Smart geo-fencing using location sensitive product affinity

1. A computer-implemented method for generating personalized geo-fences that customize electronic communications with mobile devices, the computer-implemented method comprising:

determining, by a processor and from a set of taxonomy data defining a common characteristic of a set of similar products, cosine similarities between a product at a particular geographical location and additional products;

computing, by the processor, an intrinsic affinity of a particular user for the product from a weighted browse-count sum of (i) a browse count for the product and (ii) additional browse counts for the additional products, respectively, wherein each weight in the weighted browse-count sum indicates a respective cosine similarity between the product and a respective additional product,

computing, by the processor, a user-similarity-based affinity based on:

semantic vectors for geographic locations, wherein a semantic vector includes a plurality of dimensions respectively representing types of places located within a threshold distance of a geographic location and, wherein each dimension from the plurality of dimensions has a respective value representing a respective number of places having a respective type represented by the dimension, and

the intrinsic affinity of the particular user for the product,

designing, by the processor and based on the user-similarity-based affinity, a geo-fence targeted to the particular user such that a message is transmitted to a user device of the particular user in response to an indication that the particular user is within a boundary of the geo-fence, the geo-fence defining a geographical area that includes the particular geographical location and that is associated with a provider of the product; and

causing, by the processor, a telecommunication server to transmit the message to the user device when the user device is positioned within the designed geo-fence.

2. The computer-implemented method of claim 1 , wherein the geo-fence is designed based on a location-specific affinity determined from the user-similarity-based affinity, wherein the method further comprises determining the location-specific affinity by, at least:

computing the user-similarity-based affinity for the particular user from a weighted sum of intrinsic affinities for a set of users, wherein each weight of the weighted sum of the intrinsic affinities indicates a respective similarity between the particular user and a respective one of the users; and

computing the location-specific affinity from the user-similarity-based affinity and a weighted sum of additional user-similarity-based affinities for additional geographical locations wherein each weight of the weighted sum of the additional user-similarity-based affinities indicates a respective similarity between the particular geographical location and a respective one of the additional geographical locations.

3. The method of claim 1 , wherein the processor computes the user-similarity-based affinity by performing operations comprising:

generating a multi-layer hierarchy of clusters of the semantic vectors, wherein each layer of the multi-layer hierarchy includes a respective set of clusters of the semantic vectors; and

computing similarities between the particular user and users from a set of users, respectively, wherein computing each similarity comprises, for each user:

determining layer similarities between the user and the particular user based on the sets of clusters, respectively, and

determining the similarity between the user and the particular user based on a weighted sum of the layer similarities.

4. The method of claim 2 , wherein the processor further computes the location-specific affinity by computing each similarity between the particular geographical location and respective one of the additional geographical locations by computing a respective cosine similarity between a respective pair of semantic vectors representing the particular geographical location and the respective one of the additional geographical locations.

5. The method of claim 1 , wherein designing the geo-fence targeted to the particular user comprises:

computing, for the particular user, modeled location-specific affinities for geographical locations, respectively;

selecting a subset of particular geographical locations based on each selected geographical location having a respective modeled location-specific affinity greater than a threshold affinity;

grouping the subset of the particular geographical locations into clusters; and

deriving the geo-fence from a boundary around a cluster including the particular geographical location.

6. The method of claim 5 , wherein computing the modeled location-specific affinities for geographical locations comprises:

computing location-specific affinity vectors for the geographical locations and a set of users including the particular user, wherein each location-specific affinity vector comprises, for a respective one of the users, a respective distribution of location-specific affinities across the geographical locations;

segmenting the set of users by grouping the location-specific affinity vectors into location-specific affinity clusters, wherein the location-specific affinity clusters define respective segments of the users;

averaging, for a segment corresponding to the particular user, distributions of location-specific affinities in a location-specific affinity cluster corresponding to the segment; and

selecting the averaged distribution of location-specific affinities as the modeled location-specific affinities for the particular user.

7. The method of claim 1 , wherein stored user information used to compute the intrinsic affinity comprises:

a user browsing history generated from interactions by the user device with an online service providing access to the product; and

locations tracked by a global positioning system of the user device and associated with the interactions.

8. A non-transitory computer-readable medium embodying program code that, when executed by a processing device, cause the processing device to perform operations comprising:

determining, from a set of taxonomy data defining a common characteristic of a set of similar products, cosine similarities between a product at a particular geographical location and additional products;

computing an intrinsic affinity of a particular user for the product from a weighted browse-count sum of (i) a browse count for the product and (ii) additional browse counts for the additional products, respectively, wherein each weight in the weighted browse-count sum indicates a respective cosine similarity between the product and a respective additional product,

computing a user-similarity-based affinity based on:

semantic vectors for geographic locations, wherein a semantic vector includes a plurality of dimensions respectively representing types of places located within a threshold distance of a geographic location and, wherein each dimension from the plurality of dimensions has a respective value representing a respective number of places having a respective type represented by the dimension, and

the intrinsic affinity of the particular user for the product,

designing, based on the user-similarity-based affinity, a geo-fence targeted to the particular user such that a message is transmitted to a user device of the particular user in response to an indication that the particular user is within a boundary of the geo-fence, the geo-fence defining a geographical area that includes the particular geographical location and that is associated with a provider of the product; and

causing a telecommunication server to transmit the message to the user device when the user device is positioned within the designed geo-fence.

9. The non-transitory computer-readable medium of claim 8 , wherein the geo-fence is designed based on a location-specific affinity determined from the user-similarity-based affinity, wherein the operations further comprise determining the location-specific affinity by, at least:

computing the user-similarity-based affinity for the particular user from a weighted sum of intrinsic affinities for a set of users, wherein each weight of the weighted sum of the intrinsic affinities indicates a respective similarity between the particular user and a respective one of the users; and

computing the location-specific affinity from the user-similarity-based affinity and a weighted sum of additional user-similarity-based affinities for additional geographical locations wherein each weight of the weighted sum of the additional user-similarity-based affinities indicates a respective similarity between the particular geographical location and a respective one of the additional geographical locations.

10. The non-transitory computer-readable medium of claim 8 , wherein the user-similarity-based affinity is computed by performing operations comprising:

generating a multi-layer hierarchy of clusters of the semantic vectors, wherein each layer of the multi-layer hierarchy includes a respective set of clusters of the semantic vectors; and

computing similarities between the particular user and users from a set of users, respectively, wherein computing each similarity comprises, for each user:

determining layer similarities between the user and the particular user based on the sets of clusters, respectively, and

determining the similarity between the user and the particular user based on a weighted sum of the layer similarities.

11. The non-transitory computer-readable medium of claim 9 , wherein computing the location-specific affinity further comprises computing each similarity between the particular geographical location and respective one of the additional geographical locations by computing a respective cosine similarity between a respective pair of semantic vectors representing the particular geographical location and the respective one of the additional geographical locations.

12. The non-transitory computer-readable medium of claim 8 , wherein designing the geo-fence targeted to the particular user comprises:

computing, for the particular user, modeled location-specific affinities for geographical locations, respectively, wherein computing the modeled location-specific affinities for geographical locations comprises:

computing location-specific affinity vectors for the geographical locations and a set of users including the particular user, wherein each location-specific affinity vector comprises, for a respective one of the users, a respective distribution of location-specific affinities across the geographical locations,

segmenting the set of users by grouping the location-specific affinity vectors into location-specific affinity clusters, wherein the location-specific affinity clusters define respective segments of the users,

averaging, for a segment corresponding to the particular user, distributions of location-specific affinities in a location-specific affinity cluster corresponding to the segment, and

selecting averaged distribution of location-specific affinities as the modeled location-specific affinities for the particular user;

selecting a subset of the particular geographical locations based on each selected geographical location having a respective modeled location-specific affinity greater than a threshold affinity;

grouping the subset of the particular geographical locations into clusters; and

deriving the geo-fence from a boundary around a cluster including the particular geographical location.

13. A computing system comprising:

a non-transitory computer-readable medium storing computer executable program instructions; and

processing hardware configured for executing the program instructions and thereby perform operations comprising:

determining, from a set of taxonomy data defining a common characteristic of a set of similar products, cosine similarities between a product at a particular geographical location and additional products;

computing an intrinsic affinity of a particular user for the product from a weighted browse-count sum of (i) a browse count for the product and (ii) additional browse counts for the additional products, respectively, wherein each weight in the weighted browse-count sum indicates a respective cosine similarity between the product and a respective additional product,

computing a user-similarity-based affinity based on:

(a) semantic vectors for geographic locations, wherein a semantic vector includes a plurality of dimensions respectively representing types of places located within a threshold distance of a geographic location and, wherein each dimension from the plurality of dimensions has a respective value representing a respective number of places having a respective type represented by the dimension, and

(b) the intrinsic affinity of the particular user for the product, and

designing, based on the user-similarity-based affinity, a geo-fence targeted to the particular user such that a message is transmitted to a user device of the particular user in response to an indication that the particular user is within a boundary of the geo-fence, the geo-fence defining a geographical area that includes the particular geographical location and that is associated with a provider of the product; and

a server configured for transmitting the message to the user device when the user device is positioned within the designed geo-fence.

14. The computing system of claim 13 , wherein the geo-fence is designed based on a location-specific affinity determined from the user-similarity-based affinity, wherein the operations further comprise determining the location-specific affinity by, at least:

computing the user-similarity-based affinity for the particular user from a weighted sum of intrinsic affinities for a set of users, wherein each weight of the weighted sum of the intrinsic affinities indicates a respective similarity between the particular user and a respective one of the users; and

computing the location-specific affinity from the user-similarity-based affinity and a weighted sum of additional user-similarity-based affinities for additional geographical locations wherein each weight of the weighted sum of the additional user-similarity-based affinities indicates a respective similarity between the particular geographical location and a respective one of the additional geographical locations.

15. The computing system of claim 13 , wherein the user-similarity-based affinity is computed by performing operations comprising:

generating a multi-layer hierarchy of clusters of the semantic vectors, wherein each layer of the multi-layer hierarchy includes a respective set of clusters of the semantic vectors; and

computing similarities between the particular user and users from a set of users, respectively, wherein computing each similarity comprises, for each user:

determining layer similarities between the user and the particular user based on the sets of clusters, respectively, and

determining the similarity between the user and the particular user based on a weighted sum of the layer similarities.

16. The computing system of claim 14 , wherein computing the location-specific affinity further comprises computing each similarity between the particular geographical location and respective one of the additional geographical locations by computing a respective cosine similarity between a respective pair of semantic vectors representing the particular geographical location and the respective one of the additional geographical locations.

17. The computing system of claim 13 , wherein designing the geo-fence targeted to the particular user comprises:

computing, for the particular user, modeled location-specific affinities for geographical locations, respectively, wherein computing the modeled location-specific affinities for geographical locations comprises:

computing location-specific affinity vectors for the geographical locations and a set of users including the particular user, wherein each location-specific affinity vector comprises, for a respective one of the users, a respective distribution of location-specific affinities across the geographical locations,

segmenting the set of users by grouping the location-specific affinity vectors into location-specific affinity clusters, wherein the location-specific affinity clusters define respective segments of the users,

averaging, for a segment corresponding to the particular user, distributions of location-specific affinities in a location-specific affinity cluster corresponding to the segment, and

selecting the averaged distribution of location-specific affinities as the modeled location-specific affinities for the particular user;

selecting a subset of particular geographical locations based on each selected geographical location having a respective modeled location-specific affinity greater than a threshold affinity;

grouping the subset of the particular geographical locations into clusters; and

deriving the geo-fence from a boundary around a cluster including the particular geographical location.