Industrial purposes of machine studying are generally composed of assorted gadgets which have differing information modalities or characteristic distributions. Heterogeneous graphs (HGs) supply a unified view of those multimodal information methods by defining a number of kinds of nodes (for every information sort) and edges (for the relation between information gadgets). As an illustration, e-commerce networks might need [user, product, review] nodes or video platforms might need [channel, user, video, comment] nodes. Heterogeneous graph neural networks (HGNNs) study node embeddings summarizing every node’s relationships right into a vector. Nonetheless, in actual world HGs, there may be usually a label imbalance subject between completely different node varieties. Which means label-scarce node varieties can not exploit HGNNs, which hampers the broader applicability of HGNNs.
In “Zero-shot Switch Studying inside a Heterogeneous Graph by way of Information Switch Networks”, introduced at NeurIPS 2022, we suggest a mannequin known as a Information Switch Community (KTN), which transfers data from label-abundant node varieties to zero-labeled node varieties utilizing the wealthy relational data given in a HG. We describe how we pre-train a HGNN mannequin with out the necessity for fine-tuning. KTNs outperform state-of-the-art switch studying baselines by as much as 140% on zero-shot studying duties, and can be utilized to enhance many current HGNN fashions on these duties by 24% (or extra).
KTNs remodel labels from one sort of data (squares) by a graph to a different sort (stars).
What’s a heterogeneous graph?
A HG consists of a number of node and edge varieties. The determine under exhibits an e-commerce community introduced as a HG. In e-commerce, “customers” buy “merchandise” and write “critiques”. A HG presents this ecosystem utilizing three node varieties [user, product, review] and three edge varieties [user-buy-product, user-write-review, review-on-product]. Particular person merchandise, customers, and critiques are then introduced as nodes and their relationships as edges within the HG with the corresponding node and edge varieties.
E-commerce heterogeneous graph.
Along with all connectivity data, HGs are generally given with enter node attributes that summarize every node’s data. Enter node attributes may have completely different modalities throughout completely different node varieties. As an illustration, pictures of merchandise might be given as enter node attributes for the product nodes, whereas textual content could be given as enter attributes to evaluation nodes. Node labels (e.g., the class of every product or the class that the majority pursuits every consumer) are what we wish to predict on every node.
HGNNs and label shortage points
HGNNs compute node embeddings that summarize every node’s native buildings (together with the node and its neighbor’s data). These node embeddings are utilized by a classifier to foretell every node’s label. To coach a HGNN mannequin and a classifier to foretell labels for a selected node sort, we require quantity of labels for the sort.
A typical subject in industrial purposes of deep studying is label shortage, and with their various node varieties, HGNNs are much more prone to face this problem. As an illustration, publicly accessible content material node varieties (e.g., product nodes) are abundantly labeled, whereas labels for consumer or account nodes will not be accessible as a consequence of privateness restrictions. Which means in most traditional coaching settings, HGNN fashions can solely study to make good inferences for a number of label-abundant node varieties and might normally not make any inferences for any remaining node varieties (given the absence of any labels for them).
Switch studying on heterogeneous graphs
Zero-shot switch studying is a method used to enhance the efficiency of a mannequin on a goal area with no labels by utilizing the data realized by the mannequin from one other associated supply area with adequately labeled information. To use switch studying to unravel this label shortage subject for sure node varieties in HGs, the goal area could be the zero-labeled node varieties. Then what could be the supply area? Earlier work generally units the supply area as the identical sort of nodes positioned in a distinct HG, assuming these nodes are abundantly labeled. This graph-to-graph switch studying method pre-trains a HGNN mannequin on the exterior HG after which runs the mannequin on the unique (label-scarce) HG.
Nonetheless, these approaches usually are not relevant in lots of real-world situations for 3 causes. First, any exterior HG that might be utilized in a graph-to-graph switch studying setting would nearly absolutely be proprietary, thus, seemingly unavailable. Second, even when practitioners may receive entry to an exterior HG, it’s unlikely the distribution of that supply HG would match their goal HG effectively sufficient to use switch studying. Lastly, node varieties affected by label shortage are prone to undergo the identical subject on different HGs (e.g., privateness points on consumer nodes).
Our method: Switch studying between node varieties inside a heterogeneous graph
Right here, we make clear a extra sensible supply area, different node varieties with considerable labels positioned on the identical HG. As an alternative of utilizing additional HGs, we switch data inside a single HG (assumed to be totally owned by the practitioners) throughout several types of nodes. Extra particularly, we pre-train a HGNN mannequin and a classifier on a label-abundant (supply) node sort, then reuse the fashions on the zero-labeled (goal) node varieties positioned in the identical HG with out further fine-tuning. The one requirement is that the supply and goal node varieties share the identical label set (e.g., within the e-commerce HG, product nodes have a label set describing product classes, and consumer nodes share the identical label set describing their favourite procuring classes).
Why is it difficult?
Sadly, we can not instantly reuse the pre-trained HGNN and classifier on the goal node sort. One essential attribute of HGNN architectures is that they’re composed of modules specialised to every node sort to completely study the multiplicity of HGs. HGNNs use distinct units of modules to compute embeddings for every node sort. Within the determine under, blue- and red-colored modules are used to compute node embeddings for the supply and goal node varieties, respectively.
HGNNs are composed of modules specialised to every node sort and use distinct units of modules to compute embeddings of various node varieties. Extra particulars could be discovered within the paper.
Whereas pre-training HGNNs on the supply node sort, source-specific modules within the HGNNs are effectively educated, nevertheless target-specific modules are under-trained as they’ve solely a small quantity of gradients flowing into them. That is proven under, the place we see that the L2 norm of gradients for goal node varieties (i.e., Mtt) are a lot decrease than for supply varieties (i.e., Mss). On this case a HGNN mannequin outputs poor node embeddings for the goal node sort, which leads to poor job efficiency.
In HGNNs, goal type-specific modules obtain zero or solely a small quantity of gradients throughout pre-training on the supply node sort, resulting in poor efficiency on the goal node sort.
KTN: Trainable cross-type switch studying for HGNNs
Our work focuses on reworking the (poor) goal node embeddings computed by a pre-trained HGNN mannequin to comply with the distribution of the supply node embeddings. Then the classifier, pre-trained on the supply node sort, could be reused for the goal node sort. How can we map the goal node embeddings to the supply area? To reply this query, we examine how HGNNs compute node embeddings to study the connection between supply and goal distributions.
HGNNs mixture linked node embeddings to enhance a goal node’s embeddings in every layer. In different phrases, the node embeddings for each supply and goal node varieties are up to date utilizing the identical enter — the earlier layer’s node embeddings of any linked node varieties. Which means they are often represented by one another. We show this relationship theoretically and discover there’s a mapping matrix (outlined by HGNN parameters) from the goal area to the supply area (extra particulars in Theorem 1 within the paper). Based mostly on this theorem, we introduce an auxiliary neural community, which we check with as a Information Switch Community (KTN), that receives the goal node embeddings after which transforms them by multiplying them with a (trainable) mapping matrix. We then outline a regularizer that’s minimized together with the efficiency loss within the pre-training section to coach the KTN. At check time, we map the goal embeddings computed from the pre-trained HGNN to the supply area utilizing the educated KTN for classification.
In HGNNs, the ultimate node embeddings of each supply and goal varieties are computed from completely different mathematical capabilities (f(): supply, g(): goal) which use the identical enter — the earlier layer’s node embeddings.
Experimental outcomes
To look at the effectiveness of KTNs, we ran 18 completely different zero-shot switch studying duties on two public heterogeneous graphs, Open Educational Graph and Pubmed. We evaluate KTN with eight state-of-the-art switch studying strategies (DAN, JAN, DANN, CDAN, CDAN-E, WDGRL, LP, EP). Proven under, KTN constantly outperforms all baselines on all duties, beating switch studying baselines by as much as 140% (as measured by Normalized Discounted Cumulative Acquire, a rating metric).
Zero-shot switch studying on Open Educational Graph (OAG-CS) and Pubmed datasets. The colours symbolize completely different classes of switch studying baselines towards which the outcomes are in contrast. Yellow: Use statistical properties (e.g., imply, variance) of distributions. Inexperienced: Use adversarial fashions to switch data. Orange: Switch data instantly by way of graph construction utilizing label propagation.
Most significantly, KTN could be utilized to nearly all HGNN fashions which have node and edge type-specific parameters and enhance their zero-shot efficiency on track domains. As proven under, KTN improves accuracy on zero-labeled node varieties throughout six completely different HGNN fashions(R-GCN, HAN, HGT, MAGNN, MPNN, H-MPNN) by as much as 190%.
KTN could be utilized to 6 completely different HGNN fashions and enhance their zero-shot efficiency on track domains.
Takeaways
Varied ecosystems in business could be introduced as heterogeneous graphs. HGNNs summarize heterogeneous graph data into efficient representations. Nonetheless, label shortage points on sure kinds of nodes forestall the broader software of HGNNs. On this submit, we launched KTN, the primary cross-type switch studying technique designed for HGNNs. With KTN, we are able to totally exploit the richness of heterogeneous graphs by way of HGNNs no matter label shortage. See the paper for extra particulars.
Acknowledgements
This paper is joint work with our co-authors John Palowitch (Google Analysis), Dustin Zelle (Google Analysis), Ziniu Hu (Intern, Google Analysis), and Russ Salakhutdinov (CMU). We thank Tom Small for creating the animated determine on this weblog submit.
