Abstract: In one embodiment, a method includes receiving a search query, searching a multiple verticals to identify multiple of sets of objects in each respective vertical that match the search query, wherein each vertical stores objects of a particular object-type, generating a set of blended search results by blending the sets of identified objects from each vertical, determining that greater than a threshold proportion of objects in the set of blended search results are from a first vertical, adding at least one object from a second vertical to the set of blended search results in responsive to determining that greater than the threshold proportion of objects in the blended search results are from the first vertical, wherein the second vertical is different from the first vertical; and sending, responsive to the search query, the set of blended search results for display.

Abstract: To dynamically map nodes for locality and balance, computer implemented methods, systems, and computer readable media, in an embodiment, may compute histograms for nodes in a first partition. Histograms may be computed for nodes in a second partition. The second partition may be selected as a candidate partition for a set of nodes in the first partition based on the histograms for the nodes in the first partition. The first partition may be selected as a candidate partition for a set of nodes in the second partition based on the histograms for the nodes in the second partition. At least a portion of the set of nodes in the first partition may be mapped to the second partition and at least a portion of the set of nodes in the second partition may be mapped to the first partition based on load balancing.

Abstract: In one embodiment, a method includes receiving a search query, searching a multiple verticals to identify multiple of sets of objects in each respective vertical that match the search query, wherein each vertical stores objects of a particular object-type, generating a set of blended search results by blending the sets of identified objects from each vertical, determining that greater than a threshold proportion of objects in the set of blended search results are from a first vertical, adding at least one object from a second vertical to the set of blended search results in responsive to determining that greater than the threshold proportion of objects in the blended search results are from the first vertical, wherein the second vertical is different from the first vertical; and sending, responsive to the search query, the set of blended search results for display.

Abstract: Classifications associated with a plurality of nodes may be identified. The classifications may be grouped into first level communities based on edge weights between the classifications. The first level communities may be grouped into second level communities based on edge weights between the first level communities. A sorted list of the plurality of nodes may be generated based on the classifications, the first level communities, and the second level communities. Unique identifiers (IDs) may be assigned sequentially to the sorted list of the plurality of nodes.

Abstract: In one embodiment, a method includes receiving a search query from a user of an online social network and searching multiple verticals to identify multiple sets of objects in each vertical, respectively, that match the search query, and wherein each vertical stores one or more objects associated with the online social network. The method also includes ranking, for each set of identified objects from a vertical, each identified object in the set of identified objects. The method further includes blending the multiple sets of identified objects from each vertical to form a set of blended search results that includes a threshold number of identified objects, the blending including an iterative process performed at least the threshold number of iterations. Each iteration of the iterative blending process includes determining a blender score for each top-ranked identified object in each set of identified objects.

Abstract: A method includes training a neural network having parameters on training data, in which the neural network receives an input state and processes the input state to generate a respective score for each decision in a set of decisions. The method includes receiving training data including training text sequences and, for each training text sequence, a corresponding gold decision sequence. The method includes training the neural network on the training data to determine trained values of parameters of the neural network. Training the neural network includes for each training text sequence: maintaining a beam of candidate decision sequences for the training text sequence, updating each candidate decision sequence by adding one decision at a time, determining that a gold candidate decision sequence matching a prefix of the gold decision sequence has dropped out of the beam, and in response, performing an iteration of gradient descent to optimize an objective function.

Abstract: Unique identifiers (IDs) associated with a plurality of nodes may be provided. Nodes clustered within a community may be assigned numerically proximate unique IDs. A number of partitions associated with a plurality of machines may be determined. The unique IDs may be segmented into divisions based on the number of partitions. The unique IDs may be mapped to the plurality of machines based on the divisions.

Abstract: In one embodiment, a method includes receiving a search query from a user of an online social network and searching multiple verticals to identify multiple sets of objects in each vertical, respectively, that match the search query, and wherein each vertical stores one or more objects associated with the online social network. The method also includes ranking, for each set of identified objects from a vertical, each identified object in the set of identified objects. The method further includes blending the multiple sets of identified objects from each vertical to form a set of blended search results that includes a threshold number of identified objects, the blending including an iterative process performed at least the threshold number of iterations. Each iteration of the iterative blending process includes determining a blender score for each top-ranked identified object in each set of identified objects.

Abstract: Unique identifiers (IDs) associated with a plurality of nodes may be provided. Nodes clustered within a community may be assigned numerically proximate unique IDs. A number of partitions associated with a plurality of machines may be determined. The unique IDs may be segmented into divisions based on the number of partitions. The unique IDs may be mapped to the plurality of machines based on the divisions.

Abstract: Classifications associated with a plurality of nodes may be identified. The classifications may be grouped into first level communities based on edge weights between the classifications. The first level communities may be grouped into second level communities based on edge weights between the first level communities. A sorted list of the plurality of nodes may be generated based on the classifications, the first level communities, and the second level communities. Unique identifiers (IDs) may be assigned sequentially to the sorted list of the plurality of nodes.

Abstract: To dynamically map nodes for locality and balance, computer implemented methods, systems, and computer readable media, in an embodiment, may compute histograms for nodes in a first partition. Histograms may be computed for nodes in a second partition. The second partition may be selected as a candidate partition for a set of nodes in the first partition based on the histograms for the nodes in the first partition. The first partition may be selected as a candidate partition for a set of nodes in the second partition based on the histograms for the nodes in the second partition. At least a portion of the set of nodes in the first partition may be mapped to the second partition and at least a portion of the set of nodes in the second partition may be mapped to the first partition based on load balancing.