy comprises adjusting the best performing training strategy based on historical data that identifies binning strategies that have previously been successful for training predictive models on data sets similar to the particular data set.
5. The method of claim 2 , wherein generating new training strategies based on the best performing training strategy comprises keeping one or more rules specified by the best performing training strategy constant among each new training strategy and varying other rules among the new training strategies.
6. The method of claim 1 , wherein the selected training strategy includes a strategy for determining a makeup of sets of training data to be used to train the predictive model.
7. The method of claim 6 , wherein the strategy for determining the makeup includes a rule for determining a number of feature vectors to be included in each set and a rule for determining a ratio of new feature vectors to old feature vectors to be included in each set.
8. The method of claim 1 , wherein:
the plurality of raw feature vectors comprises vectors representing completed financial product transactions including transaction prices; and
the predictive model is a model predicting a next transaction price or a next transaction price direction for one or more financial products.
9. The method of claim 1 , wherein:
the plurality of raw feature vectors comprises vectors representing completed credit card transactions or debit card transactions or both; and
the predictive model is a model classifying particular transactions as likely being anomalous or not.
10. The method of claim 1 , wherein:
the plurality of raw feature vectors comprises vectors representing financial claims processing transactions; and
the predictive model is a model classifying particular transactions as likely being anomalous or not.
11. The method of claim 1 , wherein:
the plurality of raw feature vectors comprises vectors representing prices for products or services or both at particular times or places or both; and
the predictive model is a model predicting prices for products or services or both in particular places or on particular dates or both.
12. The method of claim 1 , wherein:
the plurality of raw feature vectors comprises vectors representing purchase transactions representing purchases of products or services or both and including respective prices paid for the products or services or both; and
the predictive model is a model predicting prices for products or services in particular places or on particular dates or both.
13. The method of claim 12 , wherein:
the predictive model is further a model classifying particular prices for particular products or services as likely being anomalous or not.
14. The method of claim 13 , wherein:
the predictive model is further a model classifying particular prices for particular products or services as likely being fraudulent or not.
15. The method of claim 1 , wherein:
the plurality of raw feature vectors comprises vectors representing user actions on an interactive computer-based system; and
the predictive model is a model predicting user actions on the interactive computer-based system.
16. The method of claim 1 , wherein adjusting the binning strategy included in the best performing training strategy based on a measure of entropy of one or more features during the training of the model using the new training strategy comprises, for the particular feature:
when the absolute value of the measure of entropy of the particular feature during the training of the model using the new training strategy is greater than a threshold value, adjusting a rule for how a membership of the bins for the particular feature is determined;
when the absolute value of the measure of entropy of the particular feature is less than or equal to the threshold value and the measure of entropy of the particular feature is positive, increasing the number of bins for the particular feature; and
when the absolute value of the measure of entropy of the particular feature is less than or equal to the threshold value and the measure of entropy of the particular feature is negative, decreasing the number of bins for the particular feature.
17. The method of claim 1 ,
wherein performing a respective training experiment for each of the plurality of training strategies by training the model using the training strategy comprises training the model using the training strategy until experiment convergence criteria are satisfied,
wherein training the model on the particular data set using the selected training strategy comprises training the model on the particular data set using the selected training strategy until actual training convergence criteria are satisfied, and
wherein the experiment convergence criteria are relaxed from the actual training convergence criteria.
18. A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
receiving instructions to train a predictive model on a particular data set, wherein the particular data set includes a plurality of raw feature vectors;
performing experiments to select a training strategy for use in training the predictive model on the particular data set, wherein the selected training strategy includes a binning strategy for binning the raw feature vectors before the raw feature vectors are provided to the predictive model, and wherein performing experiments comprises:
selecting a plurality of training strategies, wherein each of the plurality of training strategies includes a respective binning strategy;
performing a respective training experiment for each of the plurality of training strategies by training the model using the training strategy;
identifying a best performing training strategy from the plurality of training strategies based on results of the respective training experiments for each of the plurality of training strategies; and
generating a first new training strategy from the best performing training strategy by adjusting the binning strategy included in the best performing training strategy based on a measure of entropy of one or more features during the training of the model using the best performing training strategy, comprising, for a particular feature in the raw feature vectors, determining whether to decrease a number of bins for the particular feature based on an absolute value of the measure of entropy of the particular feature during the training of the model using the best performing training strategy; and
training the predictive model on the particular data set using the selected training strategy.
19. The system of claim 18 , wherein performing experiments further comprises:
determining that termination criteria are not satisfied after performing the respective training experiment for each of the plurality of training strategies;
in response to determining that the termination criteria are not satisfied, generating one or more new training strategies including the first new training strategy based on the best performing training strategy; and
performing a respective training experiment for each of the new training strategies by training the model using the new training strategy.
20. The system of claim 19 , wherein performing experiments further comprises:
identifying a best performing new training strategy;
determining that the termination criteria are satisfied; and
selecting the best performing new training strategy as the training strategy for use in training the predictive model on the particular data set.
21. The system of claim 19 , wherein generating new training strategies based on the best performing training strategy comprises adjusting the best performing training strategy based on historical data that identifies binning strategies that have previously been successful for training predictive models on data sets similar to the particular data set.
22. The system of claim 19 , wherein generating new training strategies based on the best performing training strategy comprises keeping one or more rules specified by the best performing training strategy constant among each new training strategy and varying other rules among the new training strategies.
23. The system of claim 19 , wherein determining that termination criteria are not satisfied comprises determining that a performance of the model is likely to improve if one or more of: the binning strategy from the best performing training strategy is adjusted, additional training data is used, or new features are introduced.
24. A non-transitory computer storage medium encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:
receiving instructions to train a predictive model on a particular data set, wherein the particular data set includes a plurality of raw feature vectors;
performing experiments to select a training strategy for use in training the predictive model on the particular data set, wherein the selected training strategy includes a binning strategy for binning the raw feature vectors before the raw feature vectors are provided to the predictive model, and wherein performing experiments comprises:
selecting a plurality of training strategies, wherein each of the plurality of training strategies includes a respective binning strategy;
performing a respective training experiment for each of the plurality of training strategies by training the model using the training strategy;
identifying a best performing training strategy from the plurality of training strategies based on results of the respective training experiments for each of the plurality of training strategies; and
generating a first new training strategy from the best performing training strategy by adjusting the binning strategy included in the best performing training strategy based on a measure of entropy of one or more features during the training of the model using the best performing training strategy, comprising, for a particular feature in the raw feature vectors, determining whether to decrease a number of bins for the particular feature based on an absolute value of the measure of entropy of the particular feature during the training of the model using the best performing training strategy; and
training the predictive model on the particular data set using the selected training strategy.
25. The non-transitory computer storage medium of claim 24 ,
wherein performing experiments further comprises:
determining that termination criteria are not satisfied after performing the respective training experiment for each of the plurality of training strategies;
in response to determining that the termination criteria are not satisfied, generating one or more new training strategies including the first new training strategy based on the best performing training strategy; and
performing a respective training experiment for each of the new training strategies by training the model using the new training strategy.
26. The non-transitory computer storage medium of claim 25 , wherein performing experiments further comprises:
identifying a best performing new training strategy;
determining that the termination criteria are satisfied; and
selecting the best performing new training strategy as the training strategy for use in training the predictive model on the particular data set.
27. The non-transitory computer storage medium of claim 25 , wherein generating new training strategies based on the best performing training strategy comprises adjusting the best performing training strategy based on historical data that identifies binning strategies that have previously been successful for training predictive models on data sets similar to the particular data set.
28. The non-transitory computer storage medium of claim 25 , wherein generating new training strategies based on the best performing training strategy comprises keeping one or more rules specified by the best performing training strategy constant among each new training strategy and varying other rules among the new training strategies.
29. A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
receiving instructions to train a predictive model on a particular data set, wherein the particular data set includes a plurality of raw feature vectors;
performing experiments to select a training strategy for use in training the predictive model on the particular data set, wherein the selected training strategy includes a binning strategy for binning the raw feature vectors before the raw feature vectors are provided to the predictive model, and wherein performing experiments comprises:
selecting a plurality of training strategies, wherein each selected training strategy includes a respective binning strategy;
performing a respective training experiment for each of the plurality of training strategies by training the predictive model using the training strategy until experiment convergence criteria are satisfied;
identifying a best performing training strategy from the plurality of training strategies based on results of the respective training experiments for each of the plurality of training strategies;
determining that termination criteria for performing experiments are not satisfied after performing the respective training experiment for each of the plurality of training strategies;
in response to determining that the termination criteria are not satisfied:
generating one or more new training strategies based on the best performing training strategy, comprising generating a first new training strategy by adjusting a binning strategy from the best performing training strategy based on a measure of entropy of one or more features during the training of the model using the best performing training strategy, comprising, for a particular feature in the raw feature vectors, determining whether to decrease a number of bins for the particular feature based on an absolute value of the measure of entropy of the particular feature during the training of the model using the best performing training strategy; and
performing a respective training experiment for each of the new strategies by training the model using the new training strategy until the experiment convergence criteria are satisfied; and
training the predictive model on the particular data set using the selected training strategy until actual training convergence criteria are satisfied, wherein the experiment convergence criteria are relaxed from the actual training convergence criteria.
30. A method performed by one or more computers, the method comprising:
receiving instructions to train a predictive model on a particular data set, wherein the particular data set includes a plurality of raw feature vectors;
performing experiments on a plurality of training strategies and based on the experiments selecting a training strategy for use in training the predictive model on the particular data set,
wherein each training strategy of the plurality of training strategies includes a respective binning strategy for binning the raw feature vectors before the raw feature vectors are provided to the predictive model,
wherein each binning strategy includes, for each of a plurality of features of the raw feature vectors, a first rule for determining a number of bins for the feature and a second rule for determining membership of each of the bins for the feature,
wherein the respective binning strategy of each of the plurality of training strategies has a different first rule, a different second rule, or both from the respective binning strategy of each other training strategy of the plurality of training strategies, and
wherein performing the experiments comprises:
selecting a plurality of training strategies, wherein each of the plurality of training strategies includes a respective binning strategy;
performing a respective training experiment for each of the plurality of training strategies by training the model using the training strategy;
identifying a best performing training strategy from the plurality of training strategies based on results of the respective training experiments for each of the plurality of training strategies; and
generating a first new training strategy from the best performing training strategy by adjusting the binning strategy included in the best performing training strategy based on a measure of entropy of one or more features during the training of the model using the best performing training strategy, comprising, for a particular feature in the raw feature vectors, determining whether to decrease a number of bins for the particular feature based on an absolute value of a measure of entropy of the particular feature during the training of the model using the best performing training strategy; and
training the predictive model on the particular data set using the selected training strategy.