Adversarial bandits policy for crawling highly dynamic content
Techniques of generating recrawl policies for commercial offer pages include generating a multiple strategy approach using a number of different strategies. In some implementations, each strategy is an arm of a K-armed adversarial bandits algorithm with reinforcement learning. Moreover, in some implementations, the multiple strategy approach also uses a machine learning algorithm to estimate parameters such as a click rate, impression rate, and likelihood of price change, i.e., change rate, which was assumed known in the conventional approaches.
1 . A method, comprising:
receiving, from a repository, a plurality of entities, each of the plurality of entities having a respective value of a quantity that is accurate at a previous time step;
for each of the plurality of entities, generating associated values of a plurality of parameters at a current time step, the plurality of parameters including at least one of an access rate of that entity from the repository or a likelihood of a change in the respective value of the quantity of that entity;
selecting a refresh strategy of a plurality of refresh strategies for updating the respective values of the quantity for the plurality of entities according to a refresh policy, the refresh policy including a weight distribution representing a respective likelihood that each of the plurality of refresh strategies is selected, the plurality of refresh strategies including at least two of a uniform strategy, a change-weighted strategy, an access-weighted strategy, or a resource-optimized strategy;
generating a respective refresh rate for each of the plurality of entities according to the selected refresh strategy, the respective refresh rate for each entity of the plurality of entities being based on the associated values of the plurality of parameters at a sequence of times comprising the previous time step and the current time step; and
performing a refresh operation on the repository based on the respective refresh rates for the plurality of entities, the refresh operation being configured to obtain the respective values of the quantity at the current time step.
2 . The method as in claim 1 , wherein selecting the refresh strategy of the plurality of refresh strategies includes:
generating a probability distribution for the refresh strategy over the plurality of refresh strategies, the probability distribution including a respective probability corresponding to each of the plurality of refresh strategies; and
performing a random sample of the plurality of refresh strategies according to the probability distribution to produce the selected refresh strategy.
3 . The method as in claim 2 , wherein generating the probability distribution includes:
performing an average of a weight of the weight distribution and a reciprocal of a number of refresh strategies of the plurality of refresh strategies, the weight corresponding to the refresh strategy.
4 . The method as in claim 1 , wherein generating the respective refresh rate for each of the plurality of entities according to the selected refresh strategy includes:
for a parameter of the plurality of parameters, generating a respective neural network model corresponding to the parameter; and
generating the respective refresh rate for each of the plurality of entities using the respective neural network model corresponding to the parameter.
5 . The method as in claim 4 , wherein the parameter of the plurality of parameters is the likelihood of a change in the respective value of the quantity of an entity of the plurality of entities, and
wherein generating the respective neural network model corresponding to the parameter includes:
training a model based on a set of history features, the set of history features including at least one of a quantity change frequency in a previous time period and a length of time since most recent change.
6 . The method as in claim 4 , wherein the parameter of the plurality of parameters is the access rate for an entity of the plurality of entities, and
wherein generating the respective neural network model corresponding to the parameter includes:
training a model based on a set of history features, the set of history features including a number of accesses over a previous time period.
7 . The method as in claim 4 , wherein generating the respective neural network model corresponding to the parameter includes:
training a model based on metadata, the metadata including at least one of a day of a week for a prediction time, and a characteristic of each of the plurality of entities.
8 . The method as in claim 4 , wherein:
the repository includes a plurality of offer web pages;
each of the plurality of entities includes an offer web page of the plurality of offer web pages, the offer web page featuring a product offer;
the refresh operation for an entity includes recrawling that web page from a merchant web site;
the plurality of parameters for an entity of the plurality of entities include an impression rate of the offer web page and a click rate of the offer web page;
each of the plurality of entities including a brand identifier of that offer web page, a merchant identifier of that offer web page, and a country identifier of that offer web page, and
generating the respective neural network model corresponding to the parameter includes:
training a model based on metadata, the metadata including at least one of the brand identifier, the country identifier, a day of a week for a prediction time, and the merchant identifier.
9 . A computer program product comprising a nontransitory storage medium, the computer program product including code that, when executed by processing circuitry of a user device configured to perform a method of generating a refresh policy, the method comprising:
receiving, from a repository, a plurality of entities, each of the plurality of entities having a respective value of a quantity that is accurate at a previous time step;
for each of the plurality of entities, generating associated values of a plurality of parameters at a current time step, the plurality of parameters including at least one of an access rate of that entity from the repository or a likelihood of a change in the respective value of the quantity of that entity;
selecting a refresh strategy of a plurality of refresh strategies for updating the respective values of the quantity for the plurality of entities according to a refresh policy, the refresh policy including a weight distribution representing a respective likelihood that each of the plurality of refresh strategies is selected, the plurality of refresh strategies including at least two of a uniform strategy, a change-weighted strategy, an access-weighted strategy, and a resource-optimized strategy;
generating a respective refresh rate for each of the plurality of entities according to the selected refresh strategy, the respective refresh rate for an entity of the plurality of entities being based on the associated values of the plurality of parameters at a sequence of times comprising the previous time step and the current time step; and
performing a refresh operation on the repository based on the respective refresh rates for the plurality of entities, the refresh operation being configured to obtain the respective value of the quantity at the current time step.
10 . The computer program product as in claim 9 , wherein a sum of the respective refresh rates for the plurality of entities is normalized based on an entity refresh budget constraint.
11 . The computer program product as in claim 9 , wherein each of the plurality of refresh strategies is represented as an arm of a K-armed adversarial bandits algorithm.
12 . The computer program product as in claim 9 , further comprising:
generating a weight of the weight distribution using reinforcement learning to maximize a value of a reward parameter, the weight representing a likelihood that a refresh strategy of the plurality of refresh strategies is selected, the reward parameter indicating a relative usefulness of the refresh strategy.
13 . The computer program product as in claim 12 , wherein generating the weight of the weight distribution includes:
generating an exploration probability representing a likelihood that the refresh strategy is not chosen solely due to historical data.
14 . The computer program product as in claim 9 , wherein selecting the refresh strategy of the plurality of refresh strategies includes:
generating a probability distribution for the refresh strategy over the plurality of refresh strategies, the probability distribution including a respective probability corresponding to each of the plurality of refresh strategies; and
performing a random sample of the plurality of refresh strategies according to the probability distribution to produce the selected refresh strategy.
15 . The computer program product as in claim 9 , wherein generating the respective refresh rate for each of the plurality of entities according to the selected refresh strategy includes:
for a parameter of the plurality of parameters, generating a respective neural network model corresponding to the parameter; and
generating the respective refresh rate for each of the plurality of entities using the respective neural network model corresponding to the parameter.
16 . An electronic apparatus configured to generate a refresh policy, the electronic apparatus comprising:
memory; and
controlling circuitry coupled to the memory, the controlling circuitry being configured to:
receive, from a repository, a plurality of entities, each of the plurality of entities having a respective value of a quantity that is accurate at a previous time step;
for each of the plurality of entities, generate associated values of a plurality of parameters at a current time step, the plurality of parameters including at least one of an access rate of that entity from the repository or a likelihood of a change in the respective value of the quantity of that entity;
select a refresh strategy of a plurality of refresh strategies for updating the respective values of the quantity for the plurality of entities according to a refresh policy, the refresh policy including a weight distribution representing a respective likelihood that each of the plurality of refresh strategies is selected, the plurality of refresh strategies including at least two of a uniform strategy, a change-weighted strategy, an access-weighted strategy, and a resource-optimized strategy;
generate a respective refresh rate for each of the plurality of entities according to the selected refresh strategy, the respective refresh rate for each entity of the plurality of entities being based on the associated values of the plurality of parameters at a sequence of times comprising the previous time step and the current time step; and
perform a refresh operation on the repository based on the respective refresh rates for the plurality of entities, the refresh operation being configured to obtain the respective value of the quantity at the current time step.
17 . The electronic apparatus as in claim 16 , wherein the controlling circuitry configured to select the refresh strategy of the plurality of refresh strategies is further configured to:
generate a probability distribution for the refresh strategy over the plurality of refresh strategies, the probability distribution including a respective probability corresponding to each of the plurality of refresh strategies; and
perform a random sample of the plurality of refresh strategies according to the probability distribution to produce the selected refresh strategy.
18 . The electronic apparatus as in claim 17 , wherein the controlling circuitry configured to generate the probability distribution is further configured to:
perform an average of a weight of the weight distribution and a reciprocal of a number of refresh strategies of the plurality of refresh strategies, the weight corresponding to the refresh strategy.
19 . The electronic apparatus as in claim 16 , wherein the weight distribution includes a distribution of weights, each of the distribution of weights corresponding to a respective refresh strategy of the plurality of refresh strategies.
20 . The electronic apparatus as in claim 19 , wherein the controlling circuitry is further configured to:
generate a weight of the weight distribution using reinforcement learning to maximize a value of a reward parameter, the weight representing a likelihood that a refresh strategy of the plurality of refresh strategies is selected, the reward parameter indicating a relative usefulness of the refresh strategy.